Upload folder using huggingface_hub

NiT-B/README.md

@@ -0,0 +1,90 @@
+---
+license: apache-2.0
+library_name: diffusers
+pipeline_tag: unconditional-image-generation
+tags:
+  - diffusers
+  - nit
+  - image-generation
+  - class-conditional
+  - imagenet
+inference: true
+---
+
+# NiT-B
+
+Self-contained Diffusers checkpoint for **NiT-B** (131M), converted from [`GoodEnough/NiT-B-Models`](https://huggingface.co/GoodEnough/NiT-B-Models) (`model_500K.safetensors`, 500K training steps).
+
+Architecture and training settings follow the official [`nit_b_pack_merge_radio_65536.yaml`](https://github.com/WZDTHU/NiT/blob/main/configs/c2i/nit_b_pack_merge_radio_65536.yaml).
+
+## Model config
+
+| Field | Value |
+| --- | --- |
+| Parameters | 131M |
+| Depth | 12 |
+| Hidden size | 768 |
+| Attention heads | 12 |
+| Encoder depth | 4 |
+| Latent channels (`z_dim`) | 1280 |
+| Patch size | 1 |
+| Input latent channels | 32 |
+| Classes | 1000 |
+| Class dropout | 0.1 |
+| QK norm | true |
+| VAE | `mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers` |
+| Flow path type | linear |
+
+## Recommended inference (256×256)
+
+Official NiT sampling defaults for **256×256** class-conditional ImageNet generation:
+
+| Setting | Value |
+| --- | --- |
+| Resolution | 256×256 |
+| Solver | SDE (Euler–Maruyama) in the official repo |
+| Steps (NFE) | 250 |
+| CFG scale | 2.25 |
+| CFG interval | (0.0, 0.7) |
+
+This Diffusers port uses [`FlowMatchEulerDiscreteScheduler`](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete) in deterministic ODE mode (`stochastic_sampling=false`). Keep the same step count, CFG scale, and interval as the official recipe.
+
+## Usage
+
+```python
+from pathlib import Path
+import torch
+from diffusers import DiffusionPipeline
+
+model_dir = Path(".")
+pipe = DiffusionPipeline.from_pretrained(
+    str(model_dir),
+    local_files_only=True,
+    custom_pipeline=str(model_dir / "pipeline.py"),
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+).to("cuda" if torch.cuda.is_available() else "cpu")
+
+generator = torch.Generator(device=pipe.device).manual_seed(42)
+image = pipe(
+    class_labels="golden retriever",
+    height=256,
+    width=256,
+    num_inference_steps=250,
+    guidance_scale=2.25,
+    guidance_interval=(0.0, 0.7),
+    generator=generator,
+).images[0]
+image.save("demo_256.png")
+```
+
+## Components
+
+- `pipeline.py` — custom `NiTPipeline`
+- `model_index.json` — pipeline index + ImageNet `id2label`
+- `transformer/config.json`
+- `transformer/nit_transformer_2d.py`
+- `transformer/diffusion_pytorch_model.safetensors`
+- `scheduler/scheduler_config.json`
+- `vae/config.json`
+- `vae/diffusion_pytorch_model.safetensors`

NiT-B/model_index.json

@@ -0,0 +1,1021 @@
+{
+  "_class_name": [
+    "pipeline",
+    "NiTPipeline"
+  ],
+  "_diffusers_version": "0.36.0",
+  "scheduler": [
+    "diffusers",
+    "FlowMatchEulerDiscreteScheduler"
+  ],
+  "transformer": [
+    "nit_transformer_2d",
+    "NiTTransformer2DModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderDC"
+  ],
+  "id2label": {
+    "0": "tench, Tinca tinca",
+    "1": "goldfish, Carassius auratus",
+    "2": "great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias",
+    "3": "tiger shark, Galeocerdo cuvieri",
+    "4": "hammerhead, hammerhead shark",
+    "5": "electric ray, crampfish, numbfish, torpedo",
+    "6": "stingray",
+    "7": "cock",
+    "8": "hen",
+    "9": "ostrich, Struthio camelus",
+    "10": "brambling, Fringilla montifringilla",
+    "11": "goldfinch, Carduelis carduelis",
+    "12": "house finch, linnet, Carpodacus mexicanus",
+    "13": "junco, snowbird",
+    "14": "indigo bunting, indigo finch, indigo bird, Passerina cyanea",
+    "15": "robin, American robin, Turdus migratorius",
+    "16": "bulbul",
+    "17": "jay",
+    "18": "magpie",
+    "19": "chickadee",
+    "20": "water ouzel, dipper",
+    "21": "kite",
+    "22": "bald eagle, American eagle, Haliaeetus leucocephalus",
+    "23": "vulture",
+    "24": "great grey owl, great gray owl, Strix nebulosa",
+    "25": "European fire salamander, Salamandra salamandra",
+    "26": "common newt, Triturus vulgaris",
+    "27": "eft",
+    "28": "spotted salamander, Ambystoma maculatum",
+    "29": "axolotl, mud puppy, Ambystoma mexicanum",
+    "30": "bullfrog, Rana catesbeiana",
+    "31": "tree frog, tree-frog",
+    "32": "tailed frog, bell toad, ribbed toad, tailed toad, Ascaphus trui",
+    "33": "loggerhead, loggerhead turtle, Caretta caretta",
+    "34": "leatherback turtle, leatherback, leathery turtle, Dermochelys coriacea",
+    "35": "mud turtle",
+    "36": "terrapin",
+    "37": "box turtle, box tortoise",
+    "38": "banded gecko",
+    "39": "common iguana, iguana, Iguana iguana",
+    "40": "American chameleon, anole, Anolis carolinensis",
+    "41": "whiptail, whiptail lizard",
+    "42": "agama",
+    "43": "frilled lizard, Chlamydosaurus kingi",
+    "44": "alligator lizard",
+    "45": "Gila monster, Heloderma suspectum",
+    "46": "green lizard, Lacerta viridis",
+    "47": "African chameleon, Chamaeleo chamaeleon",
+    "48": "Komodo dragon, Komodo lizard, dragon lizard, giant lizard, Varanus komodoensis",
+    "49": "African crocodile, Nile crocodile, Crocodylus niloticus",
+    "50": "American alligator, Alligator mississipiensis",
+    "51": "triceratops",
+    "52": "thunder snake, worm snake, Carphophis amoenus",
+    "53": "ringneck snake, ring-necked snake, ring snake",
+    "54": "hognose snake, puff adder, sand viper",
+    "55": "green snake, grass snake",
+    "56": "king snake, kingsnake",
+    "57": "garter snake, grass snake",
+    "58": "water snake",
+    "59": "vine snake",
+    "60": "night snake, Hypsiglena torquata",
+    "61": "boa constrictor, Constrictor constrictor",
+    "62": "rock python, rock snake, Python sebae",
+    "63": "Indian cobra, Naja naja",
+    "64": "green mamba",
+    "65": "sea snake",
+    "66": "horned viper, cerastes, sand viper, horned asp, Cerastes cornutus",
+    "67": "diamondback, diamondback rattlesnake, Crotalus adamanteus",
+    "68": "sidewinder, horned rattlesnake, Crotalus cerastes",
+    "69": "trilobite",
+    "70": "harvestman, daddy longlegs, Phalangium opilio",
+    "71": "scorpion",
+    "72": "black and gold garden spider, Argiope aurantia",
+    "73": "barn spider, Araneus cavaticus",
+    "74": "garden spider, Aranea diademata",
+    "75": "black widow, Latrodectus mactans",
+    "76": "tarantula",
+    "77": "wolf spider, hunting spider",
+    "78": "tick",
+    "79": "centipede",
+    "80": "black grouse",
+    "81": "ptarmigan",
+    "82": "ruffed grouse, partridge, Bonasa umbellus",
+    "83": "prairie chicken, prairie grouse, prairie fowl",
+    "84": "peacock",
+    "85": "quail",
+    "86": "partridge",
+    "87": "African grey, African gray, Psittacus erithacus",
+    "88": "macaw",
+    "89": "sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita",
+    "90": "lorikeet",
+    "91": "coucal",
+    "92": "bee eater",
+    "93": "hornbill",
+    "94": "hummingbird",
+    "95": "jacamar",
+    "96": "toucan",
+    "97": "drake",
+    "98": "red-breasted merganser, Mergus serrator",
+    "99": "goose",
+    "100": "black swan, Cygnus atratus",
+    "101": "tusker",
+    "102": "echidna, spiny anteater, anteater",
+    "103": "platypus, duckbill, duckbilled platypus, duck-billed platypus, Ornithorhynchus anatinus",
+    "104": "wallaby, brush kangaroo",
+    "105": "koala, koala bear, kangaroo bear, native bear, Phascolarctos cinereus",
+    "106": "wombat",
+    "107": "jellyfish",
+    "108": "sea anemone, anemone",
+    "109": "brain coral",
+    "110": "flatworm, platyhelminth",
+    "111": "nematode, nematode worm, roundworm",
+    "112": "conch",
+    "113": "snail",
+    "114": "slug",
+    "115": "sea slug, nudibranch",
+    "116": "chiton, coat-of-mail shell, sea cradle, polyplacophore",
+    "117": "chambered nautilus, pearly nautilus, nautilus",
+    "118": "Dungeness crab, Cancer magister",
+    "119": "rock crab, Cancer irroratus",
+    "120": "fiddler crab",
+    "121": "king crab, Alaska crab, Alaskan king crab, Alaska king crab, Paralithodes camtschatica",
+    "122": "American lobster, Northern lobster, Maine lobster, Homarus americanus",
+    "123": "spiny lobster, langouste, rock lobster, crawfish, crayfish, sea crawfish",
+    "124": "crayfish, crawfish, crawdad, crawdaddy",
+    "125": "hermit crab",
+    "126": "isopod",
+    "127": "white stork, Ciconia ciconia",
+    "128": "black stork, Ciconia nigra",
+    "129": "spoonbill",
+    "130": "flamingo",
+    "131": "little blue heron, Egretta caerulea",
+    "132": "American egret, great white heron, Egretta albus",
+    "133": "bittern",
+    "134": "crane",
+    "135": "limpkin, Aramus pictus",
+    "136": "European gallinule, Porphyrio porphyrio",
+    "137": "American coot, marsh hen, mud hen, water hen, Fulica americana",
+    "138": "bustard",
+    "139": "ruddy turnstone, Arenaria interpres",
+    "140": "red-backed sandpiper, dunlin, Erolia alpina",
+    "141": "redshank, Tringa totanus",
+    "142": "dowitcher",
+    "143": "oystercatcher, oyster catcher",
+    "144": "pelican",
+    "145": "king penguin, Aptenodytes patagonica",
+    "146": "albatross, mollymawk",
+    "147": "grey whale, gray whale, devilfish, Eschrichtius gibbosus, Eschrichtius robustus",
+    "148": "killer whale, killer, orca, grampus, sea wolf, Orcinus orca",
+    "149": "dugong, Dugong dugon",
+    "150": "sea lion",
+    "151": "Chihuahua",
+    "152": "Japanese spaniel",
+    "153": "Maltese dog, Maltese terrier, Maltese",
+    "154": "Pekinese, Pekingese, Peke",
+    "155": "Shih-Tzu",
+    "156": "Blenheim spaniel",
+    "157": "papillon",
+    "158": "toy terrier",
+    "159": "Rhodesian ridgeback",
+    "160": "Afghan hound, Afghan",
+    "161": "basset, basset hound",
+    "162": "beagle",
+    "163": "bloodhound, sleuthhound",
+    "164": "bluetick",
+    "165": "black-and-tan coonhound",
+    "166": "Walker hound, Walker foxhound",
+    "167": "English foxhound",
+    "168": "redbone",
+    "169": "borzoi, Russian wolfhound",
+    "170": "Irish wolfhound",
+    "171": "Italian greyhound",
+    "172": "whippet",
+    "173": "Ibizan hound, Ibizan Podenco",
+    "174": "Norwegian elkhound, elkhound",
+    "175": "otterhound, otter hound",
+    "176": "Saluki, gazelle hound",
+    "177": "Scottish deerhound, deerhound",
+    "178": "Weimaraner",
+    "179": "Staffordshire bullterrier, Staffordshire bull terrier",
+    "180": "American Staffordshire terrier, Staffordshire terrier, American pit bull terrier, pit bull terrier",
+    "181": "Bedlington terrier",
+    "182": "Border terrier",
+    "183": "Kerry blue terrier",
+    "184": "Irish terrier",
+    "185": "Norfolk terrier",
+    "186": "Norwich terrier",
+    "187": "Yorkshire terrier",
+    "188": "wire-haired fox terrier",
+    "189": "Lakeland terrier",
+    "190": "Sealyham terrier, Sealyham",
+    "191": "Airedale, Airedale terrier",
+    "192": "cairn, cairn terrier",
+    "193": "Australian terrier",
+    "194": "Dandie Dinmont, Dandie Dinmont terrier",
+    "195": "Boston bull, Boston terrier",
+    "196": "miniature schnauzer",
+    "197": "giant schnauzer",
+    "198": "standard schnauzer",
+    "199": "Scotch terrier, Scottish terrier, Scottie",
+    "200": "Tibetan terrier, chrysanthemum dog",
+    "201": "silky terrier, Sydney silky",
+    "202": "soft-coated wheaten terrier",
+    "203": "West Highland white terrier",
+    "204": "Lhasa, Lhasa apso",
+    "205": "flat-coated retriever",
+    "206": "curly-coated retriever",
+    "207": "golden retriever",
+    "208": "Labrador retriever",
+    "209": "Chesapeake Bay retriever",
+    "210": "German short-haired pointer",
+    "211": "vizsla, Hungarian pointer",
+    "212": "English setter",
+    "213": "Irish setter, red setter",
+    "214": "Gordon setter",
+    "215": "Brittany spaniel",
+    "216": "clumber, clumber spaniel",
+    "217": "English springer, English springer spaniel",
+    "218": "Welsh springer spaniel",
+    "219": "cocker spaniel, English cocker spaniel, cocker",
+    "220": "Sussex spaniel",
+    "221": "Irish water spaniel",
+    "222": "kuvasz",
+    "223": "schipperke",
+    "224": "groenendael",
+    "225": "malinois",
+    "226": "briard",
+    "227": "kelpie",
+    "228": "komondor",
+    "229": "Old English sheepdog, bobtail",
+    "230": "Shetland sheepdog, Shetland sheep dog, Shetland",
+    "231": "collie",
+    "232": "Border collie",
+    "233": "Bouvier des Flandres, Bouviers des Flandres",
+    "234": "Rottweiler",
+    "235": "German shepherd, German shepherd dog, German police dog, alsatian",
+    "236": "Doberman, Doberman pinscher",
+    "237": "miniature pinscher",
+    "238": "Greater Swiss Mountain dog",
+    "239": "Bernese mountain dog",
+    "240": "Appenzeller",
+    "241": "EntleBucher",
+    "242": "boxer",
+    "243": "bull mastiff",
+    "244": "Tibetan mastiff",
+    "245": "French bulldog",
+    "246": "Great Dane",
+    "247": "Saint Bernard, St Bernard",
+    "248": "Eskimo dog, husky",
+    "249": "malamute, malemute, Alaskan malamute",
+    "250": "Siberian husky",
+    "251": "dalmatian, coach dog, carriage dog",
+    "252": "affenpinscher, monkey pinscher, monkey dog",
+    "253": "basenji",
+    "254": "pug, pug-dog",
+    "255": "Leonberg",
+    "256": "Newfoundland, Newfoundland dog",
+    "257": "Great Pyrenees",
+    "258": "Samoyed, Samoyede",
+    "259": "Pomeranian",
+    "260": "chow, chow chow",
+    "261": "keeshond",
+    "262": "Brabancon griffon",
+    "263": "Pembroke, Pembroke Welsh corgi",
+    "264": "Cardigan, Cardigan Welsh corgi",
+    "265": "toy poodle",
+    "266": "miniature poodle",
+    "267": "standard poodle",
+    "268": "Mexican hairless",
+    "269": "timber wolf, grey wolf, gray wolf, Canis lupus",
+    "270": "white wolf, Arctic wolf, Canis lupus tundrarum",
+    "271": "red wolf, maned wolf, Canis rufus, Canis niger",
+    "272": "coyote, prairie wolf, brush wolf, Canis latrans",
+    "273": "dingo, warrigal, warragal, Canis dingo",
+    "274": "dhole, Cuon alpinus",
+    "275": "African hunting dog, hyena dog, Cape hunting dog, Lycaon pictus",
+    "276": "hyena, hyaena",
+    "277": "red fox, Vulpes vulpes",
+    "278": "kit fox, Vulpes macrotis",
+    "279": "Arctic fox, white fox, Alopex lagopus",
+    "280": "grey fox, gray fox, Urocyon cinereoargenteus",
+    "281": "tabby, tabby cat",
+    "282": "tiger cat",
+    "283": "Persian cat",
+    "284": "Siamese cat, Siamese",
+    "285": "Egyptian cat",
+    "286": "cougar, puma, catamount, mountain lion, painter, panther, Felis concolor",
+    "287": "lynx, catamount",
+    "288": "leopard, Panthera pardus",
+    "289": "snow leopard, ounce, Panthera uncia",
+    "290": "jaguar, panther, Panthera onca, Felis onca",
+    "291": "lion, king of beasts, Panthera leo",
+    "292": "tiger, Panthera tigris",
+    "293": "cheetah, chetah, Acinonyx jubatus",
+    "294": "brown bear, bruin, Ursus arctos",
+    "295": "American black bear, black bear, Ursus americanus, Euarctos americanus",
+    "296": "ice bear, polar bear, Ursus Maritimus, Thalarctos maritimus",
+    "297": "sloth bear, Melursus ursinus, Ursus ursinus",
+    "298": "mongoose",
+    "299": "meerkat, mierkat",
+    "300": "tiger beetle",
+    "301": "ladybug, ladybeetle, lady beetle, ladybird, ladybird beetle",
+    "302": "ground beetle, carabid beetle",
+    "303": "long-horned beetle, longicorn, longicorn beetle",
+    "304": "leaf beetle, chrysomelid",
+    "305": "dung beetle",
+    "306": "rhinoceros beetle",
+    "307": "weevil",
+    "308": "fly",
+    "309": "bee",
+    "310": "ant, emmet, pismire",
+    "311": "grasshopper, hopper",
+    "312": "cricket",
+    "313": "walking stick, walkingstick, stick insect",
+    "314": "cockroach, roach",
+    "315": "mantis, mantid",
+    "316": "cicada, cicala",
+    "317": "leafhopper",
+    "318": "lacewing, lacewing fly",
+    "319": "dragonfly, darning needle, devils darning needle, sewing needle, snake feeder, snake doctor, mosquito hawk, skeeter hawk",
+    "320": "damselfly",
+    "321": "admiral",
+    "322": "ringlet, ringlet butterfly",
+    "323": "monarch, monarch butterfly, milkweed butterfly, Danaus plexippus",
+    "324": "cabbage butterfly",
+    "325": "sulphur butterfly, sulfur butterfly",
+    "326": "lycaenid, lycaenid butterfly",
+    "327": "starfish, sea star",
+    "328": "sea urchin",
+    "329": "sea cucumber, holothurian",
+    "330": "wood rabbit, cottontail, cottontail rabbit",
+    "331": "hare",
+    "332": "Angora, Angora rabbit",
+    "333": "hamster",
+    "334": "porcupine, hedgehog",
+    "335": "fox squirrel, eastern fox squirrel, Sciurus niger",
+    "336": "marmot",
+    "337": "beaver",
+    "338": "guinea pig, Cavia cobaya",
+    "339": "sorrel",
+    "340": "zebra",
+    "341": "hog, pig, grunter, squealer, Sus scrofa",
+    "342": "wild boar, boar, Sus scrofa",
+    "343": "warthog",
+    "344": "hippopotamus, hippo, river horse, Hippopotamus amphibius",
+    "345": "ox",
+    "346": "water buffalo, water ox, Asiatic buffalo, Bubalus bubalis",
+    "347": "bison",
+    "348": "ram, tup",
+    "349": "bighorn, bighorn sheep, cimarron, Rocky Mountain bighorn, Rocky Mountain sheep, Ovis canadensis",
+    "350": "ibex, Capra ibex",
+    "351": "hartebeest",
+    "352": "impala, Aepyceros melampus",
+    "353": "gazelle",
+    "354": "Arabian camel, dromedary, Camelus dromedarius",
+    "355": "llama",
+    "356": "weasel",
+    "357": "mink",
+    "358": "polecat, fitch, foulmart, foumart, Mustela putorius",
+    "359": "black-footed ferret, ferret, Mustela nigripes",
+    "360": "otter",
+    "361": "skunk, polecat, wood pussy",
+    "362": "badger",
+    "363": "armadillo",
+    "364": "three-toed sloth, ai, Bradypus tridactylus",
+    "365": "orangutan, orang, orangutang, Pongo pygmaeus",
+    "366": "gorilla, Gorilla gorilla",
+    "367": "chimpanzee, chimp, Pan troglodytes",
+    "368": "gibbon, Hylobates lar",
+    "369": "siamang, Hylobates syndactylus, Symphalangus syndactylus",
+    "370": "guenon, guenon monkey",
+    "371": "patas, hussar monkey, Erythrocebus patas",
+    "372": "baboon",
+    "373": "macaque",
+    "374": "langur",
+    "375": "colobus, colobus monkey",
+    "376": "proboscis monkey, Nasalis larvatus",
+    "377": "marmoset",
+    "378": "capuchin, ringtail, Cebus capucinus",
+    "379": "howler monkey, howler",
+    "380": "titi, titi monkey",
+    "381": "spider monkey, Ateles geoffroyi",
+    "382": "squirrel monkey, Saimiri sciureus",
+    "383": "Madagascar cat, ring-tailed lemur, Lemur catta",
+    "384": "indri, indris, Indri indri, Indri brevicaudatus",
+    "385": "Indian elephant, Elephas maximus",
+    "386": "African elephant, Loxodonta africana",
+    "387": "lesser panda, red panda, panda, bear cat, cat bear, Ailurus fulgens",
+    "388": "giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca",
+    "389": "barracouta, snoek",
+    "390": "eel",
+    "391": "coho, cohoe, coho salmon, blue jack, silver salmon, Oncorhynchus kisutch",
+    "392": "rock beauty, Holocanthus tricolor",
+    "393": "anemone fish",
+    "394": "sturgeon",
+    "395": "gar, garfish, garpike, billfish, Lepisosteus osseus",
+    "396": "lionfish",
+    "397": "puffer, pufferfish, blowfish, globefish",
+    "398": "abacus",
+    "399": "abaya",
+    "400": "academic gown, academic robe, judge robe",
+    "401": "accordion, piano accordion, squeeze box",
+    "402": "acoustic guitar",
+    "403": "aircraft carrier, carrier, flattop, attack aircraft carrier",
+    "404": "airliner",
+    "405": "airship, dirigible",
+    "406": "altar",
+    "407": "ambulance",
+    "408": "amphibian, amphibious vehicle",
+    "409": "analog clock",
+    "410": "apiary, bee house",
+    "411": "apron",
+    "412": "ashcan, trash can, garbage can, wastebin, ash bin, ash-bin, ashbin, dustbin, trash barrel, trash bin",
+    "413": "assault rifle, assault gun",
+    "414": "backpack, back pack, knapsack, packsack, rucksack, haversack",
+    "415": "bakery, bakeshop, bakehouse",
+    "416": "balance beam, beam",
+    "417": "balloon",
+    "418": "ballpoint, ballpoint pen, ballpen, Biro",
+    "419": "Band Aid",
+    "420": "banjo",
+    "421": "bannister, banister, balustrade, balusters, handrail",
+    "422": "barbell",
+    "423": "barber chair",
+    "424": "barbershop",
+    "425": "barn",
+    "426": "barometer",
+    "427": "barrel, cask",
+    "428": "barrow, garden cart, lawn cart, wheelbarrow",
+    "429": "baseball",
+    "430": "basketball",
+    "431": "bassinet",
+    "432": "bassoon",
+    "433": "bathing cap, swimming cap",
+    "434": "bath towel",
+    "435": "bathtub, bathing tub, bath, tub",
+    "436": "beach wagon, station wagon, wagon, estate car, beach waggon, station waggon, waggon",
+    "437": "beacon, lighthouse, beacon light, pharos",
+    "438": "beaker",
+    "439": "bearskin, busby, shako",
+    "440": "beer bottle",
+    "441": "beer glass",
+    "442": "bell cote, bell cot",
+    "443": "bib",
+    "444": "bicycle-built-for-two, tandem bicycle, tandem",
+    "445": "bikini, two-piece",
+    "446": "binder, ring-binder",
+    "447": "binoculars, field glasses, opera glasses",
+    "448": "birdhouse",
+    "449": "boathouse",
+    "450": "bobsled, bobsleigh, bob",
+    "451": "bolo tie, bolo, bola tie, bola",
+    "452": "bonnet, poke bonnet",
+    "453": "bookcase",
+    "454": "bookshop, bookstore, bookstall",
+    "455": "bottlecap",
+    "456": "bow",
+    "457": "bow tie, bow-tie, bowtie",
+    "458": "brass, memorial tablet, plaque",
+    "459": "brassiere, bra, bandeau",
+    "460": "breakwater, groin, groyne, mole, bulwark, seawall, jetty",
+    "461": "breastplate, aegis, egis",
+    "462": "broom",
+    "463": "bucket, pail",
+    "464": "buckle",
+    "465": "bulletproof vest",
+    "466": "bullet train, bullet",
+    "467": "butcher shop, meat market",
+    "468": "cab, hack, taxi, taxicab",
+    "469": "caldron, cauldron",
+    "470": "candle, taper, wax light",
+    "471": "cannon",
+    "472": "canoe",
+    "473": "can opener, tin opener",
+    "474": "cardigan",
+    "475": "car mirror",
+    "476": "carousel, carrousel, merry-go-round, roundabout, whirligig",
+    "477": "carpenters kit, tool kit",
+    "478": "carton",
+    "479": "car wheel",
+    "480": "cash machine, cash dispenser, automated teller machine, automatic teller machine, automated teller, automatic teller, ATM",
+    "481": "cassette",
+    "482": "cassette player",
+    "483": "castle",
+    "484": "catamaran",
+    "485": "CD player",
+    "486": "cello, violoncello",
+    "487": "cellular telephone, cellular phone, cellphone, cell, mobile phone",
+    "488": "chain",
+    "489": "chainlink fence",
+    "490": "chain mail, ring mail, mail, chain armor, chain armour, ring armor, ring armour",
+    "491": "chain saw, chainsaw",
+    "492": "chest",
+    "493": "chiffonier, commode",
+    "494": "chime, bell, gong",
+    "495": "china cabinet, china closet",
+    "496": "Christmas stocking",
+    "497": "church, church building",
+    "498": "cinema, movie theater, movie theatre, movie house, picture palace",
+    "499": "cleaver, meat cleaver, chopper",
+    "500": "cliff dwelling",
+    "501": "cloak",
+    "502": "clog, geta, patten, sabot",
+    "503": "cocktail shaker",
+    "504": "coffee mug",
+    "505": "coffeepot",
+    "506": "coil, spiral, volute, whorl, helix",
+    "507": "combination lock",
+    "508": "computer keyboard, keypad",
+    "509": "confectionery, confectionary, candy store",
+    "510": "container ship, containership, container vessel",
+    "511": "convertible",
+    "512": "corkscrew, bottle screw",
+    "513": "cornet, horn, trumpet, trump",
+    "514": "cowboy boot",
+    "515": "cowboy hat, ten-gallon hat",
+    "516": "cradle",
+    "517": "crane",
+    "518": "crash helmet",
+    "519": "crate",
+    "520": "crib, cot",
+    "521": "Crock Pot",
+    "522": "croquet ball",
+    "523": "crutch",
+    "524": "cuirass",
+    "525": "dam, dike, dyke",
+    "526": "desk",
+    "527": "desktop computer",
+    "528": "dial telephone, dial phone",
+    "529": "diaper, nappy, napkin",
+    "530": "digital clock",
+    "531": "digital watch",
+    "532": "dining table, board",
+    "533": "dishrag, dishcloth",
+    "534": "dishwasher, dish washer, dishwashing machine",
+    "535": "disk brake, disc brake",
+    "536": "dock, dockage, docking facility",
+    "537": "dogsled, dog sled, dog sleigh",
+    "538": "dome",
+    "539": "doormat, welcome mat",
+    "540": "drilling platform, offshore rig",
+    "541": "drum, membranophone, tympan",
+    "542": "drumstick",
+    "543": "dumbbell",
+    "544": "Dutch oven",
+    "545": "electric fan, blower",
+    "546": "electric guitar",
+    "547": "electric locomotive",
+    "548": "entertainment center",
+    "549": "envelope",
+    "550": "espresso maker",
+    "551": "face powder",
+    "552": "feather boa, boa",
+    "553": "file, file cabinet, filing cabinet",
+    "554": "fireboat",
+    "555": "fire engine, fire truck",
+    "556": "fire screen, fireguard",
+    "557": "flagpole, flagstaff",
+    "558": "flute, transverse flute",
+    "559": "folding chair",
+    "560": "football helmet",
+    "561": "forklift",
+    "562": "fountain",
+    "563": "fountain pen",
+    "564": "four-poster",
+    "565": "freight car",
+    "566": "French horn, horn",
+    "567": "frying pan, frypan, skillet",
+    "568": "fur coat",
+    "569": "garbage truck, dustcart",
+    "570": "gasmask, respirator, gas helmet",
+    "571": "gas pump, gasoline pump, petrol pump, island dispenser",
+    "572": "goblet",
+    "573": "go-kart",
+    "574": "golf ball",
+    "575": "golfcart, golf cart",
+    "576": "gondola",
+    "577": "gong, tam-tam",
+    "578": "gown",
+    "579": "grand piano, grand",
+    "580": "greenhouse, nursery, glasshouse",
+    "581": "grille, radiator grille",
+    "582": "grocery store, grocery, food market, market",
+    "583": "guillotine",
+    "584": "hair slide",
+    "585": "hair spray",
+    "586": "half track",
+    "587": "hammer",
+    "588": "hamper",
+    "589": "hand blower, blow dryer, blow drier, hair dryer, hair drier",
+    "590": "hand-held computer, hand-held microcomputer",
+    "591": "handkerchief, hankie, hanky, hankey",
+    "592": "hard disc, hard disk, fixed disk",
+    "593": "harmonica, mouth organ, harp, mouth harp",
+    "594": "harp",
+    "595": "harvester, reaper",
+    "596": "hatchet",
+    "597": "holster",
+    "598": "home theater, home theatre",
+    "599": "honeycomb",
+    "600": "hook, claw",
+    "601": "hoopskirt, crinoline",
+    "602": "horizontal bar, high bar",
+    "603": "horse cart, horse-cart",
+    "604": "hourglass",
+    "605": "iPod",
+    "606": "iron, smoothing iron",
+    "607": "jack-o-lantern",
+    "608": "jean, blue jean, denim",
+    "609": "jeep, landrover",
+    "610": "jersey, T-shirt, tee shirt",
+    "611": "jigsaw puzzle",
+    "612": "jinrikisha, ricksha, rickshaw",
+    "613": "joystick",
+    "614": "kimono",
+    "615": "knee pad",
+    "616": "knot",
+    "617": "lab coat, laboratory coat",
+    "618": "ladle",
+    "619": "lampshade, lamp shade",
+    "620": "laptop, laptop computer",
+    "621": "lawn mower, mower",
+    "622": "lens cap, lens cover",
+    "623": "letter opener, paper knife, paperknife",
+    "624": "library",
+    "625": "lifeboat",
+    "626": "lighter, light, igniter, ignitor",
+    "627": "limousine, limo",
+    "628": "liner, ocean liner",
+    "629": "lipstick, lip rouge",
+    "630": "Loafer",
+    "631": "lotion",
+    "632": "loudspeaker, speaker, speaker unit, loudspeaker system, speaker system",
+    "633": "loupe, jewelers loupe",
+    "634": "lumbermill, sawmill",
+    "635": "magnetic compass",
+    "636": "mailbag, postbag",
+    "637": "mailbox, letter box",
+    "638": "maillot",
+    "639": "maillot, tank suit",
+    "640": "manhole cover",
+    "641": "maraca",
+    "642": "marimba, xylophone",
+    "643": "mask",
+    "644": "matchstick",
+    "645": "maypole",
+    "646": "maze, labyrinth",
+    "647": "measuring cup",
+    "648": "medicine chest, medicine cabinet",
+    "649": "megalith, megalithic structure",
+    "650": "microphone, mike",
+    "651": "microwave, microwave oven",
+    "652": "military uniform",
+    "653": "milk can",
+    "654": "minibus",
+    "655": "miniskirt, mini",
+    "656": "minivan",
+    "657": "missile",
+    "658": "mitten",
+    "659": "mixing bowl",
+    "660": "mobile home, manufactured home",
+    "661": "Model T",
+    "662": "modem",
+    "663": "monastery",
+    "664": "monitor",
+    "665": "moped",
+    "666": "mortar",
+    "667": "mortarboard",
+    "668": "mosque",
+    "669": "mosquito net",
+    "670": "motor scooter, scooter",
+    "671": "mountain bike, all-terrain bike, off-roader",
+    "672": "mountain tent",
+    "673": "mouse, computer mouse",
+    "674": "mousetrap",
+    "675": "moving van",
+    "676": "muzzle",
+    "677": "nail",
+    "678": "neck brace",
+    "679": "necklace",
+    "680": "nipple",
+    "681": "notebook, notebook computer",
+    "682": "obelisk",
+    "683": "oboe, hautboy, hautbois",
+    "684": "ocarina, sweet potato",
+    "685": "odometer, hodometer, mileometer, milometer",
+    "686": "oil filter",
+    "687": "organ, pipe organ",
+    "688": "oscilloscope, scope, cathode-ray oscilloscope, CRO",
+    "689": "overskirt",
+    "690": "oxcart",
+    "691": "oxygen mask",
+    "692": "packet",
+    "693": "paddle, boat paddle",
+    "694": "paddlewheel, paddle wheel",
+    "695": "padlock",
+    "696": "paintbrush",
+    "697": "pajama, pyjama, pjs, jammies",
+    "698": "palace",
+    "699": "panpipe, pandean pipe, syrinx",
+    "700": "paper towel",
+    "701": "parachute, chute",
+    "702": "parallel bars, bars",
+    "703": "park bench",
+    "704": "parking meter",
+    "705": "passenger car, coach, carriage",
+    "706": "patio, terrace",
+    "707": "pay-phone, pay-station",
+    "708": "pedestal, plinth, footstall",
+    "709": "pencil box, pencil case",
+    "710": "pencil sharpener",
+    "711": "perfume, essence",
+    "712": "Petri dish",
+    "713": "photocopier",
+    "714": "pick, plectrum, plectron",
+    "715": "pickelhaube",
+    "716": "picket fence, paling",
+    "717": "pickup, pickup truck",
+    "718": "pier",
+    "719": "piggy bank, penny bank",
+    "720": "pill bottle",
+    "721": "pillow",
+    "722": "ping-pong ball",
+    "723": "pinwheel",
+    "724": "pirate, pirate ship",
+    "725": "pitcher, ewer",
+    "726": "plane, carpenters plane, woodworking plane",
+    "727": "planetarium",
+    "728": "plastic bag",
+    "729": "plate rack",
+    "730": "plow, plough",
+    "731": "plunger, plumbers helper",
+    "732": "Polaroid camera, Polaroid Land camera",
+    "733": "pole",
+    "734": "police van, police wagon, paddy wagon, patrol wagon, wagon, black Maria",
+    "735": "poncho",
+    "736": "pool table, billiard table, snooker table",
+    "737": "pop bottle, soda bottle",
+    "738": "pot, flowerpot",
+    "739": "potters wheel",
+    "740": "power drill",
+    "741": "prayer rug, prayer mat",
+    "742": "printer",
+    "743": "prison, prison house",
+    "744": "projectile, missile",
+    "745": "projector",
+    "746": "puck, hockey puck",
+    "747": "punching bag, punch bag, punching ball, punchball",
+    "748": "purse",
+    "749": "quill, quill pen",
+    "750": "quilt, comforter, comfort, puff",
+    "751": "racer, race car, racing car",
+    "752": "racket, racquet",
+    "753": "radiator",
+    "754": "radio, wireless",
+    "755": "radio telescope, radio reflector",
+    "756": "rain barrel",
+    "757": "recreational vehicle, RV, R.V.",
+    "758": "reel",
+    "759": "reflex camera",
+    "760": "refrigerator, icebox",
+    "761": "remote control, remote",
+    "762": "restaurant, eating house, eating place, eatery",
+    "763": "revolver, six-gun, six-shooter",
+    "764": "rifle",
+    "765": "rocking chair, rocker",
+    "766": "rotisserie",
+    "767": "rubber eraser, rubber, pencil eraser",
+    "768": "rugby ball",
+    "769": "rule, ruler",
+    "770": "running shoe",
+    "771": "safe",
+    "772": "safety pin",
+    "773": "saltshaker, salt shaker",
+    "774": "sandal",
+    "775": "sarong",
+    "776": "sax, saxophone",
+    "777": "scabbard",
+    "778": "scale, weighing machine",
+    "779": "school bus",
+    "780": "schooner",
+    "781": "scoreboard",
+    "782": "screen, CRT screen",
+    "783": "screw",
+    "784": "screwdriver",
+    "785": "seat belt, seatbelt",
+    "786": "sewing machine",
+    "787": "shield, buckler",
+    "788": "shoe shop, shoe-shop, shoe store",
+    "789": "shoji",
+    "790": "shopping basket",
+    "791": "shopping cart",
+    "792": "shovel",
+    "793": "shower cap",
+    "794": "shower curtain",
+    "795": "ski",
+    "796": "ski mask",
+    "797": "sleeping bag",
+    "798": "slide rule, slipstick",
+    "799": "sliding door",
+    "800": "slot, one-armed bandit",
+    "801": "snorkel",
+    "802": "snowmobile",
+    "803": "snowplow, snowplough",
+    "804": "soap dispenser",
+    "805": "soccer ball",
+    "806": "sock",
+    "807": "solar dish, solar collector, solar furnace",
+    "808": "sombrero",
+    "809": "soup bowl",
+    "810": "space bar",
+    "811": "space heater",
+    "812": "space shuttle",
+    "813": "spatula",
+    "814": "speedboat",
+    "815": "spider web, spiders web",
+    "816": "spindle",
+    "817": "sports car, sport car",
+    "818": "spotlight, spot",
+    "819": "stage",
+    "820": "steam locomotive",
+    "821": "steel arch bridge",
+    "822": "steel drum",
+    "823": "stethoscope",
+    "824": "stole",
+    "825": "stone wall",
+    "826": "stopwatch, stop watch",
+    "827": "stove",
+    "828": "strainer",
+    "829": "streetcar, tram, tramcar, trolley, trolley car",
+    "830": "stretcher",
+    "831": "studio couch, day bed",
+    "832": "stupa, tope",
+    "833": "submarine, pigboat, sub, U-boat",
+    "834": "suit, suit of clothes",
+    "835": "sundial",
+    "836": "sunglass",
+    "837": "sunglasses, dark glasses, shades",
+    "838": "sunscreen, sunblock, sun blocker",
+    "839": "suspension bridge",
+    "840": "swab, swob, mop",
+    "841": "sweatshirt",
+    "842": "swimming trunks, bathing trunks",
+    "843": "swing",
+    "844": "switch, electric switch, electrical switch",
+    "845": "syringe",
+    "846": "table lamp",
+    "847": "tank, army tank, armored combat vehicle, armoured combat vehicle",
+    "848": "tape player",
+    "849": "teapot",
+    "850": "teddy, teddy bear",
+    "851": "television, television system",
+    "852": "tennis ball",
+    "853": "thatch, thatched roof",
+    "854": "theater curtain, theatre curtain",
+    "855": "thimble",
+    "856": "thresher, thrasher, threshing machine",
+    "857": "throne",
+    "858": "tile roof",
+    "859": "toaster",
+    "860": "tobacco shop, tobacconist shop, tobacconist",
+    "861": "toilet seat",
+    "862": "torch",
+    "863": "totem pole",
+    "864": "tow truck, tow car, wrecker",
+    "865": "toyshop",
+    "866": "tractor",
+    "867": "trailer truck, tractor trailer, trucking rig, rig, articulated lorry, semi",
+    "868": "tray",
+    "869": "trench coat",
+    "870": "tricycle, trike, velocipede",
+    "871": "trimaran",
+    "872": "tripod",
+    "873": "triumphal arch",
+    "874": "trolleybus, trolley coach, trackless trolley",
+    "875": "trombone",
+    "876": "tub, vat",
+    "877": "turnstile",
+    "878": "typewriter keyboard",
+    "879": "umbrella",
+    "880": "unicycle, monocycle",
+    "881": "upright, upright piano",
+    "882": "vacuum, vacuum cleaner",
+    "883": "vase",
+    "884": "vault",
+    "885": "velvet",
+    "886": "vending machine",
+    "887": "vestment",
+    "888": "viaduct",
+    "889": "violin, fiddle",
+    "890": "volleyball",
+    "891": "waffle iron",
+    "892": "wall clock",
+    "893": "wallet, billfold, notecase, pocketbook",
+    "894": "wardrobe, closet, press",
+    "895": "warplane, military plane",
+    "896": "washbasin, handbasin, washbowl, lavabo, wash-hand basin",
+    "897": "washer, automatic washer, washing machine",
+    "898": "water bottle",
+    "899": "water jug",
+    "900": "water tower",
+    "901": "whiskey jug",
+    "902": "whistle",
+    "903": "wig",
+    "904": "window screen",
+    "905": "window shade",
+    "906": "Windsor tie",
+    "907": "wine bottle",
+    "908": "wing",
+    "909": "wok",
+    "910": "wooden spoon",
+    "911": "wool, woolen, woollen",
+    "912": "worm fence, snake fence, snake-rail fence, Virginia fence",
+    "913": "wreck",
+    "914": "yawl",
+    "915": "yurt",
+    "916": "web site, website, internet site, site",
+    "917": "comic book",
+    "918": "crossword puzzle, crossword",
+    "919": "street sign",
+    "920": "traffic light, traffic signal, stoplight",
+    "921": "book jacket, dust cover, dust jacket, dust wrapper",
+    "922": "menu",
+    "923": "plate",
+    "924": "guacamole",
+    "925": "consomme",
+    "926": "hot pot, hotpot",
+    "927": "trifle",
+    "928": "ice cream, icecream",
+    "929": "ice lolly, lolly, lollipop, popsicle",
+    "930": "French loaf",
+    "931": "bagel, beigel",
+    "932": "pretzel",
+    "933": "cheeseburger",
+    "934": "hotdog, hot dog, red hot",
+    "935": "mashed potato",
+    "936": "head cabbage",
+    "937": "broccoli",
+    "938": "cauliflower",
+    "939": "zucchini, courgette",
+    "940": "spaghetti squash",
+    "941": "acorn squash",
+    "942": "butternut squash",
+    "943": "cucumber, cuke",
+    "944": "artichoke, globe artichoke",
+    "945": "bell pepper",
+    "946": "cardoon",
+    "947": "mushroom",
+    "948": "Granny Smith",
+    "949": "strawberry",
+    "950": "orange",
+    "951": "lemon",
+    "952": "fig",
+    "953": "pineapple, ananas",
+    "954": "banana",
+    "955": "jackfruit, jak, jack",
+    "956": "custard apple",
+    "957": "pomegranate",
+    "958": "hay",
+    "959": "carbonara",
+    "960": "chocolate sauce, chocolate syrup",
+    "961": "dough",
+    "962": "meat loaf, meatloaf",
+    "963": "pizza, pizza pie",
+    "964": "potpie",
+    "965": "burrito",
+    "966": "red wine",
+    "967": "espresso",
+    "968": "cup",
+    "969": "eggnog",
+    "970": "alp",
+    "971": "bubble",
+    "972": "cliff, drop, drop-off",
+    "973": "coral reef",
+    "974": "geyser",
+    "975": "lakeside, lakeshore",
+    "976": "promontory, headland, head, foreland",
+    "977": "sandbar, sand bar",
+    "978": "seashore, coast, seacoast, sea-coast",
+    "979": "valley, vale",
+    "980": "volcano",
+    "981": "ballplayer, baseball player",
+    "982": "groom, bridegroom",
+    "983": "scuba diver",
+    "984": "rapeseed",
+    "985": "daisy",
+    "986": "yellow ladys slipper, yellow lady-slipper, Cypripedium calceolus, Cypripedium parviflorum",
+    "987": "corn",
+    "988": "acorn",
+    "989": "hip, rose hip, rosehip",
+    "990": "buckeye, horse chestnut, conker",
+    "991": "coral fungus",
+    "992": "agaric",
+    "993": "gyromitra",
+    "994": "stinkhorn, carrion fungus",
+    "995": "earthstar",
+    "996": "hen-of-the-woods, hen of the woods, Polyporus frondosus, Grifola frondosa",
+    "997": "bolete",
+    "998": "ear, spike, capitulum",
+    "999": "toilet tissue, toilet paper, bathroom tissue"
+  }
+}

NiT-B/pipeline.py

@@ -0,0 +1,483 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils.torch_utils import randn_tensor
+
+# Local component classes are loaded dynamically in from_pretrained.
+
+DEFAULT_NATIVE_RESOLUTION = 256
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> from pathlib import Path
+        >>> import torch
+        >>> from diffusers import DiffusionPipeline
+
+        >>> model_dir = Path("./NiT-B").resolve()
+        >>> pipe = DiffusionPipeline.from_pretrained(
+        ...     str(model_dir),
+        ...     local_files_only=True,
+        ...     custom_pipeline=str(model_dir / "pipeline.py"),
+        ...     trust_remote_code=True,
+        ...     torch_dtype=torch.bfloat16,
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> print(pipe.id2label[207])
+        >>> print(pipe.get_label_ids("golden retriever"))
+
+        >>> generator = torch.Generator(device="cuda").manual_seed(42)
+        >>> image = pipe(
+        ...     class_labels="golden retriever",
+        ...     height=256,
+        ...     width=256,
+        ...     num_inference_steps=250,
+        ...     guidance_scale=2.25,
+        ...     guidance_interval=(0.0, 0.7),
+        ...     generator=generator,
+        ... ).images[0]
+        >>> image.save("demo.png")
+        ```
+"""
+
+
+class NiTPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for native-resolution class-conditional image generation with NiT.
+
+    Uses the native [`FlowMatchEulerDiscreteScheduler`] in deterministic (ODE) mode.
+    The official NiT repo defaults to an Euler-Maruyama SDE sampler for 512×512; that SDE is
+    not the same as the scheduler's `stochastic_sampling` path, so keep
+    `scheduler.config.stochastic_sampling=False` and let the scheduler perform the ODE update
+    `x_{t+dt} = x_t + dt * v`.
+
+    Parameters:
+        transformer ([`NiTTransformer2DModel`]):
+            Class-conditional transformer that predicts flow-matching velocity in packed latent space.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            Native diffusers flow-matching Euler scheduler (`stochastic_sampling=False`).
+        vae ([`AutoencoderDC`] or [`AutoencoderKL`], *optional*):
+            Variational autoencoder used to decode packed transformer latents to pixels.
+        id2label (`dict[int, str]`, *optional*):
+            ImageNet class id to English label mapping. Values may contain comma-separated synonyms.
+    """
+
+    model_cpu_offload_seq = "transformer->vae"
+    _optional_components = ["vae"]
+
+    def __init__(
+        self,
+        transformer,
+        scheduler,
+        vae=None,
+        id2label: Optional[Dict[Union[int, str], str]] = None,
+    ):
+        super().__init__()
+        self.register_modules(transformer=transformer, scheduler=scheduler, vae=vae)
+        self.image_processor = VaeImageProcessor()
+        self._id2label = self._normalize_id2label(id2label)
+        self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = bool(self._id2label)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path=None, subfolder=None, **kwargs):
+        """Load a self-contained variant folder locally or from the Hub."""
+        import importlib
+        import sys
+
+        repo_root = Path(__file__).resolve().parent
+
+        if pretrained_model_name_or_path in (None, "", "."):
+            variant = repo_root
+        elif (
+            isinstance(pretrained_model_name_or_path, str)
+            and "/" in pretrained_model_name_or_path
+            and not Path(pretrained_model_name_or_path).exists()
+        ):
+            from huggingface_hub import snapshot_download
+
+            hub_kwargs = dict(kwargs.pop("hub_kwargs", {}))
+            if subfolder:
+                hub_kwargs.setdefault("allow_patterns", [f"{subfolder}/**"])
+            cache_dir = snapshot_download(pretrained_model_name_or_path, **hub_kwargs)
+            variant = Path(cache_dir) / subfolder if subfolder else Path(cache_dir)
+        else:
+            variant = Path(pretrained_model_name_or_path)
+            if not variant.is_absolute():
+                candidate = (Path.cwd() / variant).resolve()
+                variant = candidate if candidate.exists() else (repo_root / variant).resolve()
+            if subfolder:
+                variant = variant / subfolder
+
+        id2label_override = kwargs.pop("id2label", None)
+        model_kwargs = dict(kwargs)
+        inserted: List[str] = []
+
+        def _load_component(folder: str, module_name: str, class_name: str):
+            comp_dir = variant / folder
+            module_path = comp_dir / f"{module_name}.py"
+            has_weights = (comp_dir / "config.json").exists() or (comp_dir / "scheduler_config.json").exists()
+            if not module_path.exists() or not has_weights:
+                return None
+
+            comp_path = str(comp_dir)
+            if comp_path not in sys.path:
+                sys.path.insert(0, comp_path)
+                inserted.append(comp_path)
+
+            module = importlib.import_module(module_name)
+            component_cls = getattr(module, class_name)
+            return component_cls.from_pretrained(str(comp_dir), **model_kwargs)
+
+        try:
+            transformer = _load_component("transformer", "nit_transformer_2d", "NiTTransformer2DModel")
+            try:
+                scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(str(variant), subfolder="scheduler")
+            except Exception:
+                scheduler = FlowMatchEulerDiscreteScheduler(
+                    num_train_timesteps=1000,
+                    shift=1.0,
+                    stochastic_sampling=False,
+                )
+            if transformer is None:
+                raise ValueError(f"No loadable transformer found under {variant}")
+
+            vae = None
+            vae_dir = variant / "vae"
+            if vae_dir.exists() and (vae_dir / "config.json").exists():
+                from diffusers import AutoencoderDC, AutoencoderKL
+
+                vae_class_name = json.loads((vae_dir / "config.json").read_text(encoding="utf-8")).get(
+                    "_class_name", "AutoencoderDC"
+                )
+                vae_cls = AutoencoderDC if vae_class_name == "AutoencoderDC" else AutoencoderKL
+                vae = vae_cls.from_pretrained(str(vae_dir), **model_kwargs)
+
+            id2label = id2label_override or cls._read_id2label_from_model_index(str(variant))
+            pipe = cls(
+                transformer=transformer,
+                scheduler=scheduler,
+                vae=vae,
+                id2label=id2label,
+            )
+            if hasattr(pipe, "register_to_config"):
+                pipe.register_to_config(_name_or_path=str(variant))
+            return pipe
+        finally:
+            for comp_path in inserted:
+                if comp_path in sys.path:
+                    sys.path.remove(comp_path)
+
+    def _ensure_labels_loaded(self) -> None:
+        if self._labels_loaded_from_model_index:
+            return
+        loaded = self._read_id2label_from_model_index(getattr(self.config, "_name_or_path", None))
+        if loaded:
+            self._id2label = loaded
+            self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = True
+
+    @staticmethod
+    def _normalize_id2label(id2label: Optional[Dict[Union[int, str], str]]) -> Dict[int, str]:
+        if not id2label:
+            return {}
+        return {int(key): value for key, value in id2label.items()}
+
+    @staticmethod
+    def _read_id2label_from_model_index(variant_path: Optional[str]) -> Dict[int, str]:
+        if not variant_path:
+            return {}
+        variant_dir = Path(variant_path).resolve()
+        model_index_path = variant_dir / "model_index.json"
+        if not model_index_path.exists():
+            return {}
+        raw = json.loads(model_index_path.read_text(encoding="utf-8"))
+        id2label = raw.get("id2label")
+        if not isinstance(id2label, dict):
+            return {}
+        return {int(key): value for key, value in id2label.items()}
+
+    @staticmethod
+    def _build_label2id(id2label: Dict[int, str]) -> Dict[str, int]:
+        label2id: Dict[str, int] = {}
+        for class_id, value in id2label.items():
+            for synonym in value.split(","):
+                synonym = synonym.strip()
+                if synonym:
+                    label2id[synonym] = int(class_id)
+        return dict(sorted(label2id.items()))
+
+    @property
+    def id2label(self) -> Dict[int, str]:
+        r"""ImageNet class id to English label string (comma-separated synonyms)."""
+        self._ensure_labels_loaded()
+        return self._id2label
+
+    def get_label_ids(self, label: Union[str, List[str]]) -> List[int]:
+        r"""
+        Map ImageNet label strings to class ids.
+
+        Args:
+            label (`str` or `list[str]`):
+                One or more English label strings. Each string must match a synonym in `id2label`.
+        """
+        self._ensure_labels_loaded()
+        label2id = self.labels
+        if not label2id:
+            raise ValueError("No English labels loaded. Ensure `id2label` exists in model_index.json.")
+
+        if isinstance(label, str):
+            label = [label]
+
+        missing = [item for item in label if item not in label2id]
+        if missing:
+            preview = ", ".join(list(label2id.keys())[:8])
+            raise ValueError(f"Unknown English label(s): {missing}. Example valid labels: {preview}, ...")
+        return [label2id[item] for item in label]
+
+    def _normalize_class_labels(
+        self,
+        class_labels: Union[int, str, List[Union[int, str]], torch.LongTensor],
+    ) -> torch.LongTensor:
+        if torch.is_tensor(class_labels):
+            return class_labels.to(device=self._execution_device, dtype=torch.long).reshape(-1)
+
+        if isinstance(class_labels, int):
+            class_label_ids = [class_labels]
+        elif isinstance(class_labels, str):
+            class_label_ids = self.get_label_ids(class_labels)
+        elif class_labels and isinstance(class_labels[0], str):
+            class_label_ids = self.get_label_ids(class_labels)
+        else:
+            class_label_ids = list(class_labels)
+
+        return torch.tensor(class_label_ids, device=self._execution_device, dtype=torch.long).reshape(-1)
+
+    def _get_vae_spatial_downsample(self) -> int:
+        if self.vae is None:
+            return 1
+        if self.vae.__class__.__name__ == "AutoencoderDC" or "dc-ae" in getattr(
+            self.vae.config, "_name_or_path", ""
+        ):
+            return 32
+        block_out_channels = getattr(self.vae.config, "block_out_channels", [0, 0, 0, 0])
+        return 2 ** (len(block_out_channels) - 1)
+
+    def check_inputs(
+        self,
+        height: int,
+        width: int,
+        num_inference_steps: int,
+        output_type: str,
+    ) -> None:
+        if num_inference_steps < 1:
+            raise ValueError("num_inference_steps must be >= 1.")
+        if output_type not in {"pil", "np", "pt", "latent"}:
+            raise ValueError("output_type must be one of: 'pil', 'np', 'pt', 'latent'.")
+
+        spatial_downsample = self._get_vae_spatial_downsample()
+        if height % spatial_downsample != 0 or width % spatial_downsample != 0:
+            raise ValueError(
+                f"height and width must be divisible by the VAE downsample factor {spatial_downsample}."
+            )
+
+        patch_size = int(self.transformer.config.patch_size)
+        latent_height = height // spatial_downsample
+        latent_width = width // spatial_downsample
+        if latent_height % patch_size != 0 or latent_width % patch_size != 0:
+            raise ValueError("Latent height and width must be divisible by transformer's patch_size.")
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]],
+    ) -> Tuple[torch.Tensor, torch.LongTensor]:
+        spatial_downsample = self._get_vae_spatial_downsample()
+        latent_height = height // spatial_downsample
+        latent_width = width // spatial_downsample
+        patch_size = int(self.transformer.config.patch_size)
+        token_height = latent_height // patch_size
+        token_width = latent_width // patch_size
+        image_sizes = torch.tensor([[token_height, token_width]] * batch_size, device=device, dtype=torch.long)
+
+        packed_shape = (
+            batch_size * token_height * token_width,
+            self.transformer.config.in_channels,
+            patch_size,
+            patch_size,
+        )
+        packed_latents = randn_tensor(
+            packed_shape,
+            generator=generator,
+            device=device,
+            dtype=dtype,
+        )
+        return packed_latents, image_sizes
+
+    @staticmethod
+    def _flow_time_from_scheduler_timestep(timestep: torch.Tensor, num_train_timesteps: int) -> float:
+        """Map native scheduler timesteps (sigma * num_train_timesteps) to NiT flow time in [0, 1]."""
+        return float(timestep) / num_train_timesteps
+
+    def _apply_classifier_free_guidance(
+        self,
+        model_output: torch.Tensor,
+        guidance_scale: float,
+        guidance_active: bool,
+    ) -> torch.Tensor:
+        if guidance_scale <= 1.0 or not guidance_active:
+            return model_output
+        model_output_cond, model_output_uncond = model_output.chunk(2)
+        return model_output_uncond + guidance_scale * (model_output_cond - model_output_uncond)
+
+    def _get_vae_dtype(self, latents: torch.Tensor) -> torch.dtype:
+        vae_dtype = getattr(self.vae, "dtype", None)
+        if vae_dtype is not None:
+            return vae_dtype
+        vae_params = next(self.vae.parameters(), None)
+        return vae_params.dtype if vae_params is not None else latents.dtype
+
+    def decode_latents(self, latents: torch.Tensor, output_type: str = "pil"):
+        if self.vae is None:
+            if output_type == "latent":
+                return latents
+            raise ValueError("Cannot decode latents without a VAE.")
+
+        vae_dtype = self._get_vae_dtype(latents)
+        latents = latents.to(dtype=vae_dtype)
+        scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
+        shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
+        latents = (latents / scaling_factor) + shift_factor
+        if output_type == "latent":
+            return latents
+
+        image = self.vae.decode(latents, return_dict=False)[0]
+        return self.image_processor.postprocess(image, output_type=output_type)
+
+    @torch.inference_mode()
+    def __call__(
+        self,
+        class_labels: Union[int, str, List[Union[int, str]], torch.LongTensor],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 1.0,
+        guidance_interval: Tuple[float, float] = (0.0, 1.0),
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: str = "pil",
+        return_dict: bool = True,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        r"""
+        Generate class-conditional images at native resolution.
+
+        Args:
+            class_labels (`int`, `str`, `list[int]`, `list[str]`, or `torch.LongTensor`):
+                ImageNet class indices or human-readable English label strings.
+            height (`int`, *optional*):
+                Output image height in pixels. Defaults to `512` when a VAE is present.
+            width (`int`, *optional*):
+                Output image width in pixels. Defaults to `512` when a VAE is present.
+            num_inference_steps (`int`, defaults to `50`):
+                Number of denoising steps.
+            guidance_scale (`float`, defaults to `1.0`):
+                Classifier-free guidance scale. CFG is active when `guidance_scale > 1.0`.
+            guidance_interval (`tuple[float, float]`, defaults to `(0.0, 1.0)`):
+                Flow-time interval where CFG is applied. Uses continuous flow time
+                `timestep / num_train_timesteps`, matching the official NiT ODE sampler.
+            generator (`torch.Generator`, *optional*):
+                RNG for reproducibility.
+            output_type (`str`, defaults to `"pil"`):
+                `"pil"`, `"np"`, `"pt"`, or `"latent"`.
+            return_dict (`bool`, defaults to `True`):
+                Return [`ImagePipelineOutput`] if True.
+        """
+        default_size = DEFAULT_NATIVE_RESOLUTION if self.vae is not None else 256
+        height = int(height or default_size)
+        width = int(width or default_size)
+        self.check_inputs(height, width, num_inference_steps, output_type)
+
+        if getattr(self.scheduler.config, "stochastic_sampling", False):
+            raise ValueError(
+                "NiT expects deterministic FlowMatchEulerDiscreteScheduler stepping "
+                "(scheduler.config.stochastic_sampling=False). The scheduler's stochastic_sampling "
+                "path uses a different update rule than the official NiT Euler-Maruyama SDE and "
+                "produces salt-and-pepper noise."
+            )
+
+        device = self._execution_device
+        model_dtype = next(self.transformer.parameters()).dtype
+        class_labels_tensor = self._normalize_class_labels(class_labels)
+        batch_size = class_labels_tensor.numel()
+
+        packed_latents, image_sizes = self.prepare_latents(
+            batch_size, height, width, model_dtype, device, generator
+        )
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        num_train_timesteps = self.scheduler.config.num_train_timesteps
+
+        null_labels = torch.full_like(class_labels_tensor, self.transformer.config.num_classes)
+        guidance_low, guidance_high = guidance_interval
+
+        for t in self.progress_bar(self.scheduler.timesteps):
+            flow_time = self._flow_time_from_scheduler_timestep(t, num_train_timesteps)
+            guidance_active = guidance_low <= flow_time <= guidance_high
+            if guidance_scale > 1.0 and guidance_active:
+                model_input = torch.cat([packed_latents, packed_latents], dim=0)
+                labels = torch.cat([class_labels_tensor, null_labels], dim=0)
+                model_image_sizes = torch.cat([image_sizes, image_sizes], dim=0)
+            else:
+                model_input = packed_latents
+                labels = class_labels_tensor
+                model_image_sizes = image_sizes
+
+            timestep_batch = torch.full((labels.numel(),), flow_time, device=device, dtype=model_dtype)
+            model_output = self.transformer(
+                model_input.to(dtype=model_dtype),
+                timestep_batch,
+                labels,
+                image_sizes=model_image_sizes,
+                return_dict=True,
+            ).sample
+            model_output = self._apply_classifier_free_guidance(model_output, guidance_scale, guidance_active)
+            packed_latents = self.scheduler.step(
+                model_output,
+                t,
+                packed_latents,
+                generator=generator,
+            ).prev_sample
+
+        latents = self.transformer._unpack_latents(packed_latents, image_sizes)
+        image = self.decode_latents(latents, output_type=output_type)
+
+        self.maybe_free_model_hooks()
+        if not return_dict:
+            return (image,)
+        return ImagePipelineOutput(images=image)
+
+
+NiTPipelineOutput = ImagePipelineOutput

NiT-B/scheduler/scheduler_config.json

@@ -0,0 +1,7 @@
+{
+  "_class_name": "FlowMatchEulerDiscreteScheduler",
+  "_diffusers_version": "0.36.0",
+  "num_train_timesteps": 1000,
+  "shift": 1.0,
+  "stochastic_sampling": false
+}

NiT-B/transformer/config.json

@@ -0,0 +1,15 @@
+{
+  "_class_name": "NiTTransformer2DModel",
+  "class_dropout_prob": 0.1,
+  "depth": 12,
+  "encoder_depth": 4,
+  "hidden_size": 768,
+  "in_channels": 32,
+  "input_size": 32,
+  "num_classes": 1000,
+  "num_heads": 12,
+  "patch_size": 1,
+  "qk_norm": true,
+  "use_checkpoint": false,
+  "z_dim": 1280
+}

NiT-B/transformer/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72e838fb4dd0cdf52bd3b1407e97c8d7c2874fd527b4e9b51d7f724e98f8f35c
+size 554918480

NiT-B/transformer/nit_transformer_2d.py

@@ -0,0 +1,471 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+
+from dataclasses import dataclass
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    from diffusers.configuration_utils import ConfigMixin, register_to_config
+    from diffusers.models.modeling_utils import ModelMixin
+    from diffusers.utils import BaseOutput
+except Exception:  # pragma: no cover - lets this subtree be tested outside diffusers.
+    class BaseOutput(dict):
+        def __post_init__(self):
+            self.update(self.__dict__)
+
+    class _Config(dict):
+        def __getattr__(self, key):
+            try:
+                return self[key]
+            except KeyError as error:
+                raise AttributeError(key) from error
+
+    class ConfigMixin:
+        config_name = "config.json"
+
+    class ModelMixin(nn.Module):
+        pass
+
+    def register_to_config(init):
+        def wrapper(self, *args, **kwargs):
+            import inspect
+
+            signature = inspect.signature(init)
+            bound = signature.bind(self, *args, **kwargs)
+            bound.apply_defaults()
+            self.config = _Config({key: value for key, value in bound.arguments.items() if key != "self"})
+            init(self, *args, **kwargs)
+
+        return wrapper
+
+
+try:
+    from flash_attn import flash_attn_varlen_func
+except Exception:  # pragma: no cover - optional acceleration.
+    flash_attn_varlen_func = None
+
+
+@dataclass
+class NiTTransformer2DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+    projection_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+
+
+def _modulate(hidden_states: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
+    return hidden_states * (1 + scale) + shift
+
+
+def _rotate_half(hidden_states: torch.Tensor) -> torch.Tensor:
+    hidden_states = hidden_states.reshape(*hidden_states.shape[:-1], -1, 2)
+    hidden_states_1, hidden_states_2 = hidden_states.unbind(dim=-1)
+    return torch.stack((-hidden_states_2, hidden_states_1), dim=-1).flatten(-2)
+
+
+def _get_float_dtype_or_default(tensor: Optional[torch.Tensor] = None) -> torch.dtype:
+    if tensor is not None and tensor.is_floating_point():
+        return tensor.dtype
+    return torch.get_default_dtype()
+
+
+class NiTPatchEmbed(nn.Module):
+    def __init__(self, patch_size: int, in_channels: int, hidden_size: int):
+        super().__init__()
+        self.patch_size = (patch_size, patch_size)
+        self.proj = nn.Conv2d(in_channels, hidden_size, kernel_size=patch_size, stride=patch_size, bias=True)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.proj(hidden_states)
+        return hidden_states.flatten(2).transpose(1, 2)
+
+
+class NiTTimestepEmbedder(nn.Module):
+    def __init__(self, hidden_size: int, frequency_embedding_size: int = 256):
+        super().__init__()
+        self.frequency_embedding_size = frequency_embedding_size
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+
+    @staticmethod
+    def get_timestep_embedding(timesteps: torch.Tensor, embedding_dim: int, max_period: int = 10000):
+        half = embedding_dim // 2
+        # Keep sinusoid construction in fp32 to mirror native NiT behavior.
+        exponent = -math.log(max_period) * torch.arange(half, dtype=torch.float32, device=timesteps.device) / half
+        freqs = torch.exp(exponent)
+        args = timesteps.float()[:, None] * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if embedding_dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        timestep_freq = self.get_timestep_embedding(timesteps, self.frequency_embedding_size).to(timesteps.dtype)
+        return self.mlp(timestep_freq)
+
+
+class NiTLabelEmbedder(nn.Module):
+    def __init__(self, num_classes: int, hidden_size: int, dropout_prob: float):
+        super().__init__()
+        use_cfg_embedding = dropout_prob > 0
+        self.embedding_table = nn.Embedding(num_classes + int(use_cfg_embedding), hidden_size)
+        self.num_classes = num_classes
+        self.dropout_prob = dropout_prob
+
+    def forward(self, class_labels: torch.LongTensor) -> torch.Tensor:
+        return self.embedding_table(class_labels)
+
+
+class NiTRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        head_dim: int,
+        custom_freqs: str = "normal",
+        theta: int = 10000,
+        max_cached_len: int = 1024,
+        max_pe_len_h: Optional[int] = None,
+        max_pe_len_w: Optional[int] = None,
+        decouple: bool = False,
+        ori_max_pe_len: Optional[int] = None,
+    ):
+        super().__init__()
+        del max_pe_len_h, max_pe_len_w, decouple, ori_max_pe_len
+        if custom_freqs not in {"normal", "scale1", "scale2"}:
+            raise ValueError(
+                "This Diffusers implementation supports the trained RoPE frequencies directly. "
+                "Checkpoint conversion preserves weights; extrapolation variants should be handled "
+                "by changing the model config before loading."
+            )
+        dim = head_dim // 2
+        if dim % 2 != 0:
+            raise ValueError("NiT rotary embedding requires head_dim // 2 to be even.")
+        default_dtype = _get_float_dtype_or_default()
+        freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=default_dtype) / dim))
+        self.register_buffer("freqs_h", freqs, persistent=False)
+        self.register_buffer("freqs_w", freqs.clone(), persistent=False)
+        positions = torch.arange(max_cached_len, dtype=default_dtype)
+        freqs_h_cached = torch.einsum("n,f->nf", positions, self.freqs_h).repeat_interleave(2, dim=-1)
+        freqs_w_cached = torch.einsum("n,f->nf", positions, self.freqs_w).repeat_interleave(2, dim=-1)
+        self.register_buffer("freqs_h_cached", freqs_h_cached, persistent=False)
+        self.register_buffer("freqs_w_cached", freqs_w_cached, persistent=False)
+
+    def forward(self, image_sizes: torch.LongTensor, device: torch.device) -> Tuple[torch.Tensor, torch.Tensor]:
+        grids = []
+        for height, width in image_sizes.tolist():
+            # Match native NiT token ordering for RoPE alignment.
+            grid_h = torch.arange(height, device=device)
+            grid_w = torch.arange(width, device=device)
+            grid = torch.meshgrid(grid_h, grid_w, indexing="xy")
+            grids.append(torch.stack(grid, dim=0).reshape(2, -1))
+        grid = torch.cat(grids, dim=1)
+        freqs_h = self.freqs_h_cached.to(device)[grid[0]]
+        freqs_w = self.freqs_w_cached.to(device)[grid[1]]
+        freqs = torch.cat([freqs_h, freqs_w], dim=-1)
+        return freqs.cos().unsqueeze(1), freqs.sin().unsqueeze(1)
+
+
+class NiTAttention(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, qk_norm: bool = False):
+        super().__init__()
+        if hidden_size % num_heads != 0:
+            raise ValueError("hidden_size must be divisible by num_heads")
+        self.num_heads = num_heads
+        self.head_dim = hidden_size // num_heads
+        self.qkv = nn.Linear(hidden_size, hidden_size * 3, bias=True)
+        self.q_norm = nn.LayerNorm(self.head_dim) if qk_norm else nn.Identity()
+        self.k_norm = nn.LayerNorm(self.head_dim) if qk_norm else nn.Identity()
+        self.proj = nn.Linear(hidden_size, hidden_size)
+        self.proj_drop = nn.Dropout(0.0)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.IntTensor,
+        freqs_cos: torch.Tensor,
+        freqs_sin: torch.Tensor,
+    ) -> torch.Tensor:
+        qkv = self.qkv(hidden_states).reshape(hidden_states.shape[0], 3, self.num_heads, self.head_dim)
+        query, key, value = qkv.unbind(dim=1)
+        original_dtype = qkv.dtype
+        query = self.q_norm(query)
+        key = self.k_norm(key)
+        query = query * freqs_cos + _rotate_half(query) * freqs_sin
+        key = key * freqs_cos + _rotate_half(key) * freqs_sin
+        query = query.to(dtype=original_dtype)
+        key = key.to(dtype=original_dtype)
+
+        if flash_attn_varlen_func is not None and query.is_cuda:
+            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+            hidden_states = flash_attn_varlen_func(
+                query, key, value, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen
+            ).reshape(hidden_states.shape[0], -1)
+        else:
+            segments = []
+            for start, end in zip(cu_seqlens[:-1].tolist(), cu_seqlens[1:].tolist()):
+                q = query[start:end].transpose(0, 1).unsqueeze(0)
+                k = key[start:end].transpose(0, 1).unsqueeze(0)
+                v = value[start:end].transpose(0, 1).unsqueeze(0)
+                segments.append(F.scaled_dot_product_attention(q, k, v).squeeze(0).transpose(0, 1))
+            hidden_states = torch.cat(segments, dim=0).reshape(hidden_states.shape[0], -1)
+
+        hidden_states = self.proj(hidden_states)
+        return self.proj_drop(hidden_states)
+
+
+class NiTMLP(nn.Module):
+    def __init__(self, hidden_size: int, mlp_hidden_dim: int):
+        super().__init__()
+        self.fc1 = nn.Linear(hidden_size, mlp_hidden_dim)
+        self.act = nn.GELU(approximate="tanh")
+        self.drop1 = nn.Dropout(0.0)
+        self.norm = nn.Identity()
+        self.fc2 = nn.Linear(mlp_hidden_dim, hidden_size)
+        self.drop2 = nn.Dropout(0.0)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.drop1(hidden_states)
+        hidden_states = self.norm(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return self.drop2(hidden_states)
+
+
+class NiTBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_norm: bool = False,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 512,
+    ):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.attn = NiTAttention(hidden_size, num_heads=num_heads, qk_norm=qk_norm)
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.mlp = NiTMLP(hidden_size, mlp_hidden_dim)
+        if use_adaln_lora:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(hidden_size, adaln_lora_dim, bias=True),
+                nn.Linear(adaln_lora_dim, 6 * hidden_size, bias=True),
+            )
+        else:
+            self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True))
+
+    def forward(self, hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(conditioning).chunk(
+            6, dim=-1
+        )
+        hidden_states = hidden_states + gate_msa * self.attn(
+            _modulate(self.norm1(hidden_states), shift_msa, scale_msa), cu_seqlens, freqs_cos, freqs_sin
+        )
+        hidden_states = hidden_states + gate_mlp * self.mlp(
+            _modulate(self.norm2(hidden_states), shift_mlp, scale_mlp)
+        )
+        return hidden_states
+
+
+class NiTFinalLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
+
+    def forward(self, hidden_states: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
+        shift, scale = self.adaLN_modulation(conditioning).chunk(2, dim=-1)
+        hidden_states = _modulate(self.norm_final(hidden_states), shift, scale)
+        return self.linear(hidden_states)
+
+
+def _build_mlp(hidden_size: int, projector_dim: int, z_dim: int) -> nn.Sequential:
+    return nn.Sequential(
+        nn.Linear(hidden_size, projector_dim),
+        nn.SiLU(),
+        nn.Linear(projector_dim, projector_dim),
+        nn.SiLU(),
+        nn.Linear(projector_dim, z_dim),
+    )
+
+
+class NiTTransformer2DModel(ModelMixin, ConfigMixin):
+    config_name = "config.json"
+
+    @register_to_config
+    def __init__(
+        self,
+        input_size: int = 32,
+        patch_size: int = 1,
+        in_channels: int = 32,
+        hidden_size: int = 1152,
+        depth: int = 28,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.0,
+        class_dropout_prob: float = 0.1,
+        num_classes: int = 1000,
+        encoder_depth: int = 8,
+        projector_dim: int = 2048,
+        z_dim: int = 1280,
+        use_checkpoint: bool = False,
+        custom_freqs: str = "normal",
+        theta: int = 10000,
+        max_pe_len_h: Optional[int] = None,
+        max_pe_len_w: Optional[int] = None,
+        decouple: bool = False,
+        ori_max_pe_len: Optional[int] = None,
+        qk_norm: bool = True,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 512,
+    ):
+        super().__init__()
+        del input_size
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.num_classes = num_classes
+        self.encoder_depth = encoder_depth
+        self.use_checkpoint = use_checkpoint
+
+        self.x_embedder = NiTPatchEmbed(patch_size, in_channels, hidden_size)
+        self.t_embedder = NiTTimestepEmbedder(hidden_size)
+        self.y_embedder = NiTLabelEmbedder(num_classes, hidden_size, class_dropout_prob)
+        self.rope = NiTRotaryEmbedding(
+            hidden_size // num_heads,
+            custom_freqs=custom_freqs,
+            theta=theta,
+            max_pe_len_h=max_pe_len_h,
+            max_pe_len_w=max_pe_len_w,
+            decouple=decouple,
+            ori_max_pe_len=ori_max_pe_len,
+        )
+        self.projector = _build_mlp(hidden_size, projector_dim, z_dim)
+        self.blocks = nn.ModuleList(
+            [
+                NiTBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qk_norm=qk_norm,
+                    use_adaln_lora=use_adaln_lora,
+                    adaln_lora_dim=adaln_lora_dim,
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.final_layer = NiTFinalLayer(hidden_size, patch_size, self.out_channels)
+
+    def _pack_latents(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.LongTensor, Tuple[int, int]]:
+        batch_size, channels, height, width = hidden_states.shape
+        if channels != self.in_channels:
+            raise ValueError(f"Expected {self.in_channels} latent channels, got {channels}.")
+        if height % self.patch_size != 0 or width % self.patch_size != 0:
+            raise ValueError("Latent height and width must be divisible by patch_size.")
+        latent_h = height // self.patch_size
+        latent_w = width // self.patch_size
+        hidden_states = hidden_states.reshape(batch_size, channels, latent_h, self.patch_size, latent_w, self.patch_size)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5).reshape(
+            batch_size * latent_h * latent_w, channels, self.patch_size, self.patch_size
+        )
+        image_sizes = torch.tensor([[latent_h, latent_w]] * batch_size, device=hidden_states.device, dtype=torch.long)
+        return hidden_states, image_sizes, (height, width)
+
+    def _unpack_latents(self, hidden_states: torch.Tensor, image_sizes: torch.LongTensor) -> torch.Tensor:
+        if image_sizes.shape[0] == 1:
+            height, width = image_sizes[0].tolist()
+            hidden_states = hidden_states.reshape(height, width, self.out_channels, self.patch_size, self.patch_size)
+            return hidden_states.permute(2, 0, 3, 1, 4).reshape(
+                1, self.out_channels, height * self.patch_size, width * self.patch_size
+            )
+
+        samples = []
+        cursor = 0
+        for height, width in image_sizes.tolist():
+            length = height * width
+            sample = hidden_states[cursor : cursor + length].reshape(
+                height, width, self.out_channels, self.patch_size, self.patch_size
+            )
+            samples.append(
+                sample.permute(2, 0, 3, 1, 4).reshape(
+                    self.out_channels, height * self.patch_size, width * self.patch_size
+                )
+            )
+            cursor += length
+        if len({tuple(sample.shape) for sample in samples}) != 1:
+            return hidden_states
+        return torch.stack(samples, dim=0)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: Union[torch.Tensor, float],
+        class_labels: torch.LongTensor,
+        image_sizes: Optional[Union[torch.LongTensor, List[Tuple[int, int]]]] = None,
+        return_dict: bool = True,
+        output_projection_states: bool = False,
+    ) -> Union[NiTTransformer2DModelOutput, Tuple[torch.Tensor, ...]]:
+        input_was_image = hidden_states.dim() == 4 and image_sizes is None
+        if input_was_image:
+            hidden_states, image_sizes, _ = self._pack_latents(hidden_states)
+        elif image_sizes is None:
+            raise ValueError("image_sizes must be provided when hidden_states are already packed.")
+        elif not torch.is_tensor(image_sizes):
+            image_sizes = torch.tensor(image_sizes, device=hidden_states.device, dtype=torch.long)
+        else:
+            image_sizes = image_sizes.to(device=hidden_states.device, dtype=torch.long)
+
+        if not torch.is_tensor(timestep):
+            timestep = torch.tensor([timestep], device=hidden_states.device, dtype=hidden_states.dtype)
+        timestep = timestep.to(device=hidden_states.device, dtype=hidden_states.dtype).flatten()
+        if timestep.numel() == 1:
+            timestep = timestep.repeat(image_sizes.shape[0])
+        class_labels = class_labels.to(device=hidden_states.device, dtype=torch.long).flatten()
+
+        hidden_states = self.x_embedder(hidden_states).squeeze(1)
+        freqs_cos, freqs_sin = self.rope(image_sizes, hidden_states.device)
+
+        seqlens = image_sizes[:, 0] * image_sizes[:, 1]
+        cu_seqlens = torch.cat(
+            [torch.zeros(1, device=hidden_states.device, dtype=torch.int32), torch.cumsum(seqlens, dim=0).int()]
+        )
+
+        conditioning = self.t_embedder(timestep) + self.y_embedder(class_labels)
+        conditioning = torch.cat([conditioning[i].repeat(int(seqlens[i]), 1) for i in range(image_sizes.shape[0])], dim=0)
+
+        projection_states = []
+        for index, block in enumerate(self.blocks):
+            if self.use_checkpoint and self.training:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    block, hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin, use_reentrant=False
+                )
+            else:
+                hidden_states = block(hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin)
+            if output_projection_states and (index + 1) == self.encoder_depth:
+                projection_states.append(self.projector(hidden_states))
+
+        hidden_states = self.final_layer(hidden_states, conditioning)
+        hidden_states = hidden_states.reshape(hidden_states.shape[0], self.out_channels, self.patch_size, self.patch_size)
+        if input_was_image:
+            hidden_states = self._unpack_latents(hidden_states, image_sizes)
+
+        if not return_dict:
+            output = (hidden_states,)
+            if output_projection_states:
+                output = output + (tuple(projection_states),)
+            return output
+        return NiTTransformer2DModelOutput(
+            sample=hidden_states,
+            projection_states=tuple(projection_states) if output_projection_states else None,
+        )

NiT-B/vae/config.json

@@ -0,0 +1,88 @@
+{
+  "_class_name": "AutoencoderDC",
+  "_diffusers_version": "0.32.2",
+  "attention_head_dim": 32,
+  "decoder_act_fns": "silu",
+  "decoder_block_out_channels": [
+    128,
+    256,
+    512,
+    512,
+    1024,
+    1024
+  ],
+  "decoder_block_types": [
+    "ResBlock",
+    "ResBlock",
+    "ResBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock"
+  ],
+  "decoder_layers_per_block": [
+    3,
+    3,
+    3,
+    3,
+    3,
+    3
+  ],
+  "decoder_norm_types": "rms_norm",
+  "decoder_qkv_multiscales": [
+    [],
+    [],
+    [],
+    [
+      5
+    ],
+    [
+      5
+    ],
+    [
+      5
+    ]
+  ],
+  "downsample_block_type": "Conv",
+  "encoder_block_out_channels": [
+    128,
+    256,
+    512,
+    512,
+    1024,
+    1024
+  ],
+  "encoder_block_types": [
+    "ResBlock",
+    "ResBlock",
+    "ResBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock"
+  ],
+  "encoder_layers_per_block": [
+    2,
+    2,
+    2,
+    3,
+    3,
+    3
+  ],
+  "encoder_qkv_multiscales": [
+    [],
+    [],
+    [],
+    [
+      5
+    ],
+    [
+      5
+    ],
+    [
+      5
+    ]
+  ],
+  "in_channels": 3,
+  "latent_channels": 32,
+  "scaling_factor": 0.41407,
+  "upsample_block_type": "interpolate"
+}

NiT-B/vae/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfd991d1b54ffabf22745c5885589d8f2a7bc59930d95d92bd741c4fc64454bb
+size 1249044836

NiT-L/README.md

@@ -0,0 +1,90 @@
+---
+license: apache-2.0
+library_name: diffusers
+pipeline_tag: unconditional-image-generation
+tags:
+  - diffusers
+  - nit
+  - image-generation
+  - class-conditional
+  - imagenet
+inference: true
+---
+
+# NiT-L
+
+Self-contained Diffusers checkpoint for **NiT-L** (457M), converted from [`GoodEnough/NiT-L-Models`](https://huggingface.co/GoodEnough/NiT-L-Models) (`model_500K.safetensors`, 500K training steps).
+
+Architecture and training settings follow the official [`nit_l_pack_merge_radio_16384.yaml`](https://github.com/WZDTHU/NiT/blob/main/configs/c2i/nit_l_pack_merge_radio_16384.yaml).
+
+## Model config
+
+| Field | Value |
+| --- | --- |
+| Parameters | 457M |
+| Depth | 24 |
+| Hidden size | 1024 |
+| Attention heads | 16 |
+| Encoder depth | 6 |
+| Latent channels (`z_dim`) | 1280 |
+| Patch size | 1 |
+| Input latent channels | 32 |
+| Classes | 1000 |
+| Class dropout | 0.1 |
+| QK norm | true |
+| VAE | `mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers` |
+| Flow path type | linear |
+
+## Recommended inference (512×512)
+
+Official NiT sampling defaults for **512×512** class-conditional ImageNet generation:
+
+| Setting | Value |
+| --- | --- |
+| Resolution | 512×512 |
+| Solver | SDE (Euler–Maruyama) in the official repo |
+| Steps (NFE) | 250 |
+| CFG scale | 2.05 |
+| CFG interval | (0.0, 0.7) |
+
+This Diffusers port uses [`FlowMatchEulerDiscreteScheduler`](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete) in deterministic ODE mode (`stochastic_sampling=false`). Keep the same step count, CFG scale, and interval as the official recipe.
+
+## Usage
+
+```python
+from pathlib import Path
+import torch
+from diffusers import DiffusionPipeline
+
+model_dir = Path(".")
+pipe = DiffusionPipeline.from_pretrained(
+    str(model_dir),
+    local_files_only=True,
+    custom_pipeline=str(model_dir / "pipeline.py"),
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+).to("cuda" if torch.cuda.is_available() else "cpu")
+
+generator = torch.Generator(device=pipe.device).manual_seed(42)
+image = pipe(
+    class_labels="golden retriever",
+    height=512,
+    width=512,
+    num_inference_steps=250,
+    guidance_scale=2.05,
+    guidance_interval=(0.0, 0.7),
+    generator=generator,
+).images[0]
+image.save("demo_512.png")
+```
+
+## Components
+
+- `pipeline.py` — custom `NiTPipeline`
+- `model_index.json` — pipeline index + ImageNet `id2label`
+- `transformer/config.json`
+- `transformer/nit_transformer_2d.py`
+- `transformer/diffusion_pytorch_model.safetensors`
+- `scheduler/scheduler_config.json`
+- `vae/config.json`
+- `vae/diffusion_pytorch_model.safetensors`

NiT-L/model_index.json

@@ -0,0 +1,1021 @@
+{
+  "_class_name": [
+    "pipeline",
+    "NiTPipeline"
+  ],
+  "_diffusers_version": "0.36.0",
+  "scheduler": [
+    "diffusers",
+    "FlowMatchEulerDiscreteScheduler"
+  ],
+  "transformer": [
+    "nit_transformer_2d",
+    "NiTTransformer2DModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderDC"
+  ],
+  "id2label": {
+    "0": "tench, Tinca tinca",
+    "1": "goldfish, Carassius auratus",
+    "2": "great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias",
+    "3": "tiger shark, Galeocerdo cuvieri",
+    "4": "hammerhead, hammerhead shark",
+    "5": "electric ray, crampfish, numbfish, torpedo",
+    "6": "stingray",
+    "7": "cock",
+    "8": "hen",
+    "9": "ostrich, Struthio camelus",
+    "10": "brambling, Fringilla montifringilla",
+    "11": "goldfinch, Carduelis carduelis",
+    "12": "house finch, linnet, Carpodacus mexicanus",
+    "13": "junco, snowbird",
+    "14": "indigo bunting, indigo finch, indigo bird, Passerina cyanea",
+    "15": "robin, American robin, Turdus migratorius",
+    "16": "bulbul",
+    "17": "jay",
+    "18": "magpie",
+    "19": "chickadee",
+    "20": "water ouzel, dipper",
+    "21": "kite",
+    "22": "bald eagle, American eagle, Haliaeetus leucocephalus",
+    "23": "vulture",
+    "24": "great grey owl, great gray owl, Strix nebulosa",
+    "25": "European fire salamander, Salamandra salamandra",
+    "26": "common newt, Triturus vulgaris",
+    "27": "eft",
+    "28": "spotted salamander, Ambystoma maculatum",
+    "29": "axolotl, mud puppy, Ambystoma mexicanum",
+    "30": "bullfrog, Rana catesbeiana",
+    "31": "tree frog, tree-frog",
+    "32": "tailed frog, bell toad, ribbed toad, tailed toad, Ascaphus trui",
+    "33": "loggerhead, loggerhead turtle, Caretta caretta",
+    "34": "leatherback turtle, leatherback, leathery turtle, Dermochelys coriacea",
+    "35": "mud turtle",
+    "36": "terrapin",
+    "37": "box turtle, box tortoise",
+    "38": "banded gecko",
+    "39": "common iguana, iguana, Iguana iguana",
+    "40": "American chameleon, anole, Anolis carolinensis",
+    "41": "whiptail, whiptail lizard",
+    "42": "agama",
+    "43": "frilled lizard, Chlamydosaurus kingi",
+    "44": "alligator lizard",
+    "45": "Gila monster, Heloderma suspectum",
+    "46": "green lizard, Lacerta viridis",
+    "47": "African chameleon, Chamaeleo chamaeleon",
+    "48": "Komodo dragon, Komodo lizard, dragon lizard, giant lizard, Varanus komodoensis",
+    "49": "African crocodile, Nile crocodile, Crocodylus niloticus",
+    "50": "American alligator, Alligator mississipiensis",
+    "51": "triceratops",
+    "52": "thunder snake, worm snake, Carphophis amoenus",
+    "53": "ringneck snake, ring-necked snake, ring snake",
+    "54": "hognose snake, puff adder, sand viper",
+    "55": "green snake, grass snake",
+    "56": "king snake, kingsnake",
+    "57": "garter snake, grass snake",
+    "58": "water snake",
+    "59": "vine snake",
+    "60": "night snake, Hypsiglena torquata",
+    "61": "boa constrictor, Constrictor constrictor",
+    "62": "rock python, rock snake, Python sebae",
+    "63": "Indian cobra, Naja naja",
+    "64": "green mamba",
+    "65": "sea snake",
+    "66": "horned viper, cerastes, sand viper, horned asp, Cerastes cornutus",
+    "67": "diamondback, diamondback rattlesnake, Crotalus adamanteus",
+    "68": "sidewinder, horned rattlesnake, Crotalus cerastes",
+    "69": "trilobite",
+    "70": "harvestman, daddy longlegs, Phalangium opilio",
+    "71": "scorpion",
+    "72": "black and gold garden spider, Argiope aurantia",
+    "73": "barn spider, Araneus cavaticus",
+    "74": "garden spider, Aranea diademata",
+    "75": "black widow, Latrodectus mactans",
+    "76": "tarantula",
+    "77": "wolf spider, hunting spider",
+    "78": "tick",
+    "79": "centipede",
+    "80": "black grouse",
+    "81": "ptarmigan",
+    "82": "ruffed grouse, partridge, Bonasa umbellus",
+    "83": "prairie chicken, prairie grouse, prairie fowl",
+    "84": "peacock",
+    "85": "quail",
+    "86": "partridge",
+    "87": "African grey, African gray, Psittacus erithacus",
+    "88": "macaw",
+    "89": "sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita",
+    "90": "lorikeet",
+    "91": "coucal",
+    "92": "bee eater",
+    "93": "hornbill",
+    "94": "hummingbird",
+    "95": "jacamar",
+    "96": "toucan",
+    "97": "drake",
+    "98": "red-breasted merganser, Mergus serrator",
+    "99": "goose",
+    "100": "black swan, Cygnus atratus",
+    "101": "tusker",
+    "102": "echidna, spiny anteater, anteater",
+    "103": "platypus, duckbill, duckbilled platypus, duck-billed platypus, Ornithorhynchus anatinus",
+    "104": "wallaby, brush kangaroo",
+    "105": "koala, koala bear, kangaroo bear, native bear, Phascolarctos cinereus",
+    "106": "wombat",
+    "107": "jellyfish",
+    "108": "sea anemone, anemone",
+    "109": "brain coral",
+    "110": "flatworm, platyhelminth",
+    "111": "nematode, nematode worm, roundworm",
+    "112": "conch",
+    "113": "snail",
+    "114": "slug",
+    "115": "sea slug, nudibranch",
+    "116": "chiton, coat-of-mail shell, sea cradle, polyplacophore",
+    "117": "chambered nautilus, pearly nautilus, nautilus",
+    "118": "Dungeness crab, Cancer magister",
+    "119": "rock crab, Cancer irroratus",
+    "120": "fiddler crab",
+    "121": "king crab, Alaska crab, Alaskan king crab, Alaska king crab, Paralithodes camtschatica",
+    "122": "American lobster, Northern lobster, Maine lobster, Homarus americanus",
+    "123": "spiny lobster, langouste, rock lobster, crawfish, crayfish, sea crawfish",
+    "124": "crayfish, crawfish, crawdad, crawdaddy",
+    "125": "hermit crab",
+    "126": "isopod",
+    "127": "white stork, Ciconia ciconia",
+    "128": "black stork, Ciconia nigra",
+    "129": "spoonbill",
+    "130": "flamingo",
+    "131": "little blue heron, Egretta caerulea",
+    "132": "American egret, great white heron, Egretta albus",
+    "133": "bittern",
+    "134": "crane",
+    "135": "limpkin, Aramus pictus",
+    "136": "European gallinule, Porphyrio porphyrio",
+    "137": "American coot, marsh hen, mud hen, water hen, Fulica americana",
+    "138": "bustard",
+    "139": "ruddy turnstone, Arenaria interpres",
+    "140": "red-backed sandpiper, dunlin, Erolia alpina",
+    "141": "redshank, Tringa totanus",
+    "142": "dowitcher",
+    "143": "oystercatcher, oyster catcher",
+    "144": "pelican",
+    "145": "king penguin, Aptenodytes patagonica",
+    "146": "albatross, mollymawk",
+    "147": "grey whale, gray whale, devilfish, Eschrichtius gibbosus, Eschrichtius robustus",
+    "148": "killer whale, killer, orca, grampus, sea wolf, Orcinus orca",
+    "149": "dugong, Dugong dugon",
+    "150": "sea lion",
+    "151": "Chihuahua",
+    "152": "Japanese spaniel",
+    "153": "Maltese dog, Maltese terrier, Maltese",
+    "154": "Pekinese, Pekingese, Peke",
+    "155": "Shih-Tzu",
+    "156": "Blenheim spaniel",
+    "157": "papillon",
+    "158": "toy terrier",
+    "159": "Rhodesian ridgeback",
+    "160": "Afghan hound, Afghan",
+    "161": "basset, basset hound",
+    "162": "beagle",
+    "163": "bloodhound, sleuthhound",
+    "164": "bluetick",
+    "165": "black-and-tan coonhound",
+    "166": "Walker hound, Walker foxhound",
+    "167": "English foxhound",
+    "168": "redbone",
+    "169": "borzoi, Russian wolfhound",
+    "170": "Irish wolfhound",
+    "171": "Italian greyhound",
+    "172": "whippet",
+    "173": "Ibizan hound, Ibizan Podenco",
+    "174": "Norwegian elkhound, elkhound",
+    "175": "otterhound, otter hound",
+    "176": "Saluki, gazelle hound",
+    "177": "Scottish deerhound, deerhound",
+    "178": "Weimaraner",
+    "179": "Staffordshire bullterrier, Staffordshire bull terrier",
+    "180": "American Staffordshire terrier, Staffordshire terrier, American pit bull terrier, pit bull terrier",
+    "181": "Bedlington terrier",
+    "182": "Border terrier",
+    "183": "Kerry blue terrier",
+    "184": "Irish terrier",
+    "185": "Norfolk terrier",
+    "186": "Norwich terrier",
+    "187": "Yorkshire terrier",
+    "188": "wire-haired fox terrier",
+    "189": "Lakeland terrier",
+    "190": "Sealyham terrier, Sealyham",
+    "191": "Airedale, Airedale terrier",
+    "192": "cairn, cairn terrier",
+    "193": "Australian terrier",
+    "194": "Dandie Dinmont, Dandie Dinmont terrier",
+    "195": "Boston bull, Boston terrier",
+    "196": "miniature schnauzer",
+    "197": "giant schnauzer",
+    "198": "standard schnauzer",
+    "199": "Scotch terrier, Scottish terrier, Scottie",
+    "200": "Tibetan terrier, chrysanthemum dog",
+    "201": "silky terrier, Sydney silky",
+    "202": "soft-coated wheaten terrier",
+    "203": "West Highland white terrier",
+    "204": "Lhasa, Lhasa apso",
+    "205": "flat-coated retriever",
+    "206": "curly-coated retriever",
+    "207": "golden retriever",
+    "208": "Labrador retriever",
+    "209": "Chesapeake Bay retriever",
+    "210": "German short-haired pointer",
+    "211": "vizsla, Hungarian pointer",
+    "212": "English setter",
+    "213": "Irish setter, red setter",
+    "214": "Gordon setter",
+    "215": "Brittany spaniel",
+    "216": "clumber, clumber spaniel",
+    "217": "English springer, English springer spaniel",
+    "218": "Welsh springer spaniel",
+    "219": "cocker spaniel, English cocker spaniel, cocker",
+    "220": "Sussex spaniel",
+    "221": "Irish water spaniel",
+    "222": "kuvasz",
+    "223": "schipperke",
+    "224": "groenendael",
+    "225": "malinois",
+    "226": "briard",
+    "227": "kelpie",
+    "228": "komondor",
+    "229": "Old English sheepdog, bobtail",
+    "230": "Shetland sheepdog, Shetland sheep dog, Shetland",
+    "231": "collie",
+    "232": "Border collie",
+    "233": "Bouvier des Flandres, Bouviers des Flandres",
+    "234": "Rottweiler",
+    "235": "German shepherd, German shepherd dog, German police dog, alsatian",
+    "236": "Doberman, Doberman pinscher",
+    "237": "miniature pinscher",
+    "238": "Greater Swiss Mountain dog",
+    "239": "Bernese mountain dog",
+    "240": "Appenzeller",
+    "241": "EntleBucher",
+    "242": "boxer",
+    "243": "bull mastiff",
+    "244": "Tibetan mastiff",
+    "245": "French bulldog",
+    "246": "Great Dane",
+    "247": "Saint Bernard, St Bernard",
+    "248": "Eskimo dog, husky",
+    "249": "malamute, malemute, Alaskan malamute",
+    "250": "Siberian husky",
+    "251": "dalmatian, coach dog, carriage dog",
+    "252": "affenpinscher, monkey pinscher, monkey dog",
+    "253": "basenji",
+    "254": "pug, pug-dog",
+    "255": "Leonberg",
+    "256": "Newfoundland, Newfoundland dog",
+    "257": "Great Pyrenees",
+    "258": "Samoyed, Samoyede",
+    "259": "Pomeranian",
+    "260": "chow, chow chow",
+    "261": "keeshond",
+    "262": "Brabancon griffon",
+    "263": "Pembroke, Pembroke Welsh corgi",
+    "264": "Cardigan, Cardigan Welsh corgi",
+    "265": "toy poodle",
+    "266": "miniature poodle",
+    "267": "standard poodle",
+    "268": "Mexican hairless",
+    "269": "timber wolf, grey wolf, gray wolf, Canis lupus",
+    "270": "white wolf, Arctic wolf, Canis lupus tundrarum",
+    "271": "red wolf, maned wolf, Canis rufus, Canis niger",
+    "272": "coyote, prairie wolf, brush wolf, Canis latrans",
+    "273": "dingo, warrigal, warragal, Canis dingo",
+    "274": "dhole, Cuon alpinus",
+    "275": "African hunting dog, hyena dog, Cape hunting dog, Lycaon pictus",
+    "276": "hyena, hyaena",
+    "277": "red fox, Vulpes vulpes",
+    "278": "kit fox, Vulpes macrotis",
+    "279": "Arctic fox, white fox, Alopex lagopus",
+    "280": "grey fox, gray fox, Urocyon cinereoargenteus",
+    "281": "tabby, tabby cat",
+    "282": "tiger cat",
+    "283": "Persian cat",
+    "284": "Siamese cat, Siamese",
+    "285": "Egyptian cat",
+    "286": "cougar, puma, catamount, mountain lion, painter, panther, Felis concolor",
+    "287": "lynx, catamount",
+    "288": "leopard, Panthera pardus",
+    "289": "snow leopard, ounce, Panthera uncia",
+    "290": "jaguar, panther, Panthera onca, Felis onca",
+    "291": "lion, king of beasts, Panthera leo",
+    "292": "tiger, Panthera tigris",
+    "293": "cheetah, chetah, Acinonyx jubatus",
+    "294": "brown bear, bruin, Ursus arctos",
+    "295": "American black bear, black bear, Ursus americanus, Euarctos americanus",
+    "296": "ice bear, polar bear, Ursus Maritimus, Thalarctos maritimus",
+    "297": "sloth bear, Melursus ursinus, Ursus ursinus",
+    "298": "mongoose",
+    "299": "meerkat, mierkat",
+    "300": "tiger beetle",
+    "301": "ladybug, ladybeetle, lady beetle, ladybird, ladybird beetle",
+    "302": "ground beetle, carabid beetle",
+    "303": "long-horned beetle, longicorn, longicorn beetle",
+    "304": "leaf beetle, chrysomelid",
+    "305": "dung beetle",
+    "306": "rhinoceros beetle",
+    "307": "weevil",
+    "308": "fly",
+    "309": "bee",
+    "310": "ant, emmet, pismire",
+    "311": "grasshopper, hopper",
+    "312": "cricket",
+    "313": "walking stick, walkingstick, stick insect",
+    "314": "cockroach, roach",
+    "315": "mantis, mantid",
+    "316": "cicada, cicala",
+    "317": "leafhopper",
+    "318": "lacewing, lacewing fly",
+    "319": "dragonfly, darning needle, devils darning needle, sewing needle, snake feeder, snake doctor, mosquito hawk, skeeter hawk",
+    "320": "damselfly",
+    "321": "admiral",
+    "322": "ringlet, ringlet butterfly",
+    "323": "monarch, monarch butterfly, milkweed butterfly, Danaus plexippus",
+    "324": "cabbage butterfly",
+    "325": "sulphur butterfly, sulfur butterfly",
+    "326": "lycaenid, lycaenid butterfly",
+    "327": "starfish, sea star",
+    "328": "sea urchin",
+    "329": "sea cucumber, holothurian",
+    "330": "wood rabbit, cottontail, cottontail rabbit",
+    "331": "hare",
+    "332": "Angora, Angora rabbit",
+    "333": "hamster",
+    "334": "porcupine, hedgehog",
+    "335": "fox squirrel, eastern fox squirrel, Sciurus niger",
+    "336": "marmot",
+    "337": "beaver",
+    "338": "guinea pig, Cavia cobaya",
+    "339": "sorrel",
+    "340": "zebra",
+    "341": "hog, pig, grunter, squealer, Sus scrofa",
+    "342": "wild boar, boar, Sus scrofa",
+    "343": "warthog",
+    "344": "hippopotamus, hippo, river horse, Hippopotamus amphibius",
+    "345": "ox",
+    "346": "water buffalo, water ox, Asiatic buffalo, Bubalus bubalis",
+    "347": "bison",
+    "348": "ram, tup",
+    "349": "bighorn, bighorn sheep, cimarron, Rocky Mountain bighorn, Rocky Mountain sheep, Ovis canadensis",
+    "350": "ibex, Capra ibex",
+    "351": "hartebeest",
+    "352": "impala, Aepyceros melampus",
+    "353": "gazelle",
+    "354": "Arabian camel, dromedary, Camelus dromedarius",
+    "355": "llama",
+    "356": "weasel",
+    "357": "mink",
+    "358": "polecat, fitch, foulmart, foumart, Mustela putorius",
+    "359": "black-footed ferret, ferret, Mustela nigripes",
+    "360": "otter",
+    "361": "skunk, polecat, wood pussy",
+    "362": "badger",
+    "363": "armadillo",
+    "364": "three-toed sloth, ai, Bradypus tridactylus",
+    "365": "orangutan, orang, orangutang, Pongo pygmaeus",
+    "366": "gorilla, Gorilla gorilla",
+    "367": "chimpanzee, chimp, Pan troglodytes",
+    "368": "gibbon, Hylobates lar",
+    "369": "siamang, Hylobates syndactylus, Symphalangus syndactylus",
+    "370": "guenon, guenon monkey",
+    "371": "patas, hussar monkey, Erythrocebus patas",
+    "372": "baboon",
+    "373": "macaque",
+    "374": "langur",
+    "375": "colobus, colobus monkey",
+    "376": "proboscis monkey, Nasalis larvatus",
+    "377": "marmoset",
+    "378": "capuchin, ringtail, Cebus capucinus",
+    "379": "howler monkey, howler",
+    "380": "titi, titi monkey",
+    "381": "spider monkey, Ateles geoffroyi",
+    "382": "squirrel monkey, Saimiri sciureus",
+    "383": "Madagascar cat, ring-tailed lemur, Lemur catta",
+    "384": "indri, indris, Indri indri, Indri brevicaudatus",
+    "385": "Indian elephant, Elephas maximus",
+    "386": "African elephant, Loxodonta africana",
+    "387": "lesser panda, red panda, panda, bear cat, cat bear, Ailurus fulgens",
+    "388": "giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca",
+    "389": "barracouta, snoek",
+    "390": "eel",
+    "391": "coho, cohoe, coho salmon, blue jack, silver salmon, Oncorhynchus kisutch",
+    "392": "rock beauty, Holocanthus tricolor",
+    "393": "anemone fish",
+    "394": "sturgeon",
+    "395": "gar, garfish, garpike, billfish, Lepisosteus osseus",
+    "396": "lionfish",
+    "397": "puffer, pufferfish, blowfish, globefish",
+    "398": "abacus",
+    "399": "abaya",
+    "400": "academic gown, academic robe, judge robe",
+    "401": "accordion, piano accordion, squeeze box",
+    "402": "acoustic guitar",
+    "403": "aircraft carrier, carrier, flattop, attack aircraft carrier",
+    "404": "airliner",
+    "405": "airship, dirigible",
+    "406": "altar",
+    "407": "ambulance",
+    "408": "amphibian, amphibious vehicle",
+    "409": "analog clock",
+    "410": "apiary, bee house",
+    "411": "apron",
+    "412": "ashcan, trash can, garbage can, wastebin, ash bin, ash-bin, ashbin, dustbin, trash barrel, trash bin",
+    "413": "assault rifle, assault gun",
+    "414": "backpack, back pack, knapsack, packsack, rucksack, haversack",
+    "415": "bakery, bakeshop, bakehouse",
+    "416": "balance beam, beam",
+    "417": "balloon",
+    "418": "ballpoint, ballpoint pen, ballpen, Biro",
+    "419": "Band Aid",
+    "420": "banjo",
+    "421": "bannister, banister, balustrade, balusters, handrail",
+    "422": "barbell",
+    "423": "barber chair",
+    "424": "barbershop",
+    "425": "barn",
+    "426": "barometer",
+    "427": "barrel, cask",
+    "428": "barrow, garden cart, lawn cart, wheelbarrow",
+    "429": "baseball",
+    "430": "basketball",
+    "431": "bassinet",
+    "432": "bassoon",
+    "433": "bathing cap, swimming cap",
+    "434": "bath towel",
+    "435": "bathtub, bathing tub, bath, tub",
+    "436": "beach wagon, station wagon, wagon, estate car, beach waggon, station waggon, waggon",
+    "437": "beacon, lighthouse, beacon light, pharos",
+    "438": "beaker",
+    "439": "bearskin, busby, shako",
+    "440": "beer bottle",
+    "441": "beer glass",
+    "442": "bell cote, bell cot",
+    "443": "bib",
+    "444": "bicycle-built-for-two, tandem bicycle, tandem",
+    "445": "bikini, two-piece",
+    "446": "binder, ring-binder",
+    "447": "binoculars, field glasses, opera glasses",
+    "448": "birdhouse",
+    "449": "boathouse",
+    "450": "bobsled, bobsleigh, bob",
+    "451": "bolo tie, bolo, bola tie, bola",
+    "452": "bonnet, poke bonnet",
+    "453": "bookcase",
+    "454": "bookshop, bookstore, bookstall",
+    "455": "bottlecap",
+    "456": "bow",
+    "457": "bow tie, bow-tie, bowtie",
+    "458": "brass, memorial tablet, plaque",
+    "459": "brassiere, bra, bandeau",
+    "460": "breakwater, groin, groyne, mole, bulwark, seawall, jetty",
+    "461": "breastplate, aegis, egis",
+    "462": "broom",
+    "463": "bucket, pail",
+    "464": "buckle",
+    "465": "bulletproof vest",
+    "466": "bullet train, bullet",
+    "467": "butcher shop, meat market",
+    "468": "cab, hack, taxi, taxicab",
+    "469": "caldron, cauldron",
+    "470": "candle, taper, wax light",
+    "471": "cannon",
+    "472": "canoe",
+    "473": "can opener, tin opener",
+    "474": "cardigan",
+    "475": "car mirror",
+    "476": "carousel, carrousel, merry-go-round, roundabout, whirligig",
+    "477": "carpenters kit, tool kit",
+    "478": "carton",
+    "479": "car wheel",
+    "480": "cash machine, cash dispenser, automated teller machine, automatic teller machine, automated teller, automatic teller, ATM",
+    "481": "cassette",
+    "482": "cassette player",
+    "483": "castle",
+    "484": "catamaran",
+    "485": "CD player",
+    "486": "cello, violoncello",
+    "487": "cellular telephone, cellular phone, cellphone, cell, mobile phone",
+    "488": "chain",
+    "489": "chainlink fence",
+    "490": "chain mail, ring mail, mail, chain armor, chain armour, ring armor, ring armour",
+    "491": "chain saw, chainsaw",
+    "492": "chest",
+    "493": "chiffonier, commode",
+    "494": "chime, bell, gong",
+    "495": "china cabinet, china closet",
+    "496": "Christmas stocking",
+    "497": "church, church building",
+    "498": "cinema, movie theater, movie theatre, movie house, picture palace",
+    "499": "cleaver, meat cleaver, chopper",
+    "500": "cliff dwelling",
+    "501": "cloak",
+    "502": "clog, geta, patten, sabot",
+    "503": "cocktail shaker",
+    "504": "coffee mug",
+    "505": "coffeepot",
+    "506": "coil, spiral, volute, whorl, helix",
+    "507": "combination lock",
+    "508": "computer keyboard, keypad",
+    "509": "confectionery, confectionary, candy store",
+    "510": "container ship, containership, container vessel",
+    "511": "convertible",
+    "512": "corkscrew, bottle screw",
+    "513": "cornet, horn, trumpet, trump",
+    "514": "cowboy boot",
+    "515": "cowboy hat, ten-gallon hat",
+    "516": "cradle",
+    "517": "crane",
+    "518": "crash helmet",
+    "519": "crate",
+    "520": "crib, cot",
+    "521": "Crock Pot",
+    "522": "croquet ball",
+    "523": "crutch",
+    "524": "cuirass",
+    "525": "dam, dike, dyke",
+    "526": "desk",
+    "527": "desktop computer",
+    "528": "dial telephone, dial phone",
+    "529": "diaper, nappy, napkin",
+    "530": "digital clock",
+    "531": "digital watch",
+    "532": "dining table, board",
+    "533": "dishrag, dishcloth",
+    "534": "dishwasher, dish washer, dishwashing machine",
+    "535": "disk brake, disc brake",
+    "536": "dock, dockage, docking facility",
+    "537": "dogsled, dog sled, dog sleigh",
+    "538": "dome",
+    "539": "doormat, welcome mat",
+    "540": "drilling platform, offshore rig",
+    "541": "drum, membranophone, tympan",
+    "542": "drumstick",
+    "543": "dumbbell",
+    "544": "Dutch oven",
+    "545": "electric fan, blower",
+    "546": "electric guitar",
+    "547": "electric locomotive",
+    "548": "entertainment center",
+    "549": "envelope",
+    "550": "espresso maker",
+    "551": "face powder",
+    "552": "feather boa, boa",
+    "553": "file, file cabinet, filing cabinet",
+    "554": "fireboat",
+    "555": "fire engine, fire truck",
+    "556": "fire screen, fireguard",
+    "557": "flagpole, flagstaff",
+    "558": "flute, transverse flute",
+    "559": "folding chair",
+    "560": "football helmet",
+    "561": "forklift",
+    "562": "fountain",
+    "563": "fountain pen",
+    "564": "four-poster",
+    "565": "freight car",
+    "566": "French horn, horn",
+    "567": "frying pan, frypan, skillet",
+    "568": "fur coat",
+    "569": "garbage truck, dustcart",
+    "570": "gasmask, respirator, gas helmet",
+    "571": "gas pump, gasoline pump, petrol pump, island dispenser",
+    "572": "goblet",
+    "573": "go-kart",
+    "574": "golf ball",
+    "575": "golfcart, golf cart",
+    "576": "gondola",
+    "577": "gong, tam-tam",
+    "578": "gown",
+    "579": "grand piano, grand",
+    "580": "greenhouse, nursery, glasshouse",
+    "581": "grille, radiator grille",
+    "582": "grocery store, grocery, food market, market",
+    "583": "guillotine",
+    "584": "hair slide",
+    "585": "hair spray",
+    "586": "half track",
+    "587": "hammer",
+    "588": "hamper",
+    "589": "hand blower, blow dryer, blow drier, hair dryer, hair drier",
+    "590": "hand-held computer, hand-held microcomputer",
+    "591": "handkerchief, hankie, hanky, hankey",
+    "592": "hard disc, hard disk, fixed disk",
+    "593": "harmonica, mouth organ, harp, mouth harp",
+    "594": "harp",
+    "595": "harvester, reaper",
+    "596": "hatchet",
+    "597": "holster",
+    "598": "home theater, home theatre",
+    "599": "honeycomb",
+    "600": "hook, claw",
+    "601": "hoopskirt, crinoline",
+    "602": "horizontal bar, high bar",
+    "603": "horse cart, horse-cart",
+    "604": "hourglass",
+    "605": "iPod",
+    "606": "iron, smoothing iron",
+    "607": "jack-o-lantern",
+    "608": "jean, blue jean, denim",
+    "609": "jeep, landrover",
+    "610": "jersey, T-shirt, tee shirt",
+    "611": "jigsaw puzzle",
+    "612": "jinrikisha, ricksha, rickshaw",
+    "613": "joystick",
+    "614": "kimono",
+    "615": "knee pad",
+    "616": "knot",
+    "617": "lab coat, laboratory coat",
+    "618": "ladle",
+    "619": "lampshade, lamp shade",
+    "620": "laptop, laptop computer",
+    "621": "lawn mower, mower",
+    "622": "lens cap, lens cover",
+    "623": "letter opener, paper knife, paperknife",
+    "624": "library",
+    "625": "lifeboat",
+    "626": "lighter, light, igniter, ignitor",
+    "627": "limousine, limo",
+    "628": "liner, ocean liner",
+    "629": "lipstick, lip rouge",
+    "630": "Loafer",
+    "631": "lotion",
+    "632": "loudspeaker, speaker, speaker unit, loudspeaker system, speaker system",
+    "633": "loupe, jewelers loupe",
+    "634": "lumbermill, sawmill",
+    "635": "magnetic compass",
+    "636": "mailbag, postbag",
+    "637": "mailbox, letter box",
+    "638": "maillot",
+    "639": "maillot, tank suit",
+    "640": "manhole cover",
+    "641": "maraca",
+    "642": "marimba, xylophone",
+    "643": "mask",
+    "644": "matchstick",
+    "645": "maypole",
+    "646": "maze, labyrinth",
+    "647": "measuring cup",
+    "648": "medicine chest, medicine cabinet",
+    "649": "megalith, megalithic structure",
+    "650": "microphone, mike",
+    "651": "microwave, microwave oven",
+    "652": "military uniform",
+    "653": "milk can",
+    "654": "minibus",
+    "655": "miniskirt, mini",
+    "656": "minivan",
+    "657": "missile",
+    "658": "mitten",
+    "659": "mixing bowl",
+    "660": "mobile home, manufactured home",
+    "661": "Model T",
+    "662": "modem",
+    "663": "monastery",
+    "664": "monitor",
+    "665": "moped",
+    "666": "mortar",
+    "667": "mortarboard",
+    "668": "mosque",
+    "669": "mosquito net",
+    "670": "motor scooter, scooter",
+    "671": "mountain bike, all-terrain bike, off-roader",
+    "672": "mountain tent",
+    "673": "mouse, computer mouse",
+    "674": "mousetrap",
+    "675": "moving van",
+    "676": "muzzle",
+    "677": "nail",
+    "678": "neck brace",
+    "679": "necklace",
+    "680": "nipple",
+    "681": "notebook, notebook computer",
+    "682": "obelisk",
+    "683": "oboe, hautboy, hautbois",
+    "684": "ocarina, sweet potato",
+    "685": "odometer, hodometer, mileometer, milometer",
+    "686": "oil filter",
+    "687": "organ, pipe organ",
+    "688": "oscilloscope, scope, cathode-ray oscilloscope, CRO",
+    "689": "overskirt",
+    "690": "oxcart",
+    "691": "oxygen mask",
+    "692": "packet",
+    "693": "paddle, boat paddle",
+    "694": "paddlewheel, paddle wheel",
+    "695": "padlock",
+    "696": "paintbrush",
+    "697": "pajama, pyjama, pjs, jammies",
+    "698": "palace",
+    "699": "panpipe, pandean pipe, syrinx",
+    "700": "paper towel",
+    "701": "parachute, chute",
+    "702": "parallel bars, bars",
+    "703": "park bench",
+    "704": "parking meter",
+    "705": "passenger car, coach, carriage",
+    "706": "patio, terrace",
+    "707": "pay-phone, pay-station",
+    "708": "pedestal, plinth, footstall",
+    "709": "pencil box, pencil case",
+    "710": "pencil sharpener",
+    "711": "perfume, essence",
+    "712": "Petri dish",
+    "713": "photocopier",
+    "714": "pick, plectrum, plectron",
+    "715": "pickelhaube",
+    "716": "picket fence, paling",
+    "717": "pickup, pickup truck",
+    "718": "pier",
+    "719": "piggy bank, penny bank",
+    "720": "pill bottle",
+    "721": "pillow",
+    "722": "ping-pong ball",
+    "723": "pinwheel",
+    "724": "pirate, pirate ship",
+    "725": "pitcher, ewer",
+    "726": "plane, carpenters plane, woodworking plane",
+    "727": "planetarium",
+    "728": "plastic bag",
+    "729": "plate rack",
+    "730": "plow, plough",
+    "731": "plunger, plumbers helper",
+    "732": "Polaroid camera, Polaroid Land camera",
+    "733": "pole",
+    "734": "police van, police wagon, paddy wagon, patrol wagon, wagon, black Maria",
+    "735": "poncho",
+    "736": "pool table, billiard table, snooker table",
+    "737": "pop bottle, soda bottle",
+    "738": "pot, flowerpot",
+    "739": "potters wheel",
+    "740": "power drill",
+    "741": "prayer rug, prayer mat",
+    "742": "printer",
+    "743": "prison, prison house",
+    "744": "projectile, missile",
+    "745": "projector",
+    "746": "puck, hockey puck",
+    "747": "punching bag, punch bag, punching ball, punchball",
+    "748": "purse",
+    "749": "quill, quill pen",
+    "750": "quilt, comforter, comfort, puff",
+    "751": "racer, race car, racing car",
+    "752": "racket, racquet",
+    "753": "radiator",
+    "754": "radio, wireless",
+    "755": "radio telescope, radio reflector",
+    "756": "rain barrel",
+    "757": "recreational vehicle, RV, R.V.",
+    "758": "reel",
+    "759": "reflex camera",
+    "760": "refrigerator, icebox",
+    "761": "remote control, remote",
+    "762": "restaurant, eating house, eating place, eatery",
+    "763": "revolver, six-gun, six-shooter",
+    "764": "rifle",
+    "765": "rocking chair, rocker",
+    "766": "rotisserie",
+    "767": "rubber eraser, rubber, pencil eraser",
+    "768": "rugby ball",
+    "769": "rule, ruler",
+    "770": "running shoe",
+    "771": "safe",
+    "772": "safety pin",
+    "773": "saltshaker, salt shaker",
+    "774": "sandal",
+    "775": "sarong",
+    "776": "sax, saxophone",
+    "777": "scabbard",
+    "778": "scale, weighing machine",
+    "779": "school bus",
+    "780": "schooner",
+    "781": "scoreboard",
+    "782": "screen, CRT screen",
+    "783": "screw",
+    "784": "screwdriver",
+    "785": "seat belt, seatbelt",
+    "786": "sewing machine",
+    "787": "shield, buckler",
+    "788": "shoe shop, shoe-shop, shoe store",
+    "789": "shoji",
+    "790": "shopping basket",
+    "791": "shopping cart",
+    "792": "shovel",
+    "793": "shower cap",
+    "794": "shower curtain",
+    "795": "ski",
+    "796": "ski mask",
+    "797": "sleeping bag",
+    "798": "slide rule, slipstick",
+    "799": "sliding door",
+    "800": "slot, one-armed bandit",
+    "801": "snorkel",
+    "802": "snowmobile",
+    "803": "snowplow, snowplough",
+    "804": "soap dispenser",
+    "805": "soccer ball",
+    "806": "sock",
+    "807": "solar dish, solar collector, solar furnace",
+    "808": "sombrero",
+    "809": "soup bowl",
+    "810": "space bar",
+    "811": "space heater",
+    "812": "space shuttle",
+    "813": "spatula",
+    "814": "speedboat",
+    "815": "spider web, spiders web",
+    "816": "spindle",
+    "817": "sports car, sport car",
+    "818": "spotlight, spot",
+    "819": "stage",
+    "820": "steam locomotive",
+    "821": "steel arch bridge",
+    "822": "steel drum",
+    "823": "stethoscope",
+    "824": "stole",
+    "825": "stone wall",
+    "826": "stopwatch, stop watch",
+    "827": "stove",
+    "828": "strainer",
+    "829": "streetcar, tram, tramcar, trolley, trolley car",
+    "830": "stretcher",
+    "831": "studio couch, day bed",
+    "832": "stupa, tope",
+    "833": "submarine, pigboat, sub, U-boat",
+    "834": "suit, suit of clothes",
+    "835": "sundial",
+    "836": "sunglass",
+    "837": "sunglasses, dark glasses, shades",
+    "838": "sunscreen, sunblock, sun blocker",
+    "839": "suspension bridge",
+    "840": "swab, swob, mop",
+    "841": "sweatshirt",
+    "842": "swimming trunks, bathing trunks",
+    "843": "swing",
+    "844": "switch, electric switch, electrical switch",
+    "845": "syringe",
+    "846": "table lamp",
+    "847": "tank, army tank, armored combat vehicle, armoured combat vehicle",
+    "848": "tape player",
+    "849": "teapot",
+    "850": "teddy, teddy bear",
+    "851": "television, television system",
+    "852": "tennis ball",
+    "853": "thatch, thatched roof",
+    "854": "theater curtain, theatre curtain",
+    "855": "thimble",
+    "856": "thresher, thrasher, threshing machine",
+    "857": "throne",
+    "858": "tile roof",
+    "859": "toaster",
+    "860": "tobacco shop, tobacconist shop, tobacconist",
+    "861": "toilet seat",
+    "862": "torch",
+    "863": "totem pole",
+    "864": "tow truck, tow car, wrecker",
+    "865": "toyshop",
+    "866": "tractor",
+    "867": "trailer truck, tractor trailer, trucking rig, rig, articulated lorry, semi",
+    "868": "tray",
+    "869": "trench coat",
+    "870": "tricycle, trike, velocipede",
+    "871": "trimaran",
+    "872": "tripod",
+    "873": "triumphal arch",
+    "874": "trolleybus, trolley coach, trackless trolley",
+    "875": "trombone",
+    "876": "tub, vat",
+    "877": "turnstile",
+    "878": "typewriter keyboard",
+    "879": "umbrella",
+    "880": "unicycle, monocycle",
+    "881": "upright, upright piano",
+    "882": "vacuum, vacuum cleaner",
+    "883": "vase",
+    "884": "vault",
+    "885": "velvet",
+    "886": "vending machine",
+    "887": "vestment",
+    "888": "viaduct",
+    "889": "violin, fiddle",
+    "890": "volleyball",
+    "891": "waffle iron",
+    "892": "wall clock",
+    "893": "wallet, billfold, notecase, pocketbook",
+    "894": "wardrobe, closet, press",
+    "895": "warplane, military plane",
+    "896": "washbasin, handbasin, washbowl, lavabo, wash-hand basin",
+    "897": "washer, automatic washer, washing machine",
+    "898": "water bottle",
+    "899": "water jug",
+    "900": "water tower",
+    "901": "whiskey jug",
+    "902": "whistle",
+    "903": "wig",
+    "904": "window screen",
+    "905": "window shade",
+    "906": "Windsor tie",
+    "907": "wine bottle",
+    "908": "wing",
+    "909": "wok",
+    "910": "wooden spoon",
+    "911": "wool, woolen, woollen",
+    "912": "worm fence, snake fence, snake-rail fence, Virginia fence",
+    "913": "wreck",
+    "914": "yawl",
+    "915": "yurt",
+    "916": "web site, website, internet site, site",
+    "917": "comic book",
+    "918": "crossword puzzle, crossword",
+    "919": "street sign",
+    "920": "traffic light, traffic signal, stoplight",
+    "921": "book jacket, dust cover, dust jacket, dust wrapper",
+    "922": "menu",
+    "923": "plate",
+    "924": "guacamole",
+    "925": "consomme",
+    "926": "hot pot, hotpot",
+    "927": "trifle",
+    "928": "ice cream, icecream",
+    "929": "ice lolly, lolly, lollipop, popsicle",
+    "930": "French loaf",
+    "931": "bagel, beigel",
+    "932": "pretzel",
+    "933": "cheeseburger",
+    "934": "hotdog, hot dog, red hot",
+    "935": "mashed potato",
+    "936": "head cabbage",
+    "937": "broccoli",
+    "938": "cauliflower",
+    "939": "zucchini, courgette",
+    "940": "spaghetti squash",
+    "941": "acorn squash",
+    "942": "butternut squash",
+    "943": "cucumber, cuke",
+    "944": "artichoke, globe artichoke",
+    "945": "bell pepper",
+    "946": "cardoon",
+    "947": "mushroom",
+    "948": "Granny Smith",
+    "949": "strawberry",
+    "950": "orange",
+    "951": "lemon",
+    "952": "fig",
+    "953": "pineapple, ananas",
+    "954": "banana",
+    "955": "jackfruit, jak, jack",
+    "956": "custard apple",
+    "957": "pomegranate",
+    "958": "hay",
+    "959": "carbonara",
+    "960": "chocolate sauce, chocolate syrup",
+    "961": "dough",
+    "962": "meat loaf, meatloaf",
+    "963": "pizza, pizza pie",
+    "964": "potpie",
+    "965": "burrito",
+    "966": "red wine",
+    "967": "espresso",
+    "968": "cup",
+    "969": "eggnog",
+    "970": "alp",
+    "971": "bubble",
+    "972": "cliff, drop, drop-off",
+    "973": "coral reef",
+    "974": "geyser",
+    "975": "lakeside, lakeshore",
+    "976": "promontory, headland, head, foreland",
+    "977": "sandbar, sand bar",
+    "978": "seashore, coast, seacoast, sea-coast",
+    "979": "valley, vale",
+    "980": "volcano",
+    "981": "ballplayer, baseball player",
+    "982": "groom, bridegroom",
+    "983": "scuba diver",
+    "984": "rapeseed",
+    "985": "daisy",
+    "986": "yellow ladys slipper, yellow lady-slipper, Cypripedium calceolus, Cypripedium parviflorum",
+    "987": "corn",
+    "988": "acorn",
+    "989": "hip, rose hip, rosehip",
+    "990": "buckeye, horse chestnut, conker",
+    "991": "coral fungus",
+    "992": "agaric",
+    "993": "gyromitra",
+    "994": "stinkhorn, carrion fungus",
+    "995": "earthstar",
+    "996": "hen-of-the-woods, hen of the woods, Polyporus frondosus, Grifola frondosa",
+    "997": "bolete",
+    "998": "ear, spike, capitulum",
+    "999": "toilet tissue, toilet paper, bathroom tissue"
+  }
+}

NiT-L/pipeline.py

@@ -0,0 +1,483 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils.torch_utils import randn_tensor
+
+# Local component classes are loaded dynamically in from_pretrained.
+
+DEFAULT_NATIVE_RESOLUTION = 512
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> from pathlib import Path
+        >>> import torch
+        >>> from diffusers import DiffusionPipeline
+
+        >>> model_dir = Path("./NiT-L").resolve()
+        >>> pipe = DiffusionPipeline.from_pretrained(
+        ...     str(model_dir),
+        ...     local_files_only=True,
+        ...     custom_pipeline=str(model_dir / "pipeline.py"),
+        ...     trust_remote_code=True,
+        ...     torch_dtype=torch.bfloat16,
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> print(pipe.id2label[207])
+        >>> print(pipe.get_label_ids("golden retriever"))
+
+        >>> generator = torch.Generator(device="cuda").manual_seed(42)
+        >>> image = pipe(
+        ...     class_labels="golden retriever",
+        ...     height=512,
+        ...     width=512,
+        ...     num_inference_steps=250,
+        ...     guidance_scale=2.05,
+        ...     guidance_interval=(0.0, 0.7),
+        ...     generator=generator,
+        ... ).images[0]
+        >>> image.save("demo.png")
+        ```
+"""
+
+
+class NiTPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for native-resolution class-conditional image generation with NiT.
+
+    Uses the native [`FlowMatchEulerDiscreteScheduler`] in deterministic (ODE) mode.
+    The official NiT repo defaults to an Euler-Maruyama SDE sampler for 512×512; that SDE is
+    not the same as the scheduler's `stochastic_sampling` path, so keep
+    `scheduler.config.stochastic_sampling=False` and let the scheduler perform the ODE update
+    `x_{t+dt} = x_t + dt * v`.
+
+    Parameters:
+        transformer ([`NiTTransformer2DModel`]):
+            Class-conditional transformer that predicts flow-matching velocity in packed latent space.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            Native diffusers flow-matching Euler scheduler (`stochastic_sampling=False`).
+        vae ([`AutoencoderDC`] or [`AutoencoderKL`], *optional*):
+            Variational autoencoder used to decode packed transformer latents to pixels.
+        id2label (`dict[int, str]`, *optional*):
+            ImageNet class id to English label mapping. Values may contain comma-separated synonyms.
+    """
+
+    model_cpu_offload_seq = "transformer->vae"
+    _optional_components = ["vae"]
+
+    def __init__(
+        self,
+        transformer,
+        scheduler,
+        vae=None,
+        id2label: Optional[Dict[Union[int, str], str]] = None,
+    ):
+        super().__init__()
+        self.register_modules(transformer=transformer, scheduler=scheduler, vae=vae)
+        self.image_processor = VaeImageProcessor()
+        self._id2label = self._normalize_id2label(id2label)
+        self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = bool(self._id2label)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path=None, subfolder=None, **kwargs):
+        """Load a self-contained variant folder locally or from the Hub."""
+        import importlib
+        import sys
+
+        repo_root = Path(__file__).resolve().parent
+
+        if pretrained_model_name_or_path in (None, "", "."):
+            variant = repo_root
+        elif (
+            isinstance(pretrained_model_name_or_path, str)
+            and "/" in pretrained_model_name_or_path
+            and not Path(pretrained_model_name_or_path).exists()
+        ):
+            from huggingface_hub import snapshot_download
+
+            hub_kwargs = dict(kwargs.pop("hub_kwargs", {}))
+            if subfolder:
+                hub_kwargs.setdefault("allow_patterns", [f"{subfolder}/**"])
+            cache_dir = snapshot_download(pretrained_model_name_or_path, **hub_kwargs)
+            variant = Path(cache_dir) / subfolder if subfolder else Path(cache_dir)
+        else:
+            variant = Path(pretrained_model_name_or_path)
+            if not variant.is_absolute():
+                candidate = (Path.cwd() / variant).resolve()
+                variant = candidate if candidate.exists() else (repo_root / variant).resolve()
+            if subfolder:
+                variant = variant / subfolder
+
+        id2label_override = kwargs.pop("id2label", None)
+        model_kwargs = dict(kwargs)
+        inserted: List[str] = []
+
+        def _load_component(folder: str, module_name: str, class_name: str):
+            comp_dir = variant / folder
+            module_path = comp_dir / f"{module_name}.py"
+            has_weights = (comp_dir / "config.json").exists() or (comp_dir / "scheduler_config.json").exists()
+            if not module_path.exists() or not has_weights:
+                return None
+
+            comp_path = str(comp_dir)
+            if comp_path not in sys.path:
+                sys.path.insert(0, comp_path)
+                inserted.append(comp_path)
+
+            module = importlib.import_module(module_name)
+            component_cls = getattr(module, class_name)
+            return component_cls.from_pretrained(str(comp_dir), **model_kwargs)
+
+        try:
+            transformer = _load_component("transformer", "nit_transformer_2d", "NiTTransformer2DModel")
+            try:
+                scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(str(variant), subfolder="scheduler")
+            except Exception:
+                scheduler = FlowMatchEulerDiscreteScheduler(
+                    num_train_timesteps=1000,
+                    shift=1.0,
+                    stochastic_sampling=False,
+                )
+            if transformer is None:
+                raise ValueError(f"No loadable transformer found under {variant}")
+
+            vae = None
+            vae_dir = variant / "vae"
+            if vae_dir.exists() and (vae_dir / "config.json").exists():
+                from diffusers import AutoencoderDC, AutoencoderKL
+
+                vae_class_name = json.loads((vae_dir / "config.json").read_text(encoding="utf-8")).get(
+                    "_class_name", "AutoencoderDC"
+                )
+                vae_cls = AutoencoderDC if vae_class_name == "AutoencoderDC" else AutoencoderKL
+                vae = vae_cls.from_pretrained(str(vae_dir), **model_kwargs)
+
+            id2label = id2label_override or cls._read_id2label_from_model_index(str(variant))
+            pipe = cls(
+                transformer=transformer,
+                scheduler=scheduler,
+                vae=vae,
+                id2label=id2label,
+            )
+            if hasattr(pipe, "register_to_config"):
+                pipe.register_to_config(_name_or_path=str(variant))
+            return pipe
+        finally:
+            for comp_path in inserted:
+                if comp_path in sys.path:
+                    sys.path.remove(comp_path)
+
+    def _ensure_labels_loaded(self) -> None:
+        if self._labels_loaded_from_model_index:
+            return
+        loaded = self._read_id2label_from_model_index(getattr(self.config, "_name_or_path", None))
+        if loaded:
+            self._id2label = loaded
+            self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = True
+
+    @staticmethod
+    def _normalize_id2label(id2label: Optional[Dict[Union[int, str], str]]) -> Dict[int, str]:
+        if not id2label:
+            return {}
+        return {int(key): value for key, value in id2label.items()}
+
+    @staticmethod
+    def _read_id2label_from_model_index(variant_path: Optional[str]) -> Dict[int, str]:
+        if not variant_path:
+            return {}
+        variant_dir = Path(variant_path).resolve()
+        model_index_path = variant_dir / "model_index.json"
+        if not model_index_path.exists():
+            return {}
+        raw = json.loads(model_index_path.read_text(encoding="utf-8"))
+        id2label = raw.get("id2label")
+        if not isinstance(id2label, dict):
+            return {}
+        return {int(key): value for key, value in id2label.items()}
+
+    @staticmethod
+    def _build_label2id(id2label: Dict[int, str]) -> Dict[str, int]:
+        label2id: Dict[str, int] = {}
+        for class_id, value in id2label.items():
+            for synonym in value.split(","):
+                synonym = synonym.strip()
+                if synonym:
+                    label2id[synonym] = int(class_id)
+        return dict(sorted(label2id.items()))
+
+    @property
+    def id2label(self) -> Dict[int, str]:
+        r"""ImageNet class id to English label string (comma-separated synonyms)."""
+        self._ensure_labels_loaded()
+        return self._id2label
+
+    def get_label_ids(self, label: Union[str, List[str]]) -> List[int]:
+        r"""
+        Map ImageNet label strings to class ids.
+
+        Args:
+            label (`str` or `list[str]`):
+                One or more English label strings. Each string must match a synonym in `id2label`.
+        """
+        self._ensure_labels_loaded()
+        label2id = self.labels
+        if not label2id:
+            raise ValueError("No English labels loaded. Ensure `id2label` exists in model_index.json.")
+
+        if isinstance(label, str):
+            label = [label]
+
+        missing = [item for item in label if item not in label2id]
+        if missing:
+            preview = ", ".join(list(label2id.keys())[:8])
+            raise ValueError(f"Unknown English label(s): {missing}. Example valid labels: {preview}, ...")
+        return [label2id[item] for item in label]
+
+    def _normalize_class_labels(
+        self,
+        class_labels: Union[int, str, List[Union[int, str]], torch.LongTensor],
+    ) -> torch.LongTensor:
+        if torch.is_tensor(class_labels):
+            return class_labels.to(device=self._execution_device, dtype=torch.long).reshape(-1)
+
+        if isinstance(class_labels, int):
+            class_label_ids = [class_labels]
+        elif isinstance(class_labels, str):
+            class_label_ids = self.get_label_ids(class_labels)
+        elif class_labels and isinstance(class_labels[0], str):
+            class_label_ids = self.get_label_ids(class_labels)
+        else:
+            class_label_ids = list(class_labels)
+
+        return torch.tensor(class_label_ids, device=self._execution_device, dtype=torch.long).reshape(-1)
+
+    def _get_vae_spatial_downsample(self) -> int:
+        if self.vae is None:
+            return 1
+        if self.vae.__class__.__name__ == "AutoencoderDC" or "dc-ae" in getattr(
+            self.vae.config, "_name_or_path", ""
+        ):
+            return 32
+        block_out_channels = getattr(self.vae.config, "block_out_channels", [0, 0, 0, 0])
+        return 2 ** (len(block_out_channels) - 1)
+
+    def check_inputs(
+        self,
+        height: int,
+        width: int,
+        num_inference_steps: int,
+        output_type: str,
+    ) -> None:
+        if num_inference_steps < 1:
+            raise ValueError("num_inference_steps must be >= 1.")
+        if output_type not in {"pil", "np", "pt", "latent"}:
+            raise ValueError("output_type must be one of: 'pil', 'np', 'pt', 'latent'.")
+
+        spatial_downsample = self._get_vae_spatial_downsample()
+        if height % spatial_downsample != 0 or width % spatial_downsample != 0:
+            raise ValueError(
+                f"height and width must be divisible by the VAE downsample factor {spatial_downsample}."
+            )
+
+        patch_size = int(self.transformer.config.patch_size)
+        latent_height = height // spatial_downsample
+        latent_width = width // spatial_downsample
+        if latent_height % patch_size != 0 or latent_width % patch_size != 0:
+            raise ValueError("Latent height and width must be divisible by transformer's patch_size.")
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]],
+    ) -> Tuple[torch.Tensor, torch.LongTensor]:
+        spatial_downsample = self._get_vae_spatial_downsample()
+        latent_height = height // spatial_downsample
+        latent_width = width // spatial_downsample
+        patch_size = int(self.transformer.config.patch_size)
+        token_height = latent_height // patch_size
+        token_width = latent_width // patch_size
+        image_sizes = torch.tensor([[token_height, token_width]] * batch_size, device=device, dtype=torch.long)
+
+        packed_shape = (
+            batch_size * token_height * token_width,
+            self.transformer.config.in_channels,
+            patch_size,
+            patch_size,
+        )
+        packed_latents = randn_tensor(
+            packed_shape,
+            generator=generator,
+            device=device,
+            dtype=dtype,
+        )
+        return packed_latents, image_sizes
+
+    @staticmethod
+    def _flow_time_from_scheduler_timestep(timestep: torch.Tensor, num_train_timesteps: int) -> float:
+        """Map native scheduler timesteps (sigma * num_train_timesteps) to NiT flow time in [0, 1]."""
+        return float(timestep) / num_train_timesteps
+
+    def _apply_classifier_free_guidance(
+        self,
+        model_output: torch.Tensor,
+        guidance_scale: float,
+        guidance_active: bool,
+    ) -> torch.Tensor:
+        if guidance_scale <= 1.0 or not guidance_active:
+            return model_output
+        model_output_cond, model_output_uncond = model_output.chunk(2)
+        return model_output_uncond + guidance_scale * (model_output_cond - model_output_uncond)
+
+    def _get_vae_dtype(self, latents: torch.Tensor) -> torch.dtype:
+        vae_dtype = getattr(self.vae, "dtype", None)
+        if vae_dtype is not None:
+            return vae_dtype
+        vae_params = next(self.vae.parameters(), None)
+        return vae_params.dtype if vae_params is not None else latents.dtype
+
+    def decode_latents(self, latents: torch.Tensor, output_type: str = "pil"):
+        if self.vae is None:
+            if output_type == "latent":
+                return latents
+            raise ValueError("Cannot decode latents without a VAE.")
+
+        vae_dtype = self._get_vae_dtype(latents)
+        latents = latents.to(dtype=vae_dtype)
+        scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
+        shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
+        latents = (latents / scaling_factor) + shift_factor
+        if output_type == "latent":
+            return latents
+
+        image = self.vae.decode(latents, return_dict=False)[0]
+        return self.image_processor.postprocess(image, output_type=output_type)
+
+    @torch.inference_mode()
+    def __call__(
+        self,
+        class_labels: Union[int, str, List[Union[int, str]], torch.LongTensor],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 1.0,
+        guidance_interval: Tuple[float, float] = (0.0, 1.0),
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: str = "pil",
+        return_dict: bool = True,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        r"""
+        Generate class-conditional images at native resolution.
+
+        Args:
+            class_labels (`int`, `str`, `list[int]`, `list[str]`, or `torch.LongTensor`):
+                ImageNet class indices or human-readable English label strings.
+            height (`int`, *optional*):
+                Output image height in pixels. Defaults to `512` when a VAE is present.
+            width (`int`, *optional*):
+                Output image width in pixels. Defaults to `512` when a VAE is present.
+            num_inference_steps (`int`, defaults to `50`):
+                Number of denoising steps.
+            guidance_scale (`float`, defaults to `1.0`):
+                Classifier-free guidance scale. CFG is active when `guidance_scale > 1.0`.
+            guidance_interval (`tuple[float, float]`, defaults to `(0.0, 1.0)`):
+                Flow-time interval where CFG is applied. Uses continuous flow time
+                `timestep / num_train_timesteps`, matching the official NiT ODE sampler.
+            generator (`torch.Generator`, *optional*):
+                RNG for reproducibility.
+            output_type (`str`, defaults to `"pil"`):
+                `"pil"`, `"np"`, `"pt"`, or `"latent"`.
+            return_dict (`bool`, defaults to `True`):
+                Return [`ImagePipelineOutput`] if True.
+        """
+        default_size = DEFAULT_NATIVE_RESOLUTION if self.vae is not None else 256
+        height = int(height or default_size)
+        width = int(width or default_size)
+        self.check_inputs(height, width, num_inference_steps, output_type)
+
+        if getattr(self.scheduler.config, "stochastic_sampling", False):
+            raise ValueError(
+                "NiT expects deterministic FlowMatchEulerDiscreteScheduler stepping "
+                "(scheduler.config.stochastic_sampling=False). The scheduler's stochastic_sampling "
+                "path uses a different update rule than the official NiT Euler-Maruyama SDE and "
+                "produces salt-and-pepper noise."
+            )
+
+        device = self._execution_device
+        model_dtype = next(self.transformer.parameters()).dtype
+        class_labels_tensor = self._normalize_class_labels(class_labels)
+        batch_size = class_labels_tensor.numel()
+
+        packed_latents, image_sizes = self.prepare_latents(
+            batch_size, height, width, model_dtype, device, generator
+        )
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        num_train_timesteps = self.scheduler.config.num_train_timesteps
+
+        null_labels = torch.full_like(class_labels_tensor, self.transformer.config.num_classes)
+        guidance_low, guidance_high = guidance_interval
+
+        for t in self.progress_bar(self.scheduler.timesteps):
+            flow_time = self._flow_time_from_scheduler_timestep(t, num_train_timesteps)
+            guidance_active = guidance_low <= flow_time <= guidance_high
+            if guidance_scale > 1.0 and guidance_active:
+                model_input = torch.cat([packed_latents, packed_latents], dim=0)
+                labels = torch.cat([class_labels_tensor, null_labels], dim=0)
+                model_image_sizes = torch.cat([image_sizes, image_sizes], dim=0)
+            else:
+                model_input = packed_latents
+                labels = class_labels_tensor
+                model_image_sizes = image_sizes
+
+            timestep_batch = torch.full((labels.numel(),), flow_time, device=device, dtype=model_dtype)
+            model_output = self.transformer(
+                model_input.to(dtype=model_dtype),
+                timestep_batch,
+                labels,
+                image_sizes=model_image_sizes,
+                return_dict=True,
+            ).sample
+            model_output = self._apply_classifier_free_guidance(model_output, guidance_scale, guidance_active)
+            packed_latents = self.scheduler.step(
+                model_output,
+                t,
+                packed_latents,
+                generator=generator,
+            ).prev_sample
+
+        latents = self.transformer._unpack_latents(packed_latents, image_sizes)
+        image = self.decode_latents(latents, output_type=output_type)
+
+        self.maybe_free_model_hooks()
+        if not return_dict:
+            return (image,)
+        return ImagePipelineOutput(images=image)
+
+
+NiTPipelineOutput = ImagePipelineOutput

NiT-L/scheduler/scheduler_config.json

@@ -0,0 +1,7 @@
+{
+  "_class_name": "FlowMatchEulerDiscreteScheduler",
+  "_diffusers_version": "0.36.0",
+  "num_train_timesteps": 1000,
+  "shift": 1.0,
+  "stochastic_sampling": false
+}

NiT-L/transformer/config.json

@@ -0,0 +1,15 @@
+{
+  "_class_name": "NiTTransformer2DModel",
+  "class_dropout_prob": 0.1,
+  "depth": 24,
+  "encoder_depth": 6,
+  "hidden_size": 1024,
+  "in_channels": 32,
+  "input_size": 32,
+  "num_classes": 1000,
+  "num_heads": 16,
+  "patch_size": 1,
+  "qk_norm": true,
+  "use_checkpoint": false,
+  "z_dim": 1280
+}

NiT-L/transformer/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:155de5763cfd2e43ad6352105e032ee195d1492c7de5b428727ec59fdfded154
+size 1867160768

NiT-L/transformer/nit_transformer_2d.py

@@ -0,0 +1,471 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+
+from dataclasses import dataclass
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    from diffusers.configuration_utils import ConfigMixin, register_to_config
+    from diffusers.models.modeling_utils import ModelMixin
+    from diffusers.utils import BaseOutput
+except Exception:  # pragma: no cover - lets this subtree be tested outside diffusers.
+    class BaseOutput(dict):
+        def __post_init__(self):
+            self.update(self.__dict__)
+
+    class _Config(dict):
+        def __getattr__(self, key):
+            try:
+                return self[key]
+            except KeyError as error:
+                raise AttributeError(key) from error
+
+    class ConfigMixin:
+        config_name = "config.json"
+
+    class ModelMixin(nn.Module):
+        pass
+
+    def register_to_config(init):
+        def wrapper(self, *args, **kwargs):
+            import inspect
+
+            signature = inspect.signature(init)
+            bound = signature.bind(self, *args, **kwargs)
+            bound.apply_defaults()
+            self.config = _Config({key: value for key, value in bound.arguments.items() if key != "self"})
+            init(self, *args, **kwargs)
+
+        return wrapper
+
+
+try:
+    from flash_attn import flash_attn_varlen_func
+except Exception:  # pragma: no cover - optional acceleration.
+    flash_attn_varlen_func = None
+
+
+@dataclass
+class NiTTransformer2DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+    projection_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+
+
+def _modulate(hidden_states: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
+    return hidden_states * (1 + scale) + shift
+
+
+def _rotate_half(hidden_states: torch.Tensor) -> torch.Tensor:
+    hidden_states = hidden_states.reshape(*hidden_states.shape[:-1], -1, 2)
+    hidden_states_1, hidden_states_2 = hidden_states.unbind(dim=-1)
+    return torch.stack((-hidden_states_2, hidden_states_1), dim=-1).flatten(-2)
+
+
+def _get_float_dtype_or_default(tensor: Optional[torch.Tensor] = None) -> torch.dtype:
+    if tensor is not None and tensor.is_floating_point():
+        return tensor.dtype
+    return torch.get_default_dtype()
+
+
+class NiTPatchEmbed(nn.Module):
+    def __init__(self, patch_size: int, in_channels: int, hidden_size: int):
+        super().__init__()
+        self.patch_size = (patch_size, patch_size)
+        self.proj = nn.Conv2d(in_channels, hidden_size, kernel_size=patch_size, stride=patch_size, bias=True)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.proj(hidden_states)
+        return hidden_states.flatten(2).transpose(1, 2)
+
+
+class NiTTimestepEmbedder(nn.Module):
+    def __init__(self, hidden_size: int, frequency_embedding_size: int = 256):
+        super().__init__()
+        self.frequency_embedding_size = frequency_embedding_size
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+
+    @staticmethod
+    def get_timestep_embedding(timesteps: torch.Tensor, embedding_dim: int, max_period: int = 10000):
+        half = embedding_dim // 2
+        # Keep sinusoid construction in fp32 to mirror native NiT behavior.
+        exponent = -math.log(max_period) * torch.arange(half, dtype=torch.float32, device=timesteps.device) / half
+        freqs = torch.exp(exponent)
+        args = timesteps.float()[:, None] * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if embedding_dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        timestep_freq = self.get_timestep_embedding(timesteps, self.frequency_embedding_size).to(timesteps.dtype)
+        return self.mlp(timestep_freq)
+
+
+class NiTLabelEmbedder(nn.Module):
+    def __init__(self, num_classes: int, hidden_size: int, dropout_prob: float):
+        super().__init__()
+        use_cfg_embedding = dropout_prob > 0
+        self.embedding_table = nn.Embedding(num_classes + int(use_cfg_embedding), hidden_size)
+        self.num_classes = num_classes
+        self.dropout_prob = dropout_prob
+
+    def forward(self, class_labels: torch.LongTensor) -> torch.Tensor:
+        return self.embedding_table(class_labels)
+
+
+class NiTRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        head_dim: int,
+        custom_freqs: str = "normal",
+        theta: int = 10000,
+        max_cached_len: int = 1024,
+        max_pe_len_h: Optional[int] = None,
+        max_pe_len_w: Optional[int] = None,
+        decouple: bool = False,
+        ori_max_pe_len: Optional[int] = None,
+    ):
+        super().__init__()
+        del max_pe_len_h, max_pe_len_w, decouple, ori_max_pe_len
+        if custom_freqs not in {"normal", "scale1", "scale2"}:
+            raise ValueError(
+                "This Diffusers implementation supports the trained RoPE frequencies directly. "
+                "Checkpoint conversion preserves weights; extrapolation variants should be handled "
+                "by changing the model config before loading."
+            )
+        dim = head_dim // 2
+        if dim % 2 != 0:
+            raise ValueError("NiT rotary embedding requires head_dim // 2 to be even.")
+        default_dtype = _get_float_dtype_or_default()
+        freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=default_dtype) / dim))
+        self.register_buffer("freqs_h", freqs, persistent=False)
+        self.register_buffer("freqs_w", freqs.clone(), persistent=False)
+        positions = torch.arange(max_cached_len, dtype=default_dtype)
+        freqs_h_cached = torch.einsum("n,f->nf", positions, self.freqs_h).repeat_interleave(2, dim=-1)
+        freqs_w_cached = torch.einsum("n,f->nf", positions, self.freqs_w).repeat_interleave(2, dim=-1)
+        self.register_buffer("freqs_h_cached", freqs_h_cached, persistent=False)
+        self.register_buffer("freqs_w_cached", freqs_w_cached, persistent=False)
+
+    def forward(self, image_sizes: torch.LongTensor, device: torch.device) -> Tuple[torch.Tensor, torch.Tensor]:
+        grids = []
+        for height, width in image_sizes.tolist():
+            # Match native NiT token ordering for RoPE alignment.
+            grid_h = torch.arange(height, device=device)
+            grid_w = torch.arange(width, device=device)
+            grid = torch.meshgrid(grid_h, grid_w, indexing="xy")
+            grids.append(torch.stack(grid, dim=0).reshape(2, -1))
+        grid = torch.cat(grids, dim=1)
+        freqs_h = self.freqs_h_cached.to(device)[grid[0]]
+        freqs_w = self.freqs_w_cached.to(device)[grid[1]]
+        freqs = torch.cat([freqs_h, freqs_w], dim=-1)
+        return freqs.cos().unsqueeze(1), freqs.sin().unsqueeze(1)
+
+
+class NiTAttention(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, qk_norm: bool = False):
+        super().__init__()
+        if hidden_size % num_heads != 0:
+            raise ValueError("hidden_size must be divisible by num_heads")
+        self.num_heads = num_heads
+        self.head_dim = hidden_size // num_heads
+        self.qkv = nn.Linear(hidden_size, hidden_size * 3, bias=True)
+        self.q_norm = nn.LayerNorm(self.head_dim) if qk_norm else nn.Identity()
+        self.k_norm = nn.LayerNorm(self.head_dim) if qk_norm else nn.Identity()
+        self.proj = nn.Linear(hidden_size, hidden_size)
+        self.proj_drop = nn.Dropout(0.0)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.IntTensor,
+        freqs_cos: torch.Tensor,
+        freqs_sin: torch.Tensor,
+    ) -> torch.Tensor:
+        qkv = self.qkv(hidden_states).reshape(hidden_states.shape[0], 3, self.num_heads, self.head_dim)
+        query, key, value = qkv.unbind(dim=1)
+        original_dtype = qkv.dtype
+        query = self.q_norm(query)
+        key = self.k_norm(key)
+        query = query * freqs_cos + _rotate_half(query) * freqs_sin
+        key = key * freqs_cos + _rotate_half(key) * freqs_sin
+        query = query.to(dtype=original_dtype)
+        key = key.to(dtype=original_dtype)
+
+        if flash_attn_varlen_func is not None and query.is_cuda:
+            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+            hidden_states = flash_attn_varlen_func(
+                query, key, value, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen
+            ).reshape(hidden_states.shape[0], -1)
+        else:
+            segments = []
+            for start, end in zip(cu_seqlens[:-1].tolist(), cu_seqlens[1:].tolist()):
+                q = query[start:end].transpose(0, 1).unsqueeze(0)
+                k = key[start:end].transpose(0, 1).unsqueeze(0)
+                v = value[start:end].transpose(0, 1).unsqueeze(0)
+                segments.append(F.scaled_dot_product_attention(q, k, v).squeeze(0).transpose(0, 1))
+            hidden_states = torch.cat(segments, dim=0).reshape(hidden_states.shape[0], -1)
+
+        hidden_states = self.proj(hidden_states)
+        return self.proj_drop(hidden_states)
+
+
+class NiTMLP(nn.Module):
+    def __init__(self, hidden_size: int, mlp_hidden_dim: int):
+        super().__init__()
+        self.fc1 = nn.Linear(hidden_size, mlp_hidden_dim)
+        self.act = nn.GELU(approximate="tanh")
+        self.drop1 = nn.Dropout(0.0)
+        self.norm = nn.Identity()
+        self.fc2 = nn.Linear(mlp_hidden_dim, hidden_size)
+        self.drop2 = nn.Dropout(0.0)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.drop1(hidden_states)
+        hidden_states = self.norm(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return self.drop2(hidden_states)
+
+
+class NiTBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_norm: bool = False,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 512,
+    ):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.attn = NiTAttention(hidden_size, num_heads=num_heads, qk_norm=qk_norm)
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.mlp = NiTMLP(hidden_size, mlp_hidden_dim)
+        if use_adaln_lora:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(hidden_size, adaln_lora_dim, bias=True),
+                nn.Linear(adaln_lora_dim, 6 * hidden_size, bias=True),
+            )
+        else:
+            self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True))
+
+    def forward(self, hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(conditioning).chunk(
+            6, dim=-1
+        )
+        hidden_states = hidden_states + gate_msa * self.attn(
+            _modulate(self.norm1(hidden_states), shift_msa, scale_msa), cu_seqlens, freqs_cos, freqs_sin
+        )
+        hidden_states = hidden_states + gate_mlp * self.mlp(
+            _modulate(self.norm2(hidden_states), shift_mlp, scale_mlp)
+        )
+        return hidden_states
+
+
+class NiTFinalLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
+
+    def forward(self, hidden_states: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
+        shift, scale = self.adaLN_modulation(conditioning).chunk(2, dim=-1)
+        hidden_states = _modulate(self.norm_final(hidden_states), shift, scale)
+        return self.linear(hidden_states)
+
+
+def _build_mlp(hidden_size: int, projector_dim: int, z_dim: int) -> nn.Sequential:
+    return nn.Sequential(
+        nn.Linear(hidden_size, projector_dim),
+        nn.SiLU(),
+        nn.Linear(projector_dim, projector_dim),
+        nn.SiLU(),
+        nn.Linear(projector_dim, z_dim),
+    )
+
+
+class NiTTransformer2DModel(ModelMixin, ConfigMixin):
+    config_name = "config.json"
+
+    @register_to_config
+    def __init__(
+        self,
+        input_size: int = 32,
+        patch_size: int = 1,
+        in_channels: int = 32,
+        hidden_size: int = 1152,
+        depth: int = 28,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.0,
+        class_dropout_prob: float = 0.1,
+        num_classes: int = 1000,
+        encoder_depth: int = 8,
+        projector_dim: int = 2048,
+        z_dim: int = 1280,
+        use_checkpoint: bool = False,
+        custom_freqs: str = "normal",
+        theta: int = 10000,
+        max_pe_len_h: Optional[int] = None,
+        max_pe_len_w: Optional[int] = None,
+        decouple: bool = False,
+        ori_max_pe_len: Optional[int] = None,
+        qk_norm: bool = True,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 512,
+    ):
+        super().__init__()
+        del input_size
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.num_classes = num_classes
+        self.encoder_depth = encoder_depth
+        self.use_checkpoint = use_checkpoint
+
+        self.x_embedder = NiTPatchEmbed(patch_size, in_channels, hidden_size)
+        self.t_embedder = NiTTimestepEmbedder(hidden_size)
+        self.y_embedder = NiTLabelEmbedder(num_classes, hidden_size, class_dropout_prob)
+        self.rope = NiTRotaryEmbedding(
+            hidden_size // num_heads,
+            custom_freqs=custom_freqs,
+            theta=theta,
+            max_pe_len_h=max_pe_len_h,
+            max_pe_len_w=max_pe_len_w,
+            decouple=decouple,
+            ori_max_pe_len=ori_max_pe_len,
+        )
+        self.projector = _build_mlp(hidden_size, projector_dim, z_dim)
+        self.blocks = nn.ModuleList(
+            [
+                NiTBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qk_norm=qk_norm,
+                    use_adaln_lora=use_adaln_lora,
+                    adaln_lora_dim=adaln_lora_dim,
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.final_layer = NiTFinalLayer(hidden_size, patch_size, self.out_channels)
+
+    def _pack_latents(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.LongTensor, Tuple[int, int]]:
+        batch_size, channels, height, width = hidden_states.shape
+        if channels != self.in_channels:
+            raise ValueError(f"Expected {self.in_channels} latent channels, got {channels}.")
+        if height % self.patch_size != 0 or width % self.patch_size != 0:
+            raise ValueError("Latent height and width must be divisible by patch_size.")
+        latent_h = height // self.patch_size
+        latent_w = width // self.patch_size
+        hidden_states = hidden_states.reshape(batch_size, channels, latent_h, self.patch_size, latent_w, self.patch_size)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5).reshape(
+            batch_size * latent_h * latent_w, channels, self.patch_size, self.patch_size
+        )
+        image_sizes = torch.tensor([[latent_h, latent_w]] * batch_size, device=hidden_states.device, dtype=torch.long)
+        return hidden_states, image_sizes, (height, width)
+
+    def _unpack_latents(self, hidden_states: torch.Tensor, image_sizes: torch.LongTensor) -> torch.Tensor:
+        if image_sizes.shape[0] == 1:
+            height, width = image_sizes[0].tolist()
+            hidden_states = hidden_states.reshape(height, width, self.out_channels, self.patch_size, self.patch_size)
+            return hidden_states.permute(2, 0, 3, 1, 4).reshape(
+                1, self.out_channels, height * self.patch_size, width * self.patch_size
+            )
+
+        samples = []
+        cursor = 0
+        for height, width in image_sizes.tolist():
+            length = height * width
+            sample = hidden_states[cursor : cursor + length].reshape(
+                height, width, self.out_channels, self.patch_size, self.patch_size
+            )
+            samples.append(
+                sample.permute(2, 0, 3, 1, 4).reshape(
+                    self.out_channels, height * self.patch_size, width * self.patch_size
+                )
+            )
+            cursor += length
+        if len({tuple(sample.shape) for sample in samples}) != 1:
+            return hidden_states
+        return torch.stack(samples, dim=0)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: Union[torch.Tensor, float],
+        class_labels: torch.LongTensor,
+        image_sizes: Optional[Union[torch.LongTensor, List[Tuple[int, int]]]] = None,
+        return_dict: bool = True,
+        output_projection_states: bool = False,
+    ) -> Union[NiTTransformer2DModelOutput, Tuple[torch.Tensor, ...]]:
+        input_was_image = hidden_states.dim() == 4 and image_sizes is None
+        if input_was_image:
+            hidden_states, image_sizes, _ = self._pack_latents(hidden_states)
+        elif image_sizes is None:
+            raise ValueError("image_sizes must be provided when hidden_states are already packed.")
+        elif not torch.is_tensor(image_sizes):
+            image_sizes = torch.tensor(image_sizes, device=hidden_states.device, dtype=torch.long)
+        else:
+            image_sizes = image_sizes.to(device=hidden_states.device, dtype=torch.long)
+
+        if not torch.is_tensor(timestep):
+            timestep = torch.tensor([timestep], device=hidden_states.device, dtype=hidden_states.dtype)
+        timestep = timestep.to(device=hidden_states.device, dtype=hidden_states.dtype).flatten()
+        if timestep.numel() == 1:
+            timestep = timestep.repeat(image_sizes.shape[0])
+        class_labels = class_labels.to(device=hidden_states.device, dtype=torch.long).flatten()
+
+        hidden_states = self.x_embedder(hidden_states).squeeze(1)
+        freqs_cos, freqs_sin = self.rope(image_sizes, hidden_states.device)
+
+        seqlens = image_sizes[:, 0] * image_sizes[:, 1]
+        cu_seqlens = torch.cat(
+            [torch.zeros(1, device=hidden_states.device, dtype=torch.int32), torch.cumsum(seqlens, dim=0).int()]
+        )
+
+        conditioning = self.t_embedder(timestep) + self.y_embedder(class_labels)
+        conditioning = torch.cat([conditioning[i].repeat(int(seqlens[i]), 1) for i in range(image_sizes.shape[0])], dim=0)
+
+        projection_states = []
+        for index, block in enumerate(self.blocks):
+            if self.use_checkpoint and self.training:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    block, hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin, use_reentrant=False
+                )
+            else:
+                hidden_states = block(hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin)
+            if output_projection_states and (index + 1) == self.encoder_depth:
+                projection_states.append(self.projector(hidden_states))
+
+        hidden_states = self.final_layer(hidden_states, conditioning)
+        hidden_states = hidden_states.reshape(hidden_states.shape[0], self.out_channels, self.patch_size, self.patch_size)
+        if input_was_image:
+            hidden_states = self._unpack_latents(hidden_states, image_sizes)
+
+        if not return_dict:
+            output = (hidden_states,)
+            if output_projection_states:
+                output = output + (tuple(projection_states),)
+            return output
+        return NiTTransformer2DModelOutput(
+            sample=hidden_states,
+            projection_states=tuple(projection_states) if output_projection_states else None,
+        )

NiT-L/vae/config.json

@@ -0,0 +1,88 @@
+{
+  "_class_name": "AutoencoderDC",
+  "_diffusers_version": "0.32.2",
+  "attention_head_dim": 32,
+  "decoder_act_fns": "silu",
+  "decoder_block_out_channels": [
+    128,
+    256,
+    512,
+    512,
+    1024,
+    1024
+  ],
+  "decoder_block_types": [
+    "ResBlock",
+    "ResBlock",
+    "ResBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock"
+  ],
+  "decoder_layers_per_block": [
+    3,
+    3,
+    3,
+    3,
+    3,
+    3
+  ],
+  "decoder_norm_types": "rms_norm",
+  "decoder_qkv_multiscales": [
+    [],
+    [],
+    [],
+    [
+      5
+    ],
+    [
+      5
+    ],
+    [
+      5
+    ]
+  ],
+  "downsample_block_type": "Conv",
+  "encoder_block_out_channels": [
+    128,
+    256,
+    512,
+    512,
+    1024,
+    1024
+  ],
+  "encoder_block_types": [
+    "ResBlock",
+    "ResBlock",
+    "ResBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock"
+  ],
+  "encoder_layers_per_block": [
+    2,
+    2,
+    2,
+    3,
+    3,
+    3
+  ],
+  "encoder_qkv_multiscales": [
+    [],
+    [],
+    [],
+    [
+      5
+    ],
+    [
+      5
+    ],
+    [
+      5
+    ]
+  ],
+  "in_channels": 3,
+  "latent_channels": 32,
+  "scaling_factor": 0.41407,
+  "upsample_block_type": "interpolate"
+}

NiT-L/vae/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfd991d1b54ffabf22745c5885589d8f2a7bc59930d95d92bd741c4fc64454bb
+size 1249044836

NiT-S/README.md

@@ -0,0 +1,91 @@
+---
+license: apache-2.0
+library_name: diffusers
+pipeline_tag: unconditional-image-generation
+tags:
+  - diffusers
+  - nit
+  - image-generation
+  - class-conditional
+  - imagenet
+inference: true
+---
+
+# NiT-S
+
+Self-contained Diffusers checkpoint for **NiT-S** (33M), converted from [`GoodEnough/NiT-S-Models`](https://huggingface.co/GoodEnough/NiT-S-Models) (`model_500K.safetensors`, 500K training steps).
+
+Architecture and training settings follow the official [`nit_s_pack_merge_radio_65536.yaml`](https://github.com/WZDTHU/NiT/blob/main/configs/c2i/nit_s_pack_merge_radio_65536.yaml).
+
+## Model config
+
+| Field | Value |
+| --- | --- |
+| Parameters | 33M |
+| Depth | 12 |
+| Hidden size | 384 |
+| Attention heads | 6 |
+| Encoder depth | 4 |
+| Latent channels (`z_dim`) | 768 |
+| Projector dim | 768 |
+| Patch size | 1 |
+| Input latent channels | 32 |
+| Classes | 1000 |
+| Class dropout | 0.1 |
+| QK norm | true |
+| VAE | `mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers` |
+| Flow path type | linear |
+
+## Recommended inference (256×256)
+
+Official NiT sampling defaults for **256×256** class-conditional ImageNet generation:
+
+| Setting | Value |
+| --- | --- |
+| Resolution | 256×256 |
+| Solver | SDE (Euler–Maruyama) in the official repo |
+| Steps (NFE) | 250 |
+| CFG scale | 2.25 |
+| CFG interval | (0.0, 0.7) |
+
+This Diffusers port uses [`FlowMatchEulerDiscreteScheduler`](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete) in deterministic ODE mode (`stochastic_sampling=false`). Keep the same step count, CFG scale, and interval as the official recipe.
+
+## Usage
+
+```python
+from pathlib import Path
+import torch
+from diffusers import DiffusionPipeline
+
+model_dir = Path(".")
+pipe = DiffusionPipeline.from_pretrained(
+    str(model_dir),
+    local_files_only=True,
+    custom_pipeline=str(model_dir / "pipeline.py"),
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+).to("cuda" if torch.cuda.is_available() else "cpu")
+
+generator = torch.Generator(device=pipe.device).manual_seed(42)
+image = pipe(
+    class_labels="golden retriever",
+    height=256,
+    width=256,
+    num_inference_steps=250,
+    guidance_scale=2.25,
+    guidance_interval=(0.0, 0.7),
+    generator=generator,
+).images[0]
+image.save("demo_256.png")
+```
+
+## Components
+
+- `pipeline.py` — custom `NiTPipeline`
+- `model_index.json` — pipeline index + ImageNet `id2label`
+- `transformer/config.json`
+- `transformer/nit_transformer_2d.py`
+- `transformer/diffusion_pytorch_model.safetensors`
+- `scheduler/scheduler_config.json`
+- `vae/config.json`
+- `vae/diffusion_pytorch_model.safetensors`

NiT-S/model_index.json

@@ -0,0 +1,1021 @@
+{
+  "_class_name": [
+    "pipeline",
+    "NiTPipeline"
+  ],
+  "_diffusers_version": "0.36.0",
+  "scheduler": [
+    "diffusers",
+    "FlowMatchEulerDiscreteScheduler"
+  ],
+  "transformer": [
+    "nit_transformer_2d",
+    "NiTTransformer2DModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderDC"
+  ],
+  "id2label": {
+    "0": "tench, Tinca tinca",
+    "1": "goldfish, Carassius auratus",
+    "2": "great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias",
+    "3": "tiger shark, Galeocerdo cuvieri",
+    "4": "hammerhead, hammerhead shark",
+    "5": "electric ray, crampfish, numbfish, torpedo",
+    "6": "stingray",
+    "7": "cock",
+    "8": "hen",
+    "9": "ostrich, Struthio camelus",
+    "10": "brambling, Fringilla montifringilla",
+    "11": "goldfinch, Carduelis carduelis",
+    "12": "house finch, linnet, Carpodacus mexicanus",
+    "13": "junco, snowbird",
+    "14": "indigo bunting, indigo finch, indigo bird, Passerina cyanea",
+    "15": "robin, American robin, Turdus migratorius",
+    "16": "bulbul",
+    "17": "jay",
+    "18": "magpie",
+    "19": "chickadee",
+    "20": "water ouzel, dipper",
+    "21": "kite",
+    "22": "bald eagle, American eagle, Haliaeetus leucocephalus",
+    "23": "vulture",
+    "24": "great grey owl, great gray owl, Strix nebulosa",
+    "25": "European fire salamander, Salamandra salamandra",
+    "26": "common newt, Triturus vulgaris",
+    "27": "eft",
+    "28": "spotted salamander, Ambystoma maculatum",
+    "29": "axolotl, mud puppy, Ambystoma mexicanum",
+    "30": "bullfrog, Rana catesbeiana",
+    "31": "tree frog, tree-frog",
+    "32": "tailed frog, bell toad, ribbed toad, tailed toad, Ascaphus trui",
+    "33": "loggerhead, loggerhead turtle, Caretta caretta",
+    "34": "leatherback turtle, leatherback, leathery turtle, Dermochelys coriacea",
+    "35": "mud turtle",
+    "36": "terrapin",
+    "37": "box turtle, box tortoise",
+    "38": "banded gecko",
+    "39": "common iguana, iguana, Iguana iguana",
+    "40": "American chameleon, anole, Anolis carolinensis",
+    "41": "whiptail, whiptail lizard",
+    "42": "agama",
+    "43": "frilled lizard, Chlamydosaurus kingi",
+    "44": "alligator lizard",
+    "45": "Gila monster, Heloderma suspectum",
+    "46": "green lizard, Lacerta viridis",
+    "47": "African chameleon, Chamaeleo chamaeleon",
+    "48": "Komodo dragon, Komodo lizard, dragon lizard, giant lizard, Varanus komodoensis",
+    "49": "African crocodile, Nile crocodile, Crocodylus niloticus",
+    "50": "American alligator, Alligator mississipiensis",
+    "51": "triceratops",
+    "52": "thunder snake, worm snake, Carphophis amoenus",
+    "53": "ringneck snake, ring-necked snake, ring snake",
+    "54": "hognose snake, puff adder, sand viper",
+    "55": "green snake, grass snake",
+    "56": "king snake, kingsnake",
+    "57": "garter snake, grass snake",
+    "58": "water snake",
+    "59": "vine snake",
+    "60": "night snake, Hypsiglena torquata",
+    "61": "boa constrictor, Constrictor constrictor",
+    "62": "rock python, rock snake, Python sebae",
+    "63": "Indian cobra, Naja naja",
+    "64": "green mamba",
+    "65": "sea snake",
+    "66": "horned viper, cerastes, sand viper, horned asp, Cerastes cornutus",
+    "67": "diamondback, diamondback rattlesnake, Crotalus adamanteus",
+    "68": "sidewinder, horned rattlesnake, Crotalus cerastes",
+    "69": "trilobite",
+    "70": "harvestman, daddy longlegs, Phalangium opilio",
+    "71": "scorpion",
+    "72": "black and gold garden spider, Argiope aurantia",
+    "73": "barn spider, Araneus cavaticus",
+    "74": "garden spider, Aranea diademata",
+    "75": "black widow, Latrodectus mactans",
+    "76": "tarantula",
+    "77": "wolf spider, hunting spider",
+    "78": "tick",
+    "79": "centipede",
+    "80": "black grouse",
+    "81": "ptarmigan",
+    "82": "ruffed grouse, partridge, Bonasa umbellus",
+    "83": "prairie chicken, prairie grouse, prairie fowl",
+    "84": "peacock",
+    "85": "quail",
+    "86": "partridge",
+    "87": "African grey, African gray, Psittacus erithacus",
+    "88": "macaw",
+    "89": "sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita",
+    "90": "lorikeet",
+    "91": "coucal",
+    "92": "bee eater",
+    "93": "hornbill",
+    "94": "hummingbird",
+    "95": "jacamar",
+    "96": "toucan",
+    "97": "drake",
+    "98": "red-breasted merganser, Mergus serrator",
+    "99": "goose",
+    "100": "black swan, Cygnus atratus",
+    "101": "tusker",
+    "102": "echidna, spiny anteater, anteater",
+    "103": "platypus, duckbill, duckbilled platypus, duck-billed platypus, Ornithorhynchus anatinus",
+    "104": "wallaby, brush kangaroo",
+    "105": "koala, koala bear, kangaroo bear, native bear, Phascolarctos cinereus",
+    "106": "wombat",
+    "107": "jellyfish",
+    "108": "sea anemone, anemone",
+    "109": "brain coral",
+    "110": "flatworm, platyhelminth",
+    "111": "nematode, nematode worm, roundworm",
+    "112": "conch",
+    "113": "snail",
+    "114": "slug",
+    "115": "sea slug, nudibranch",
+    "116": "chiton, coat-of-mail shell, sea cradle, polyplacophore",
+    "117": "chambered nautilus, pearly nautilus, nautilus",
+    "118": "Dungeness crab, Cancer magister",
+    "119": "rock crab, Cancer irroratus",
+    "120": "fiddler crab",
+    "121": "king crab, Alaska crab, Alaskan king crab, Alaska king crab, Paralithodes camtschatica",
+    "122": "American lobster, Northern lobster, Maine lobster, Homarus americanus",
+    "123": "spiny lobster, langouste, rock lobster, crawfish, crayfish, sea crawfish",
+    "124": "crayfish, crawfish, crawdad, crawdaddy",
+    "125": "hermit crab",
+    "126": "isopod",
+    "127": "white stork, Ciconia ciconia",
+    "128": "black stork, Ciconia nigra",
+    "129": "spoonbill",
+    "130": "flamingo",
+    "131": "little blue heron, Egretta caerulea",
+    "132": "American egret, great white heron, Egretta albus",
+    "133": "bittern",
+    "134": "crane",
+    "135": "limpkin, Aramus pictus",
+    "136": "European gallinule, Porphyrio porphyrio",
+    "137": "American coot, marsh hen, mud hen, water hen, Fulica americana",
+    "138": "bustard",
+    "139": "ruddy turnstone, Arenaria interpres",
+    "140": "red-backed sandpiper, dunlin, Erolia alpina",
+    "141": "redshank, Tringa totanus",
+    "142": "dowitcher",
+    "143": "oystercatcher, oyster catcher",
+    "144": "pelican",
+    "145": "king penguin, Aptenodytes patagonica",
+    "146": "albatross, mollymawk",
+    "147": "grey whale, gray whale, devilfish, Eschrichtius gibbosus, Eschrichtius robustus",
+    "148": "killer whale, killer, orca, grampus, sea wolf, Orcinus orca",
+    "149": "dugong, Dugong dugon",
+    "150": "sea lion",
+    "151": "Chihuahua",
+    "152": "Japanese spaniel",
+    "153": "Maltese dog, Maltese terrier, Maltese",
+    "154": "Pekinese, Pekingese, Peke",
+    "155": "Shih-Tzu",
+    "156": "Blenheim spaniel",
+    "157": "papillon",
+    "158": "toy terrier",
+    "159": "Rhodesian ridgeback",
+    "160": "Afghan hound, Afghan",
+    "161": "basset, basset hound",
+    "162": "beagle",
+    "163": "bloodhound, sleuthhound",
+    "164": "bluetick",
+    "165": "black-and-tan coonhound",
+    "166": "Walker hound, Walker foxhound",
+    "167": "English foxhound",
+    "168": "redbone",
+    "169": "borzoi, Russian wolfhound",
+    "170": "Irish wolfhound",
+    "171": "Italian greyhound",
+    "172": "whippet",
+    "173": "Ibizan hound, Ibizan Podenco",
+    "174": "Norwegian elkhound, elkhound",
+    "175": "otterhound, otter hound",
+    "176": "Saluki, gazelle hound",
+    "177": "Scottish deerhound, deerhound",
+    "178": "Weimaraner",
+    "179": "Staffordshire bullterrier, Staffordshire bull terrier",
+    "180": "American Staffordshire terrier, Staffordshire terrier, American pit bull terrier, pit bull terrier",
+    "181": "Bedlington terrier",
+    "182": "Border terrier",
+    "183": "Kerry blue terrier",
+    "184": "Irish terrier",
+    "185": "Norfolk terrier",
+    "186": "Norwich terrier",
+    "187": "Yorkshire terrier",
+    "188": "wire-haired fox terrier",
+    "189": "Lakeland terrier",
+    "190": "Sealyham terrier, Sealyham",
+    "191": "Airedale, Airedale terrier",
+    "192": "cairn, cairn terrier",
+    "193": "Australian terrier",
+    "194": "Dandie Dinmont, Dandie Dinmont terrier",
+    "195": "Boston bull, Boston terrier",
+    "196": "miniature schnauzer",
+    "197": "giant schnauzer",
+    "198": "standard schnauzer",
+    "199": "Scotch terrier, Scottish terrier, Scottie",
+    "200": "Tibetan terrier, chrysanthemum dog",
+    "201": "silky terrier, Sydney silky",
+    "202": "soft-coated wheaten terrier",
+    "203": "West Highland white terrier",
+    "204": "Lhasa, Lhasa apso",
+    "205": "flat-coated retriever",
+    "206": "curly-coated retriever",
+    "207": "golden retriever",
+    "208": "Labrador retriever",
+    "209": "Chesapeake Bay retriever",
+    "210": "German short-haired pointer",
+    "211": "vizsla, Hungarian pointer",
+    "212": "English setter",
+    "213": "Irish setter, red setter",
+    "214": "Gordon setter",
+    "215": "Brittany spaniel",
+    "216": "clumber, clumber spaniel",
+    "217": "English springer, English springer spaniel",
+    "218": "Welsh springer spaniel",
+    "219": "cocker spaniel, English cocker spaniel, cocker",
+    "220": "Sussex spaniel",
+    "221": "Irish water spaniel",
+    "222": "kuvasz",
+    "223": "schipperke",
+    "224": "groenendael",
+    "225": "malinois",
+    "226": "briard",
+    "227": "kelpie",
+    "228": "komondor",
+    "229": "Old English sheepdog, bobtail",
+    "230": "Shetland sheepdog, Shetland sheep dog, Shetland",
+    "231": "collie",
+    "232": "Border collie",
+    "233": "Bouvier des Flandres, Bouviers des Flandres",
+    "234": "Rottweiler",
+    "235": "German shepherd, German shepherd dog, German police dog, alsatian",
+    "236": "Doberman, Doberman pinscher",
+    "237": "miniature pinscher",
+    "238": "Greater Swiss Mountain dog",
+    "239": "Bernese mountain dog",
+    "240": "Appenzeller",
+    "241": "EntleBucher",
+    "242": "boxer",
+    "243": "bull mastiff",
+    "244": "Tibetan mastiff",
+    "245": "French bulldog",
+    "246": "Great Dane",
+    "247": "Saint Bernard, St Bernard",
+    "248": "Eskimo dog, husky",
+    "249": "malamute, malemute, Alaskan malamute",
+    "250": "Siberian husky",
+    "251": "dalmatian, coach dog, carriage dog",
+    "252": "affenpinscher, monkey pinscher, monkey dog",
+    "253": "basenji",
+    "254": "pug, pug-dog",
+    "255": "Leonberg",
+    "256": "Newfoundland, Newfoundland dog",
+    "257": "Great Pyrenees",
+    "258": "Samoyed, Samoyede",
+    "259": "Pomeranian",
+    "260": "chow, chow chow",
+    "261": "keeshond",
+    "262": "Brabancon griffon",
+    "263": "Pembroke, Pembroke Welsh corgi",
+    "264": "Cardigan, Cardigan Welsh corgi",
+    "265": "toy poodle",
+    "266": "miniature poodle",
+    "267": "standard poodle",
+    "268": "Mexican hairless",
+    "269": "timber wolf, grey wolf, gray wolf, Canis lupus",
+    "270": "white wolf, Arctic wolf, Canis lupus tundrarum",
+    "271": "red wolf, maned wolf, Canis rufus, Canis niger",
+    "272": "coyote, prairie wolf, brush wolf, Canis latrans",
+    "273": "dingo, warrigal, warragal, Canis dingo",
+    "274": "dhole, Cuon alpinus",
+    "275": "African hunting dog, hyena dog, Cape hunting dog, Lycaon pictus",
+    "276": "hyena, hyaena",
+    "277": "red fox, Vulpes vulpes",
+    "278": "kit fox, Vulpes macrotis",
+    "279": "Arctic fox, white fox, Alopex lagopus",
+    "280": "grey fox, gray fox, Urocyon cinereoargenteus",
+    "281": "tabby, tabby cat",
+    "282": "tiger cat",
+    "283": "Persian cat",
+    "284": "Siamese cat, Siamese",
+    "285": "Egyptian cat",
+    "286": "cougar, puma, catamount, mountain lion, painter, panther, Felis concolor",
+    "287": "lynx, catamount",
+    "288": "leopard, Panthera pardus",
+    "289": "snow leopard, ounce, Panthera uncia",
+    "290": "jaguar, panther, Panthera onca, Felis onca",
+    "291": "lion, king of beasts, Panthera leo",
+    "292": "tiger, Panthera tigris",
+    "293": "cheetah, chetah, Acinonyx jubatus",
+    "294": "brown bear, bruin, Ursus arctos",
+    "295": "American black bear, black bear, Ursus americanus, Euarctos americanus",
+    "296": "ice bear, polar bear, Ursus Maritimus, Thalarctos maritimus",
+    "297": "sloth bear, Melursus ursinus, Ursus ursinus",
+    "298": "mongoose",
+    "299": "meerkat, mierkat",
+    "300": "tiger beetle",
+    "301": "ladybug, ladybeetle, lady beetle, ladybird, ladybird beetle",
+    "302": "ground beetle, carabid beetle",
+    "303": "long-horned beetle, longicorn, longicorn beetle",
+    "304": "leaf beetle, chrysomelid",
+    "305": "dung beetle",
+    "306": "rhinoceros beetle",
+    "307": "weevil",
+    "308": "fly",
+    "309": "bee",
+    "310": "ant, emmet, pismire",
+    "311": "grasshopper, hopper",
+    "312": "cricket",
+    "313": "walking stick, walkingstick, stick insect",
+    "314": "cockroach, roach",
+    "315": "mantis, mantid",
+    "316": "cicada, cicala",
+    "317": "leafhopper",
+    "318": "lacewing, lacewing fly",
+    "319": "dragonfly, darning needle, devils darning needle, sewing needle, snake feeder, snake doctor, mosquito hawk, skeeter hawk",
+    "320": "damselfly",
+    "321": "admiral",
+    "322": "ringlet, ringlet butterfly",
+    "323": "monarch, monarch butterfly, milkweed butterfly, Danaus plexippus",
+    "324": "cabbage butterfly",
+    "325": "sulphur butterfly, sulfur butterfly",
+    "326": "lycaenid, lycaenid butterfly",
+    "327": "starfish, sea star",
+    "328": "sea urchin",
+    "329": "sea cucumber, holothurian",
+    "330": "wood rabbit, cottontail, cottontail rabbit",
+    "331": "hare",
+    "332": "Angora, Angora rabbit",
+    "333": "hamster",
+    "334": "porcupine, hedgehog",
+    "335": "fox squirrel, eastern fox squirrel, Sciurus niger",
+    "336": "marmot",
+    "337": "beaver",
+    "338": "guinea pig, Cavia cobaya",
+    "339": "sorrel",
+    "340": "zebra",
+    "341": "hog, pig, grunter, squealer, Sus scrofa",
+    "342": "wild boar, boar, Sus scrofa",
+    "343": "warthog",
+    "344": "hippopotamus, hippo, river horse, Hippopotamus amphibius",
+    "345": "ox",
+    "346": "water buffalo, water ox, Asiatic buffalo, Bubalus bubalis",
+    "347": "bison",
+    "348": "ram, tup",
+    "349": "bighorn, bighorn sheep, cimarron, Rocky Mountain bighorn, Rocky Mountain sheep, Ovis canadensis",
+    "350": "ibex, Capra ibex",
+    "351": "hartebeest",
+    "352": "impala, Aepyceros melampus",
+    "353": "gazelle",
+    "354": "Arabian camel, dromedary, Camelus dromedarius",
+    "355": "llama",
+    "356": "weasel",
+    "357": "mink",
+    "358": "polecat, fitch, foulmart, foumart, Mustela putorius",
+    "359": "black-footed ferret, ferret, Mustela nigripes",
+    "360": "otter",
+    "361": "skunk, polecat, wood pussy",
+    "362": "badger",
+    "363": "armadillo",
+    "364": "three-toed sloth, ai, Bradypus tridactylus",
+    "365": "orangutan, orang, orangutang, Pongo pygmaeus",
+    "366": "gorilla, Gorilla gorilla",
+    "367": "chimpanzee, chimp, Pan troglodytes",
+    "368": "gibbon, Hylobates lar",
+    "369": "siamang, Hylobates syndactylus, Symphalangus syndactylus",
+    "370": "guenon, guenon monkey",
+    "371": "patas, hussar monkey, Erythrocebus patas",
+    "372": "baboon",
+    "373": "macaque",
+    "374": "langur",
+    "375": "colobus, colobus monkey",
+    "376": "proboscis monkey, Nasalis larvatus",
+    "377": "marmoset",
+    "378": "capuchin, ringtail, Cebus capucinus",
+    "379": "howler monkey, howler",
+    "380": "titi, titi monkey",
+    "381": "spider monkey, Ateles geoffroyi",
+    "382": "squirrel monkey, Saimiri sciureus",
+    "383": "Madagascar cat, ring-tailed lemur, Lemur catta",
+    "384": "indri, indris, Indri indri, Indri brevicaudatus",
+    "385": "Indian elephant, Elephas maximus",
+    "386": "African elephant, Loxodonta africana",
+    "387": "lesser panda, red panda, panda, bear cat, cat bear, Ailurus fulgens",
+    "388": "giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca",
+    "389": "barracouta, snoek",
+    "390": "eel",
+    "391": "coho, cohoe, coho salmon, blue jack, silver salmon, Oncorhynchus kisutch",
+    "392": "rock beauty, Holocanthus tricolor",
+    "393": "anemone fish",
+    "394": "sturgeon",
+    "395": "gar, garfish, garpike, billfish, Lepisosteus osseus",
+    "396": "lionfish",
+    "397": "puffer, pufferfish, blowfish, globefish",
+    "398": "abacus",
+    "399": "abaya",
+    "400": "academic gown, academic robe, judge robe",
+    "401": "accordion, piano accordion, squeeze box",
+    "402": "acoustic guitar",
+    "403": "aircraft carrier, carrier, flattop, attack aircraft carrier",
+    "404": "airliner",
+    "405": "airship, dirigible",
+    "406": "altar",
+    "407": "ambulance",
+    "408": "amphibian, amphibious vehicle",
+    "409": "analog clock",
+    "410": "apiary, bee house",
+    "411": "apron",
+    "412": "ashcan, trash can, garbage can, wastebin, ash bin, ash-bin, ashbin, dustbin, trash barrel, trash bin",
+    "413": "assault rifle, assault gun",
+    "414": "backpack, back pack, knapsack, packsack, rucksack, haversack",
+    "415": "bakery, bakeshop, bakehouse",
+    "416": "balance beam, beam",
+    "417": "balloon",
+    "418": "ballpoint, ballpoint pen, ballpen, Biro",
+    "419": "Band Aid",
+    "420": "banjo",
+    "421": "bannister, banister, balustrade, balusters, handrail",
+    "422": "barbell",
+    "423": "barber chair",
+    "424": "barbershop",
+    "425": "barn",
+    "426": "barometer",
+    "427": "barrel, cask",
+    "428": "barrow, garden cart, lawn cart, wheelbarrow",
+    "429": "baseball",
+    "430": "basketball",
+    "431": "bassinet",
+    "432": "bassoon",
+    "433": "bathing cap, swimming cap",
+    "434": "bath towel",
+    "435": "bathtub, bathing tub, bath, tub",
+    "436": "beach wagon, station wagon, wagon, estate car, beach waggon, station waggon, waggon",
+    "437": "beacon, lighthouse, beacon light, pharos",
+    "438": "beaker",
+    "439": "bearskin, busby, shako",
+    "440": "beer bottle",
+    "441": "beer glass",
+    "442": "bell cote, bell cot",
+    "443": "bib",
+    "444": "bicycle-built-for-two, tandem bicycle, tandem",
+    "445": "bikini, two-piece",
+    "446": "binder, ring-binder",
+    "447": "binoculars, field glasses, opera glasses",
+    "448": "birdhouse",
+    "449": "boathouse",
+    "450": "bobsled, bobsleigh, bob",
+    "451": "bolo tie, bolo, bola tie, bola",
+    "452": "bonnet, poke bonnet",
+    "453": "bookcase",
+    "454": "bookshop, bookstore, bookstall",
+    "455": "bottlecap",
+    "456": "bow",
+    "457": "bow tie, bow-tie, bowtie",
+    "458": "brass, memorial tablet, plaque",
+    "459": "brassiere, bra, bandeau",
+    "460": "breakwater, groin, groyne, mole, bulwark, seawall, jetty",
+    "461": "breastplate, aegis, egis",
+    "462": "broom",
+    "463": "bucket, pail",
+    "464": "buckle",
+    "465": "bulletproof vest",
+    "466": "bullet train, bullet",
+    "467": "butcher shop, meat market",
+    "468": "cab, hack, taxi, taxicab",
+    "469": "caldron, cauldron",
+    "470": "candle, taper, wax light",
+    "471": "cannon",
+    "472": "canoe",
+    "473": "can opener, tin opener",
+    "474": "cardigan",
+    "475": "car mirror",
+    "476": "carousel, carrousel, merry-go-round, roundabout, whirligig",
+    "477": "carpenters kit, tool kit",
+    "478": "carton",
+    "479": "car wheel",
+    "480": "cash machine, cash dispenser, automated teller machine, automatic teller machine, automated teller, automatic teller, ATM",
+    "481": "cassette",
+    "482": "cassette player",
+    "483": "castle",
+    "484": "catamaran",
+    "485": "CD player",
+    "486": "cello, violoncello",
+    "487": "cellular telephone, cellular phone, cellphone, cell, mobile phone",
+    "488": "chain",
+    "489": "chainlink fence",
+    "490": "chain mail, ring mail, mail, chain armor, chain armour, ring armor, ring armour",
+    "491": "chain saw, chainsaw",
+    "492": "chest",
+    "493": "chiffonier, commode",
+    "494": "chime, bell, gong",
+    "495": "china cabinet, china closet",
+    "496": "Christmas stocking",
+    "497": "church, church building",
+    "498": "cinema, movie theater, movie theatre, movie house, picture palace",
+    "499": "cleaver, meat cleaver, chopper",
+    "500": "cliff dwelling",
+    "501": "cloak",
+    "502": "clog, geta, patten, sabot",
+    "503": "cocktail shaker",
+    "504": "coffee mug",
+    "505": "coffeepot",
+    "506": "coil, spiral, volute, whorl, helix",
+    "507": "combination lock",
+    "508": "computer keyboard, keypad",
+    "509": "confectionery, confectionary, candy store",
+    "510": "container ship, containership, container vessel",
+    "511": "convertible",
+    "512": "corkscrew, bottle screw",
+    "513": "cornet, horn, trumpet, trump",
+    "514": "cowboy boot",
+    "515": "cowboy hat, ten-gallon hat",
+    "516": "cradle",
+    "517": "crane",
+    "518": "crash helmet",
+    "519": "crate",
+    "520": "crib, cot",
+    "521": "Crock Pot",
+    "522": "croquet ball",
+    "523": "crutch",
+    "524": "cuirass",
+    "525": "dam, dike, dyke",
+    "526": "desk",
+    "527": "desktop computer",
+    "528": "dial telephone, dial phone",
+    "529": "diaper, nappy, napkin",
+    "530": "digital clock",
+    "531": "digital watch",
+    "532": "dining table, board",
+    "533": "dishrag, dishcloth",
+    "534": "dishwasher, dish washer, dishwashing machine",
+    "535": "disk brake, disc brake",
+    "536": "dock, dockage, docking facility",
+    "537": "dogsled, dog sled, dog sleigh",
+    "538": "dome",
+    "539": "doormat, welcome mat",
+    "540": "drilling platform, offshore rig",
+    "541": "drum, membranophone, tympan",
+    "542": "drumstick",
+    "543": "dumbbell",
+    "544": "Dutch oven",
+    "545": "electric fan, blower",
+    "546": "electric guitar",
+    "547": "electric locomotive",
+    "548": "entertainment center",
+    "549": "envelope",
+    "550": "espresso maker",
+    "551": "face powder",
+    "552": "feather boa, boa",
+    "553": "file, file cabinet, filing cabinet",
+    "554": "fireboat",
+    "555": "fire engine, fire truck",
+    "556": "fire screen, fireguard",
+    "557": "flagpole, flagstaff",
+    "558": "flute, transverse flute",
+    "559": "folding chair",
+    "560": "football helmet",
+    "561": "forklift",
+    "562": "fountain",
+    "563": "fountain pen",
+    "564": "four-poster",
+    "565": "freight car",
+    "566": "French horn, horn",
+    "567": "frying pan, frypan, skillet",
+    "568": "fur coat",
+    "569": "garbage truck, dustcart",
+    "570": "gasmask, respirator, gas helmet",
+    "571": "gas pump, gasoline pump, petrol pump, island dispenser",
+    "572": "goblet",
+    "573": "go-kart",
+    "574": "golf ball",
+    "575": "golfcart, golf cart",
+    "576": "gondola",
+    "577": "gong, tam-tam",
+    "578": "gown",
+    "579": "grand piano, grand",
+    "580": "greenhouse, nursery, glasshouse",
+    "581": "grille, radiator grille",
+    "582": "grocery store, grocery, food market, market",
+    "583": "guillotine",
+    "584": "hair slide",
+    "585": "hair spray",
+    "586": "half track",
+    "587": "hammer",
+    "588": "hamper",
+    "589": "hand blower, blow dryer, blow drier, hair dryer, hair drier",
+    "590": "hand-held computer, hand-held microcomputer",
+    "591": "handkerchief, hankie, hanky, hankey",
+    "592": "hard disc, hard disk, fixed disk",
+    "593": "harmonica, mouth organ, harp, mouth harp",
+    "594": "harp",
+    "595": "harvester, reaper",
+    "596": "hatchet",
+    "597": "holster",
+    "598": "home theater, home theatre",
+    "599": "honeycomb",
+    "600": "hook, claw",
+    "601": "hoopskirt, crinoline",
+    "602": "horizontal bar, high bar",
+    "603": "horse cart, horse-cart",
+    "604": "hourglass",
+    "605": "iPod",
+    "606": "iron, smoothing iron",
+    "607": "jack-o-lantern",
+    "608": "jean, blue jean, denim",
+    "609": "jeep, landrover",
+    "610": "jersey, T-shirt, tee shirt",
+    "611": "jigsaw puzzle",
+    "612": "jinrikisha, ricksha, rickshaw",
+    "613": "joystick",
+    "614": "kimono",
+    "615": "knee pad",
+    "616": "knot",
+    "617": "lab coat, laboratory coat",
+    "618": "ladle",
+    "619": "lampshade, lamp shade",
+    "620": "laptop, laptop computer",
+    "621": "lawn mower, mower",
+    "622": "lens cap, lens cover",
+    "623": "letter opener, paper knife, paperknife",
+    "624": "library",
+    "625": "lifeboat",
+    "626": "lighter, light, igniter, ignitor",
+    "627": "limousine, limo",
+    "628": "liner, ocean liner",
+    "629": "lipstick, lip rouge",
+    "630": "Loafer",
+    "631": "lotion",
+    "632": "loudspeaker, speaker, speaker unit, loudspeaker system, speaker system",
+    "633": "loupe, jewelers loupe",
+    "634": "lumbermill, sawmill",
+    "635": "magnetic compass",
+    "636": "mailbag, postbag",
+    "637": "mailbox, letter box",
+    "638": "maillot",
+    "639": "maillot, tank suit",
+    "640": "manhole cover",
+    "641": "maraca",
+    "642": "marimba, xylophone",
+    "643": "mask",
+    "644": "matchstick",
+    "645": "maypole",
+    "646": "maze, labyrinth",
+    "647": "measuring cup",
+    "648": "medicine chest, medicine cabinet",
+    "649": "megalith, megalithic structure",
+    "650": "microphone, mike",
+    "651": "microwave, microwave oven",
+    "652": "military uniform",
+    "653": "milk can",
+    "654": "minibus",
+    "655": "miniskirt, mini",
+    "656": "minivan",
+    "657": "missile",
+    "658": "mitten",
+    "659": "mixing bowl",
+    "660": "mobile home, manufactured home",
+    "661": "Model T",
+    "662": "modem",
+    "663": "monastery",
+    "664": "monitor",
+    "665": "moped",
+    "666": "mortar",
+    "667": "mortarboard",
+    "668": "mosque",
+    "669": "mosquito net",
+    "670": "motor scooter, scooter",
+    "671": "mountain bike, all-terrain bike, off-roader",
+    "672": "mountain tent",
+    "673": "mouse, computer mouse",
+    "674": "mousetrap",
+    "675": "moving van",
+    "676": "muzzle",
+    "677": "nail",
+    "678": "neck brace",
+    "679": "necklace",
+    "680": "nipple",
+    "681": "notebook, notebook computer",
+    "682": "obelisk",
+    "683": "oboe, hautboy, hautbois",
+    "684": "ocarina, sweet potato",
+    "685": "odometer, hodometer, mileometer, milometer",
+    "686": "oil filter",
+    "687": "organ, pipe organ",
+    "688": "oscilloscope, scope, cathode-ray oscilloscope, CRO",
+    "689": "overskirt",
+    "690": "oxcart",
+    "691": "oxygen mask",
+    "692": "packet",
+    "693": "paddle, boat paddle",
+    "694": "paddlewheel, paddle wheel",
+    "695": "padlock",
+    "696": "paintbrush",
+    "697": "pajama, pyjama, pjs, jammies",
+    "698": "palace",
+    "699": "panpipe, pandean pipe, syrinx",
+    "700": "paper towel",
+    "701": "parachute, chute",
+    "702": "parallel bars, bars",
+    "703": "park bench",
+    "704": "parking meter",
+    "705": "passenger car, coach, carriage",
+    "706": "patio, terrace",
+    "707": "pay-phone, pay-station",
+    "708": "pedestal, plinth, footstall",
+    "709": "pencil box, pencil case",
+    "710": "pencil sharpener",
+    "711": "perfume, essence",
+    "712": "Petri dish",
+    "713": "photocopier",
+    "714": "pick, plectrum, plectron",
+    "715": "pickelhaube",
+    "716": "picket fence, paling",
+    "717": "pickup, pickup truck",
+    "718": "pier",
+    "719": "piggy bank, penny bank",
+    "720": "pill bottle",
+    "721": "pillow",
+    "722": "ping-pong ball",
+    "723": "pinwheel",
+    "724": "pirate, pirate ship",
+    "725": "pitcher, ewer",
+    "726": "plane, carpenters plane, woodworking plane",
+    "727": "planetarium",
+    "728": "plastic bag",
+    "729": "plate rack",
+    "730": "plow, plough",
+    "731": "plunger, plumbers helper",
+    "732": "Polaroid camera, Polaroid Land camera",
+    "733": "pole",
+    "734": "police van, police wagon, paddy wagon, patrol wagon, wagon, black Maria",
+    "735": "poncho",
+    "736": "pool table, billiard table, snooker table",
+    "737": "pop bottle, soda bottle",
+    "738": "pot, flowerpot",
+    "739": "potters wheel",
+    "740": "power drill",
+    "741": "prayer rug, prayer mat",
+    "742": "printer",
+    "743": "prison, prison house",
+    "744": "projectile, missile",
+    "745": "projector",
+    "746": "puck, hockey puck",
+    "747": "punching bag, punch bag, punching ball, punchball",
+    "748": "purse",
+    "749": "quill, quill pen",
+    "750": "quilt, comforter, comfort, puff",
+    "751": "racer, race car, racing car",
+    "752": "racket, racquet",
+    "753": "radiator",
+    "754": "radio, wireless",
+    "755": "radio telescope, radio reflector",
+    "756": "rain barrel",
+    "757": "recreational vehicle, RV, R.V.",
+    "758": "reel",
+    "759": "reflex camera",
+    "760": "refrigerator, icebox",
+    "761": "remote control, remote",
+    "762": "restaurant, eating house, eating place, eatery",
+    "763": "revolver, six-gun, six-shooter",
+    "764": "rifle",
+    "765": "rocking chair, rocker",
+    "766": "rotisserie",
+    "767": "rubber eraser, rubber, pencil eraser",
+    "768": "rugby ball",
+    "769": "rule, ruler",
+    "770": "running shoe",
+    "771": "safe",
+    "772": "safety pin",
+    "773": "saltshaker, salt shaker",
+    "774": "sandal",
+    "775": "sarong",
+    "776": "sax, saxophone",
+    "777": "scabbard",
+    "778": "scale, weighing machine",
+    "779": "school bus",
+    "780": "schooner",
+    "781": "scoreboard",
+    "782": "screen, CRT screen",
+    "783": "screw",
+    "784": "screwdriver",
+    "785": "seat belt, seatbelt",
+    "786": "sewing machine",
+    "787": "shield, buckler",
+    "788": "shoe shop, shoe-shop, shoe store",
+    "789": "shoji",
+    "790": "shopping basket",
+    "791": "shopping cart",
+    "792": "shovel",
+    "793": "shower cap",
+    "794": "shower curtain",
+    "795": "ski",
+    "796": "ski mask",
+    "797": "sleeping bag",
+    "798": "slide rule, slipstick",
+    "799": "sliding door",
+    "800": "slot, one-armed bandit",
+    "801": "snorkel",
+    "802": "snowmobile",
+    "803": "snowplow, snowplough",
+    "804": "soap dispenser",
+    "805": "soccer ball",
+    "806": "sock",
+    "807": "solar dish, solar collector, solar furnace",
+    "808": "sombrero",
+    "809": "soup bowl",
+    "810": "space bar",
+    "811": "space heater",
+    "812": "space shuttle",
+    "813": "spatula",
+    "814": "speedboat",
+    "815": "spider web, spiders web",
+    "816": "spindle",
+    "817": "sports car, sport car",
+    "818": "spotlight, spot",
+    "819": "stage",
+    "820": "steam locomotive",
+    "821": "steel arch bridge",
+    "822": "steel drum",
+    "823": "stethoscope",
+    "824": "stole",
+    "825": "stone wall",
+    "826": "stopwatch, stop watch",
+    "827": "stove",
+    "828": "strainer",
+    "829": "streetcar, tram, tramcar, trolley, trolley car",
+    "830": "stretcher",
+    "831": "studio couch, day bed",
+    "832": "stupa, tope",
+    "833": "submarine, pigboat, sub, U-boat",
+    "834": "suit, suit of clothes",
+    "835": "sundial",
+    "836": "sunglass",
+    "837": "sunglasses, dark glasses, shades",
+    "838": "sunscreen, sunblock, sun blocker",
+    "839": "suspension bridge",
+    "840": "swab, swob, mop",
+    "841": "sweatshirt",
+    "842": "swimming trunks, bathing trunks",
+    "843": "swing",
+    "844": "switch, electric switch, electrical switch",
+    "845": "syringe",
+    "846": "table lamp",
+    "847": "tank, army tank, armored combat vehicle, armoured combat vehicle",
+    "848": "tape player",
+    "849": "teapot",
+    "850": "teddy, teddy bear",
+    "851": "television, television system",
+    "852": "tennis ball",
+    "853": "thatch, thatched roof",
+    "854": "theater curtain, theatre curtain",
+    "855": "thimble",
+    "856": "thresher, thrasher, threshing machine",
+    "857": "throne",
+    "858": "tile roof",
+    "859": "toaster",
+    "860": "tobacco shop, tobacconist shop, tobacconist",
+    "861": "toilet seat",
+    "862": "torch",
+    "863": "totem pole",
+    "864": "tow truck, tow car, wrecker",
+    "865": "toyshop",
+    "866": "tractor",
+    "867": "trailer truck, tractor trailer, trucking rig, rig, articulated lorry, semi",
+    "868": "tray",
+    "869": "trench coat",
+    "870": "tricycle, trike, velocipede",
+    "871": "trimaran",
+    "872": "tripod",
+    "873": "triumphal arch",
+    "874": "trolleybus, trolley coach, trackless trolley",
+    "875": "trombone",
+    "876": "tub, vat",
+    "877": "turnstile",
+    "878": "typewriter keyboard",
+    "879": "umbrella",
+    "880": "unicycle, monocycle",
+    "881": "upright, upright piano",
+    "882": "vacuum, vacuum cleaner",
+    "883": "vase",
+    "884": "vault",
+    "885": "velvet",
+    "886": "vending machine",
+    "887": "vestment",
+    "888": "viaduct",
+    "889": "violin, fiddle",
+    "890": "volleyball",
+    "891": "waffle iron",
+    "892": "wall clock",
+    "893": "wallet, billfold, notecase, pocketbook",
+    "894": "wardrobe, closet, press",
+    "895": "warplane, military plane",
+    "896": "washbasin, handbasin, washbowl, lavabo, wash-hand basin",
+    "897": "washer, automatic washer, washing machine",
+    "898": "water bottle",
+    "899": "water jug",
+    "900": "water tower",
+    "901": "whiskey jug",
+    "902": "whistle",
+    "903": "wig",
+    "904": "window screen",
+    "905": "window shade",
+    "906": "Windsor tie",
+    "907": "wine bottle",
+    "908": "wing",
+    "909": "wok",
+    "910": "wooden spoon",
+    "911": "wool, woolen, woollen",
+    "912": "worm fence, snake fence, snake-rail fence, Virginia fence",
+    "913": "wreck",
+    "914": "yawl",
+    "915": "yurt",
+    "916": "web site, website, internet site, site",
+    "917": "comic book",
+    "918": "crossword puzzle, crossword",
+    "919": "street sign",
+    "920": "traffic light, traffic signal, stoplight",
+    "921": "book jacket, dust cover, dust jacket, dust wrapper",
+    "922": "menu",
+    "923": "plate",
+    "924": "guacamole",
+    "925": "consomme",
+    "926": "hot pot, hotpot",
+    "927": "trifle",
+    "928": "ice cream, icecream",
+    "929": "ice lolly, lolly, lollipop, popsicle",
+    "930": "French loaf",
+    "931": "bagel, beigel",
+    "932": "pretzel",
+    "933": "cheeseburger",
+    "934": "hotdog, hot dog, red hot",
+    "935": "mashed potato",
+    "936": "head cabbage",
+    "937": "broccoli",
+    "938": "cauliflower",
+    "939": "zucchini, courgette",
+    "940": "spaghetti squash",
+    "941": "acorn squash",
+    "942": "butternut squash",
+    "943": "cucumber, cuke",
+    "944": "artichoke, globe artichoke",
+    "945": "bell pepper",
+    "946": "cardoon",
+    "947": "mushroom",
+    "948": "Granny Smith",
+    "949": "strawberry",
+    "950": "orange",
+    "951": "lemon",
+    "952": "fig",
+    "953": "pineapple, ananas",
+    "954": "banana",
+    "955": "jackfruit, jak, jack",
+    "956": "custard apple",
+    "957": "pomegranate",
+    "958": "hay",
+    "959": "carbonara",
+    "960": "chocolate sauce, chocolate syrup",
+    "961": "dough",
+    "962": "meat loaf, meatloaf",
+    "963": "pizza, pizza pie",
+    "964": "potpie",
+    "965": "burrito",
+    "966": "red wine",
+    "967": "espresso",
+    "968": "cup",
+    "969": "eggnog",
+    "970": "alp",
+    "971": "bubble",
+    "972": "cliff, drop, drop-off",
+    "973": "coral reef",
+    "974": "geyser",
+    "975": "lakeside, lakeshore",
+    "976": "promontory, headland, head, foreland",
+    "977": "sandbar, sand bar",
+    "978": "seashore, coast, seacoast, sea-coast",
+    "979": "valley, vale",
+    "980": "volcano",
+    "981": "ballplayer, baseball player",
+    "982": "groom, bridegroom",
+    "983": "scuba diver",
+    "984": "rapeseed",
+    "985": "daisy",
+    "986": "yellow ladys slipper, yellow lady-slipper, Cypripedium calceolus, Cypripedium parviflorum",
+    "987": "corn",
+    "988": "acorn",
+    "989": "hip, rose hip, rosehip",
+    "990": "buckeye, horse chestnut, conker",
+    "991": "coral fungus",
+    "992": "agaric",
+    "993": "gyromitra",
+    "994": "stinkhorn, carrion fungus",
+    "995": "earthstar",
+    "996": "hen-of-the-woods, hen of the woods, Polyporus frondosus, Grifola frondosa",
+    "997": "bolete",
+    "998": "ear, spike, capitulum",
+    "999": "toilet tissue, toilet paper, bathroom tissue"
+  }
+}

NiT-S/pipeline.py

@@ -0,0 +1,483 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils.torch_utils import randn_tensor
+
+# Local component classes are loaded dynamically in from_pretrained.
+
+DEFAULT_NATIVE_RESOLUTION = 256
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> from pathlib import Path
+        >>> import torch
+        >>> from diffusers import DiffusionPipeline
+
+        >>> model_dir = Path("./NiT-S").resolve()
+        >>> pipe = DiffusionPipeline.from_pretrained(
+        ...     str(model_dir),
+        ...     local_files_only=True,
+        ...     custom_pipeline=str(model_dir / "pipeline.py"),
+        ...     trust_remote_code=True,
+        ...     torch_dtype=torch.bfloat16,
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> print(pipe.id2label[207])
+        >>> print(pipe.get_label_ids("golden retriever"))
+
+        >>> generator = torch.Generator(device="cuda").manual_seed(42)
+        >>> image = pipe(
+        ...     class_labels="golden retriever",
+        ...     height=256,
+        ...     width=256,
+        ...     num_inference_steps=250,
+        ...     guidance_scale=2.25,
+        ...     guidance_interval=(0.0, 0.7),
+        ...     generator=generator,
+        ... ).images[0]
+        >>> image.save("demo.png")
+        ```
+"""
+
+
+class NiTPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for native-resolution class-conditional image generation with NiT.
+
+    Uses the native [`FlowMatchEulerDiscreteScheduler`] in deterministic (ODE) mode.
+    The official NiT repo defaults to an Euler-Maruyama SDE sampler for 512×512; that SDE is
+    not the same as the scheduler's `stochastic_sampling` path, so keep
+    `scheduler.config.stochastic_sampling=False` and let the scheduler perform the ODE update
+    `x_{t+dt} = x_t + dt * v`.
+
+    Parameters:
+        transformer ([`NiTTransformer2DModel`]):
+            Class-conditional transformer that predicts flow-matching velocity in packed latent space.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            Native diffusers flow-matching Euler scheduler (`stochastic_sampling=False`).
+        vae ([`AutoencoderDC`] or [`AutoencoderKL`], *optional*):
+            Variational autoencoder used to decode packed transformer latents to pixels.
+        id2label (`dict[int, str]`, *optional*):
+            ImageNet class id to English label mapping. Values may contain comma-separated synonyms.
+    """
+
+    model_cpu_offload_seq = "transformer->vae"
+    _optional_components = ["vae"]
+
+    def __init__(
+        self,
+        transformer,
+        scheduler,
+        vae=None,
+        id2label: Optional[Dict[Union[int, str], str]] = None,
+    ):
+        super().__init__()
+        self.register_modules(transformer=transformer, scheduler=scheduler, vae=vae)
+        self.image_processor = VaeImageProcessor()
+        self._id2label = self._normalize_id2label(id2label)
+        self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = bool(self._id2label)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path=None, subfolder=None, **kwargs):
+        """Load a self-contained variant folder locally or from the Hub."""
+        import importlib
+        import sys
+
+        repo_root = Path(__file__).resolve().parent
+
+        if pretrained_model_name_or_path in (None, "", "."):
+            variant = repo_root
+        elif (
+            isinstance(pretrained_model_name_or_path, str)
+            and "/" in pretrained_model_name_or_path
+            and not Path(pretrained_model_name_or_path).exists()
+        ):
+            from huggingface_hub import snapshot_download
+
+            hub_kwargs = dict(kwargs.pop("hub_kwargs", {}))
+            if subfolder:
+                hub_kwargs.setdefault("allow_patterns", [f"{subfolder}/**"])
+            cache_dir = snapshot_download(pretrained_model_name_or_path, **hub_kwargs)
+            variant = Path(cache_dir) / subfolder if subfolder else Path(cache_dir)
+        else:
+            variant = Path(pretrained_model_name_or_path)
+            if not variant.is_absolute():
+                candidate = (Path.cwd() / variant).resolve()
+                variant = candidate if candidate.exists() else (repo_root / variant).resolve()
+            if subfolder:
+                variant = variant / subfolder
+
+        id2label_override = kwargs.pop("id2label", None)
+        model_kwargs = dict(kwargs)
+        inserted: List[str] = []
+
+        def _load_component(folder: str, module_name: str, class_name: str):
+            comp_dir = variant / folder
+            module_path = comp_dir / f"{module_name}.py"
+            has_weights = (comp_dir / "config.json").exists() or (comp_dir / "scheduler_config.json").exists()
+            if not module_path.exists() or not has_weights:
+                return None
+
+            comp_path = str(comp_dir)
+            if comp_path not in sys.path:
+                sys.path.insert(0, comp_path)
+                inserted.append(comp_path)
+
+            module = importlib.import_module(module_name)
+            component_cls = getattr(module, class_name)
+            return component_cls.from_pretrained(str(comp_dir), **model_kwargs)
+
+        try:
+            transformer = _load_component("transformer", "nit_transformer_2d", "NiTTransformer2DModel")
+            try:
+                scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(str(variant), subfolder="scheduler")
+            except Exception:
+                scheduler = FlowMatchEulerDiscreteScheduler(
+                    num_train_timesteps=1000,
+                    shift=1.0,
+                    stochastic_sampling=False,
+                )
+            if transformer is None:
+                raise ValueError(f"No loadable transformer found under {variant}")
+
+            vae = None
+            vae_dir = variant / "vae"
+            if vae_dir.exists() and (vae_dir / "config.json").exists():
+                from diffusers import AutoencoderDC, AutoencoderKL
+
+                vae_class_name = json.loads((vae_dir / "config.json").read_text(encoding="utf-8")).get(
+                    "_class_name", "AutoencoderDC"
+                )
+                vae_cls = AutoencoderDC if vae_class_name == "AutoencoderDC" else AutoencoderKL
+                vae = vae_cls.from_pretrained(str(vae_dir), **model_kwargs)
+
+            id2label = id2label_override or cls._read_id2label_from_model_index(str(variant))
+            pipe = cls(
+                transformer=transformer,
+                scheduler=scheduler,
+                vae=vae,
+                id2label=id2label,
+            )
+            if hasattr(pipe, "register_to_config"):
+                pipe.register_to_config(_name_or_path=str(variant))
+            return pipe
+        finally:
+            for comp_path in inserted:
+                if comp_path in sys.path:
+                    sys.path.remove(comp_path)
+
+    def _ensure_labels_loaded(self) -> None:
+        if self._labels_loaded_from_model_index:
+            return
+        loaded = self._read_id2label_from_model_index(getattr(self.config, "_name_or_path", None))
+        if loaded:
+            self._id2label = loaded
+            self.labels = self._build_label2id(self._id2label)
+        self._labels_loaded_from_model_index = True
+
+    @staticmethod
+    def _normalize_id2label(id2label: Optional[Dict[Union[int, str], str]]) -> Dict[int, str]:
+        if not id2label:
+            return {}
+        return {int(key): value for key, value in id2label.items()}
+
+    @staticmethod
+    def _read_id2label_from_model_index(variant_path: Optional[str]) -> Dict[int, str]:
+        if not variant_path:
+            return {}
+        variant_dir = Path(variant_path).resolve()
+        model_index_path = variant_dir / "model_index.json"
+        if not model_index_path.exists():
+            return {}
+        raw = json.loads(model_index_path.read_text(encoding="utf-8"))
+        id2label = raw.get("id2label")
+        if not isinstance(id2label, dict):
+            return {}
+        return {int(key): value for key, value in id2label.items()}
+
+    @staticmethod
+    def _build_label2id(id2label: Dict[int, str]) -> Dict[str, int]:
+        label2id: Dict[str, int] = {}
+        for class_id, value in id2label.items():
+            for synonym in value.split(","):
+                synonym = synonym.strip()
+                if synonym:
+                    label2id[synonym] = int(class_id)
+        return dict(sorted(label2id.items()))
+
+    @property
+    def id2label(self) -> Dict[int, str]:
+        r"""ImageNet class id to English label string (comma-separated synonyms)."""
+        self._ensure_labels_loaded()
+        return self._id2label
+
+    def get_label_ids(self, label: Union[str, List[str]]) -> List[int]:
+        r"""
+        Map ImageNet label strings to class ids.
+
+        Args:
+            label (`str` or `list[str]`):
+                One or more English label strings. Each string must match a synonym in `id2label`.
+        """
+        self._ensure_labels_loaded()
+        label2id = self.labels
+        if not label2id:
+            raise ValueError("No English labels loaded. Ensure `id2label` exists in model_index.json.")
+
+        if isinstance(label, str):
+            label = [label]
+
+        missing = [item for item in label if item not in label2id]
+        if missing:
+            preview = ", ".join(list(label2id.keys())[:8])
+            raise ValueError(f"Unknown English label(s): {missing}. Example valid labels: {preview}, ...")
+        return [label2id[item] for item in label]
+
+    def _normalize_class_labels(
+        self,
+        class_labels: Union[int, str, List[Union[int, str]], torch.LongTensor],
+    ) -> torch.LongTensor:
+        if torch.is_tensor(class_labels):
+            return class_labels.to(device=self._execution_device, dtype=torch.long).reshape(-1)
+
+        if isinstance(class_labels, int):
+            class_label_ids = [class_labels]
+        elif isinstance(class_labels, str):
+            class_label_ids = self.get_label_ids(class_labels)
+        elif class_labels and isinstance(class_labels[0], str):
+            class_label_ids = self.get_label_ids(class_labels)
+        else:
+            class_label_ids = list(class_labels)
+
+        return torch.tensor(class_label_ids, device=self._execution_device, dtype=torch.long).reshape(-1)
+
+    def _get_vae_spatial_downsample(self) -> int:
+        if self.vae is None:
+            return 1
+        if self.vae.__class__.__name__ == "AutoencoderDC" or "dc-ae" in getattr(
+            self.vae.config, "_name_or_path", ""
+        ):
+            return 32
+        block_out_channels = getattr(self.vae.config, "block_out_channels", [0, 0, 0, 0])
+        return 2 ** (len(block_out_channels) - 1)
+
+    def check_inputs(
+        self,
+        height: int,
+        width: int,
+        num_inference_steps: int,
+        output_type: str,
+    ) -> None:
+        if num_inference_steps < 1:
+            raise ValueError("num_inference_steps must be >= 1.")
+        if output_type not in {"pil", "np", "pt", "latent"}:
+            raise ValueError("output_type must be one of: 'pil', 'np', 'pt', 'latent'.")
+
+        spatial_downsample = self._get_vae_spatial_downsample()
+        if height % spatial_downsample != 0 or width % spatial_downsample != 0:
+            raise ValueError(
+                f"height and width must be divisible by the VAE downsample factor {spatial_downsample}."
+            )
+
+        patch_size = int(self.transformer.config.patch_size)
+        latent_height = height // spatial_downsample
+        latent_width = width // spatial_downsample
+        if latent_height % patch_size != 0 or latent_width % patch_size != 0:
+            raise ValueError("Latent height and width must be divisible by transformer's patch_size.")
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]],
+    ) -> Tuple[torch.Tensor, torch.LongTensor]:
+        spatial_downsample = self._get_vae_spatial_downsample()
+        latent_height = height // spatial_downsample
+        latent_width = width // spatial_downsample
+        patch_size = int(self.transformer.config.patch_size)
+        token_height = latent_height // patch_size
+        token_width = latent_width // patch_size
+        image_sizes = torch.tensor([[token_height, token_width]] * batch_size, device=device, dtype=torch.long)
+
+        packed_shape = (
+            batch_size * token_height * token_width,
+            self.transformer.config.in_channels,
+            patch_size,
+            patch_size,
+        )
+        packed_latents = randn_tensor(
+            packed_shape,
+            generator=generator,
+            device=device,
+            dtype=dtype,
+        )
+        return packed_latents, image_sizes
+
+    @staticmethod
+    def _flow_time_from_scheduler_timestep(timestep: torch.Tensor, num_train_timesteps: int) -> float:
+        """Map native scheduler timesteps (sigma * num_train_timesteps) to NiT flow time in [0, 1]."""
+        return float(timestep) / num_train_timesteps
+
+    def _apply_classifier_free_guidance(
+        self,
+        model_output: torch.Tensor,
+        guidance_scale: float,
+        guidance_active: bool,
+    ) -> torch.Tensor:
+        if guidance_scale <= 1.0 or not guidance_active:
+            return model_output
+        model_output_cond, model_output_uncond = model_output.chunk(2)
+        return model_output_uncond + guidance_scale * (model_output_cond - model_output_uncond)
+
+    def _get_vae_dtype(self, latents: torch.Tensor) -> torch.dtype:
+        vae_dtype = getattr(self.vae, "dtype", None)
+        if vae_dtype is not None:
+            return vae_dtype
+        vae_params = next(self.vae.parameters(), None)
+        return vae_params.dtype if vae_params is not None else latents.dtype
+
+    def decode_latents(self, latents: torch.Tensor, output_type: str = "pil"):
+        if self.vae is None:
+            if output_type == "latent":
+                return latents
+            raise ValueError("Cannot decode latents without a VAE.")
+
+        vae_dtype = self._get_vae_dtype(latents)
+        latents = latents.to(dtype=vae_dtype)
+        scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
+        shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
+        latents = (latents / scaling_factor) + shift_factor
+        if output_type == "latent":
+            return latents
+
+        image = self.vae.decode(latents, return_dict=False)[0]
+        return self.image_processor.postprocess(image, output_type=output_type)
+
+    @torch.inference_mode()
+    def __call__(
+        self,
+        class_labels: Union[int, str, List[Union[int, str]], torch.LongTensor],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 1.0,
+        guidance_interval: Tuple[float, float] = (0.0, 1.0),
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: str = "pil",
+        return_dict: bool = True,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        r"""
+        Generate class-conditional images at native resolution.
+
+        Args:
+            class_labels (`int`, `str`, `list[int]`, `list[str]`, or `torch.LongTensor`):
+                ImageNet class indices or human-readable English label strings.
+            height (`int`, *optional*):
+                Output image height in pixels. Defaults to `512` when a VAE is present.
+            width (`int`, *optional*):
+                Output image width in pixels. Defaults to `512` when a VAE is present.
+            num_inference_steps (`int`, defaults to `50`):
+                Number of denoising steps.
+            guidance_scale (`float`, defaults to `1.0`):
+                Classifier-free guidance scale. CFG is active when `guidance_scale > 1.0`.
+            guidance_interval (`tuple[float, float]`, defaults to `(0.0, 1.0)`):
+                Flow-time interval where CFG is applied. Uses continuous flow time
+                `timestep / num_train_timesteps`, matching the official NiT ODE sampler.
+            generator (`torch.Generator`, *optional*):
+                RNG for reproducibility.
+            output_type (`str`, defaults to `"pil"`):
+                `"pil"`, `"np"`, `"pt"`, or `"latent"`.
+            return_dict (`bool`, defaults to `True`):
+                Return [`ImagePipelineOutput`] if True.
+        """
+        default_size = DEFAULT_NATIVE_RESOLUTION if self.vae is not None else 256
+        height = int(height or default_size)
+        width = int(width or default_size)
+        self.check_inputs(height, width, num_inference_steps, output_type)
+
+        if getattr(self.scheduler.config, "stochastic_sampling", False):
+            raise ValueError(
+                "NiT expects deterministic FlowMatchEulerDiscreteScheduler stepping "
+                "(scheduler.config.stochastic_sampling=False). The scheduler's stochastic_sampling "
+                "path uses a different update rule than the official NiT Euler-Maruyama SDE and "
+                "produces salt-and-pepper noise."
+            )
+
+        device = self._execution_device
+        model_dtype = next(self.transformer.parameters()).dtype
+        class_labels_tensor = self._normalize_class_labels(class_labels)
+        batch_size = class_labels_tensor.numel()
+
+        packed_latents, image_sizes = self.prepare_latents(
+            batch_size, height, width, model_dtype, device, generator
+        )
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        num_train_timesteps = self.scheduler.config.num_train_timesteps
+
+        null_labels = torch.full_like(class_labels_tensor, self.transformer.config.num_classes)
+        guidance_low, guidance_high = guidance_interval
+
+        for t in self.progress_bar(self.scheduler.timesteps):
+            flow_time = self._flow_time_from_scheduler_timestep(t, num_train_timesteps)
+            guidance_active = guidance_low <= flow_time <= guidance_high
+            if guidance_scale > 1.0 and guidance_active:
+                model_input = torch.cat([packed_latents, packed_latents], dim=0)
+                labels = torch.cat([class_labels_tensor, null_labels], dim=0)
+                model_image_sizes = torch.cat([image_sizes, image_sizes], dim=0)
+            else:
+                model_input = packed_latents
+                labels = class_labels_tensor
+                model_image_sizes = image_sizes
+
+            timestep_batch = torch.full((labels.numel(),), flow_time, device=device, dtype=model_dtype)
+            model_output = self.transformer(
+                model_input.to(dtype=model_dtype),
+                timestep_batch,
+                labels,
+                image_sizes=model_image_sizes,
+                return_dict=True,
+            ).sample
+            model_output = self._apply_classifier_free_guidance(model_output, guidance_scale, guidance_active)
+            packed_latents = self.scheduler.step(
+                model_output,
+                t,
+                packed_latents,
+                generator=generator,
+            ).prev_sample
+
+        latents = self.transformer._unpack_latents(packed_latents, image_sizes)
+        image = self.decode_latents(latents, output_type=output_type)
+
+        self.maybe_free_model_hooks()
+        if not return_dict:
+            return (image,)
+        return ImagePipelineOutput(images=image)
+
+
+NiTPipelineOutput = ImagePipelineOutput

NiT-S/scheduler/scheduler_config.json

@@ -0,0 +1,7 @@
+{
+  "_class_name": "FlowMatchEulerDiscreteScheduler",
+  "_diffusers_version": "0.36.0",
+  "num_train_timesteps": 1000,
+  "shift": 1.0,
+  "stochastic_sampling": false
+}

NiT-S/transformer/config.json

@@ -0,0 +1,16 @@
+{
+  "_class_name": "NiTTransformer2DModel",
+  "class_dropout_prob": 0.1,
+  "depth": 12,
+  "encoder_depth": 4,
+  "hidden_size": 384,
+  "in_channels": 32,
+  "input_size": 32,
+  "num_classes": 1000,
+  "num_heads": 6,
+  "patch_size": 1,
+  "projector_dim": 768,
+  "qk_norm": true,
+  "use_checkpoint": false,
+  "z_dim": 768
+}

NiT-S/transformer/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15681510284ed290f7f9b2979909d364a64a4c57f7ed8e1474a8574d8bc81a09
+size 137422176

NiT-S/transformer/nit_transformer_2d.py

@@ -0,0 +1,471 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+
+from dataclasses import dataclass
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    from diffusers.configuration_utils import ConfigMixin, register_to_config
+    from diffusers.models.modeling_utils import ModelMixin
+    from diffusers.utils import BaseOutput
+except Exception:  # pragma: no cover - lets this subtree be tested outside diffusers.
+    class BaseOutput(dict):
+        def __post_init__(self):
+            self.update(self.__dict__)
+
+    class _Config(dict):
+        def __getattr__(self, key):
+            try:
+                return self[key]
+            except KeyError as error:
+                raise AttributeError(key) from error
+
+    class ConfigMixin:
+        config_name = "config.json"
+
+    class ModelMixin(nn.Module):
+        pass
+
+    def register_to_config(init):
+        def wrapper(self, *args, **kwargs):
+            import inspect
+
+            signature = inspect.signature(init)
+            bound = signature.bind(self, *args, **kwargs)
+            bound.apply_defaults()
+            self.config = _Config({key: value for key, value in bound.arguments.items() if key != "self"})
+            init(self, *args, **kwargs)
+
+        return wrapper
+
+
+try:
+    from flash_attn import flash_attn_varlen_func
+except Exception:  # pragma: no cover - optional acceleration.
+    flash_attn_varlen_func = None
+
+
+@dataclass
+class NiTTransformer2DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+    projection_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+
+
+def _modulate(hidden_states: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
+    return hidden_states * (1 + scale) + shift
+
+
+def _rotate_half(hidden_states: torch.Tensor) -> torch.Tensor:
+    hidden_states = hidden_states.reshape(*hidden_states.shape[:-1], -1, 2)
+    hidden_states_1, hidden_states_2 = hidden_states.unbind(dim=-1)
+    return torch.stack((-hidden_states_2, hidden_states_1), dim=-1).flatten(-2)
+
+
+def _get_float_dtype_or_default(tensor: Optional[torch.Tensor] = None) -> torch.dtype:
+    if tensor is not None and tensor.is_floating_point():
+        return tensor.dtype
+    return torch.get_default_dtype()
+
+
+class NiTPatchEmbed(nn.Module):
+    def __init__(self, patch_size: int, in_channels: int, hidden_size: int):
+        super().__init__()
+        self.patch_size = (patch_size, patch_size)
+        self.proj = nn.Conv2d(in_channels, hidden_size, kernel_size=patch_size, stride=patch_size, bias=True)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.proj(hidden_states)
+        return hidden_states.flatten(2).transpose(1, 2)
+
+
+class NiTTimestepEmbedder(nn.Module):
+    def __init__(self, hidden_size: int, frequency_embedding_size: int = 256):
+        super().__init__()
+        self.frequency_embedding_size = frequency_embedding_size
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+
+    @staticmethod
+    def get_timestep_embedding(timesteps: torch.Tensor, embedding_dim: int, max_period: int = 10000):
+        half = embedding_dim // 2
+        # Keep sinusoid construction in fp32 to mirror native NiT behavior.
+        exponent = -math.log(max_period) * torch.arange(half, dtype=torch.float32, device=timesteps.device) / half
+        freqs = torch.exp(exponent)
+        args = timesteps.float()[:, None] * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if embedding_dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        timestep_freq = self.get_timestep_embedding(timesteps, self.frequency_embedding_size).to(timesteps.dtype)
+        return self.mlp(timestep_freq)
+
+
+class NiTLabelEmbedder(nn.Module):
+    def __init__(self, num_classes: int, hidden_size: int, dropout_prob: float):
+        super().__init__()
+        use_cfg_embedding = dropout_prob > 0
+        self.embedding_table = nn.Embedding(num_classes + int(use_cfg_embedding), hidden_size)
+        self.num_classes = num_classes
+        self.dropout_prob = dropout_prob
+
+    def forward(self, class_labels: torch.LongTensor) -> torch.Tensor:
+        return self.embedding_table(class_labels)
+
+
+class NiTRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        head_dim: int,
+        custom_freqs: str = "normal",
+        theta: int = 10000,
+        max_cached_len: int = 1024,
+        max_pe_len_h: Optional[int] = None,
+        max_pe_len_w: Optional[int] = None,
+        decouple: bool = False,
+        ori_max_pe_len: Optional[int] = None,
+    ):
+        super().__init__()
+        del max_pe_len_h, max_pe_len_w, decouple, ori_max_pe_len
+        if custom_freqs not in {"normal", "scale1", "scale2"}:
+            raise ValueError(
+                "This Diffusers implementation supports the trained RoPE frequencies directly. "
+                "Checkpoint conversion preserves weights; extrapolation variants should be handled "
+                "by changing the model config before loading."
+            )
+        dim = head_dim // 2
+        if dim % 2 != 0:
+            raise ValueError("NiT rotary embedding requires head_dim // 2 to be even.")
+        default_dtype = _get_float_dtype_or_default()
+        freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=default_dtype) / dim))
+        self.register_buffer("freqs_h", freqs, persistent=False)
+        self.register_buffer("freqs_w", freqs.clone(), persistent=False)
+        positions = torch.arange(max_cached_len, dtype=default_dtype)
+        freqs_h_cached = torch.einsum("n,f->nf", positions, self.freqs_h).repeat_interleave(2, dim=-1)
+        freqs_w_cached = torch.einsum("n,f->nf", positions, self.freqs_w).repeat_interleave(2, dim=-1)
+        self.register_buffer("freqs_h_cached", freqs_h_cached, persistent=False)
+        self.register_buffer("freqs_w_cached", freqs_w_cached, persistent=False)
+
+    def forward(self, image_sizes: torch.LongTensor, device: torch.device) -> Tuple[torch.Tensor, torch.Tensor]:
+        grids = []
+        for height, width in image_sizes.tolist():
+            # Match native NiT token ordering for RoPE alignment.
+            grid_h = torch.arange(height, device=device)
+            grid_w = torch.arange(width, device=device)
+            grid = torch.meshgrid(grid_h, grid_w, indexing="xy")
+            grids.append(torch.stack(grid, dim=0).reshape(2, -1))
+        grid = torch.cat(grids, dim=1)
+        freqs_h = self.freqs_h_cached.to(device)[grid[0]]
+        freqs_w = self.freqs_w_cached.to(device)[grid[1]]
+        freqs = torch.cat([freqs_h, freqs_w], dim=-1)
+        return freqs.cos().unsqueeze(1), freqs.sin().unsqueeze(1)
+
+
+class NiTAttention(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, qk_norm: bool = False):
+        super().__init__()
+        if hidden_size % num_heads != 0:
+            raise ValueError("hidden_size must be divisible by num_heads")
+        self.num_heads = num_heads
+        self.head_dim = hidden_size // num_heads
+        self.qkv = nn.Linear(hidden_size, hidden_size * 3, bias=True)
+        self.q_norm = nn.LayerNorm(self.head_dim) if qk_norm else nn.Identity()
+        self.k_norm = nn.LayerNorm(self.head_dim) if qk_norm else nn.Identity()
+        self.proj = nn.Linear(hidden_size, hidden_size)
+        self.proj_drop = nn.Dropout(0.0)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.IntTensor,
+        freqs_cos: torch.Tensor,
+        freqs_sin: torch.Tensor,
+    ) -> torch.Tensor:
+        qkv = self.qkv(hidden_states).reshape(hidden_states.shape[0], 3, self.num_heads, self.head_dim)
+        query, key, value = qkv.unbind(dim=1)
+        original_dtype = qkv.dtype
+        query = self.q_norm(query)
+        key = self.k_norm(key)
+        query = query * freqs_cos + _rotate_half(query) * freqs_sin
+        key = key * freqs_cos + _rotate_half(key) * freqs_sin
+        query = query.to(dtype=original_dtype)
+        key = key.to(dtype=original_dtype)
+
+        if flash_attn_varlen_func is not None and query.is_cuda:
+            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+            hidden_states = flash_attn_varlen_func(
+                query, key, value, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen
+            ).reshape(hidden_states.shape[0], -1)
+        else:
+            segments = []
+            for start, end in zip(cu_seqlens[:-1].tolist(), cu_seqlens[1:].tolist()):
+                q = query[start:end].transpose(0, 1).unsqueeze(0)
+                k = key[start:end].transpose(0, 1).unsqueeze(0)
+                v = value[start:end].transpose(0, 1).unsqueeze(0)
+                segments.append(F.scaled_dot_product_attention(q, k, v).squeeze(0).transpose(0, 1))
+            hidden_states = torch.cat(segments, dim=0).reshape(hidden_states.shape[0], -1)
+
+        hidden_states = self.proj(hidden_states)
+        return self.proj_drop(hidden_states)
+
+
+class NiTMLP(nn.Module):
+    def __init__(self, hidden_size: int, mlp_hidden_dim: int):
+        super().__init__()
+        self.fc1 = nn.Linear(hidden_size, mlp_hidden_dim)
+        self.act = nn.GELU(approximate="tanh")
+        self.drop1 = nn.Dropout(0.0)
+        self.norm = nn.Identity()
+        self.fc2 = nn.Linear(mlp_hidden_dim, hidden_size)
+        self.drop2 = nn.Dropout(0.0)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.drop1(hidden_states)
+        hidden_states = self.norm(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return self.drop2(hidden_states)
+
+
+class NiTBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_norm: bool = False,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 512,
+    ):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.attn = NiTAttention(hidden_size, num_heads=num_heads, qk_norm=qk_norm)
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.mlp = NiTMLP(hidden_size, mlp_hidden_dim)
+        if use_adaln_lora:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(hidden_size, adaln_lora_dim, bias=True),
+                nn.Linear(adaln_lora_dim, 6 * hidden_size, bias=True),
+            )
+        else:
+            self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True))
+
+    def forward(self, hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(conditioning).chunk(
+            6, dim=-1
+        )
+        hidden_states = hidden_states + gate_msa * self.attn(
+            _modulate(self.norm1(hidden_states), shift_msa, scale_msa), cu_seqlens, freqs_cos, freqs_sin
+        )
+        hidden_states = hidden_states + gate_mlp * self.mlp(
+            _modulate(self.norm2(hidden_states), shift_mlp, scale_mlp)
+        )
+        return hidden_states
+
+
+class NiTFinalLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
+
+    def forward(self, hidden_states: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
+        shift, scale = self.adaLN_modulation(conditioning).chunk(2, dim=-1)
+        hidden_states = _modulate(self.norm_final(hidden_states), shift, scale)
+        return self.linear(hidden_states)
+
+
+def _build_mlp(hidden_size: int, projector_dim: int, z_dim: int) -> nn.Sequential:
+    return nn.Sequential(
+        nn.Linear(hidden_size, projector_dim),
+        nn.SiLU(),
+        nn.Linear(projector_dim, projector_dim),
+        nn.SiLU(),
+        nn.Linear(projector_dim, z_dim),
+    )
+
+
+class NiTTransformer2DModel(ModelMixin, ConfigMixin):
+    config_name = "config.json"
+
+    @register_to_config
+    def __init__(
+        self,
+        input_size: int = 32,
+        patch_size: int = 1,
+        in_channels: int = 32,
+        hidden_size: int = 1152,
+        depth: int = 28,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.0,
+        class_dropout_prob: float = 0.1,
+        num_classes: int = 1000,
+        encoder_depth: int = 8,
+        projector_dim: int = 2048,
+        z_dim: int = 1280,
+        use_checkpoint: bool = False,
+        custom_freqs: str = "normal",
+        theta: int = 10000,
+        max_pe_len_h: Optional[int] = None,
+        max_pe_len_w: Optional[int] = None,
+        decouple: bool = False,
+        ori_max_pe_len: Optional[int] = None,
+        qk_norm: bool = True,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 512,
+    ):
+        super().__init__()
+        del input_size
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.num_classes = num_classes
+        self.encoder_depth = encoder_depth
+        self.use_checkpoint = use_checkpoint
+
+        self.x_embedder = NiTPatchEmbed(patch_size, in_channels, hidden_size)
+        self.t_embedder = NiTTimestepEmbedder(hidden_size)
+        self.y_embedder = NiTLabelEmbedder(num_classes, hidden_size, class_dropout_prob)
+        self.rope = NiTRotaryEmbedding(
+            hidden_size // num_heads,
+            custom_freqs=custom_freqs,
+            theta=theta,
+            max_pe_len_h=max_pe_len_h,
+            max_pe_len_w=max_pe_len_w,
+            decouple=decouple,
+            ori_max_pe_len=ori_max_pe_len,
+        )
+        self.projector = _build_mlp(hidden_size, projector_dim, z_dim)
+        self.blocks = nn.ModuleList(
+            [
+                NiTBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qk_norm=qk_norm,
+                    use_adaln_lora=use_adaln_lora,
+                    adaln_lora_dim=adaln_lora_dim,
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.final_layer = NiTFinalLayer(hidden_size, patch_size, self.out_channels)
+
+    def _pack_latents(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.LongTensor, Tuple[int, int]]:
+        batch_size, channels, height, width = hidden_states.shape
+        if channels != self.in_channels:
+            raise ValueError(f"Expected {self.in_channels} latent channels, got {channels}.")
+        if height % self.patch_size != 0 or width % self.patch_size != 0:
+            raise ValueError("Latent height and width must be divisible by patch_size.")
+        latent_h = height // self.patch_size
+        latent_w = width // self.patch_size
+        hidden_states = hidden_states.reshape(batch_size, channels, latent_h, self.patch_size, latent_w, self.patch_size)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5).reshape(
+            batch_size * latent_h * latent_w, channels, self.patch_size, self.patch_size
+        )
+        image_sizes = torch.tensor([[latent_h, latent_w]] * batch_size, device=hidden_states.device, dtype=torch.long)
+        return hidden_states, image_sizes, (height, width)
+
+    def _unpack_latents(self, hidden_states: torch.Tensor, image_sizes: torch.LongTensor) -> torch.Tensor:
+        if image_sizes.shape[0] == 1:
+            height, width = image_sizes[0].tolist()
+            hidden_states = hidden_states.reshape(height, width, self.out_channels, self.patch_size, self.patch_size)
+            return hidden_states.permute(2, 0, 3, 1, 4).reshape(
+                1, self.out_channels, height * self.patch_size, width * self.patch_size
+            )
+
+        samples = []
+        cursor = 0
+        for height, width in image_sizes.tolist():
+            length = height * width
+            sample = hidden_states[cursor : cursor + length].reshape(
+                height, width, self.out_channels, self.patch_size, self.patch_size
+            )
+            samples.append(
+                sample.permute(2, 0, 3, 1, 4).reshape(
+                    self.out_channels, height * self.patch_size, width * self.patch_size
+                )
+            )
+            cursor += length
+        if len({tuple(sample.shape) for sample in samples}) != 1:
+            return hidden_states
+        return torch.stack(samples, dim=0)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: Union[torch.Tensor, float],
+        class_labels: torch.LongTensor,
+        image_sizes: Optional[Union[torch.LongTensor, List[Tuple[int, int]]]] = None,
+        return_dict: bool = True,
+        output_projection_states: bool = False,
+    ) -> Union[NiTTransformer2DModelOutput, Tuple[torch.Tensor, ...]]:
+        input_was_image = hidden_states.dim() == 4 and image_sizes is None
+        if input_was_image:
+            hidden_states, image_sizes, _ = self._pack_latents(hidden_states)
+        elif image_sizes is None:
+            raise ValueError("image_sizes must be provided when hidden_states are already packed.")
+        elif not torch.is_tensor(image_sizes):
+            image_sizes = torch.tensor(image_sizes, device=hidden_states.device, dtype=torch.long)
+        else:
+            image_sizes = image_sizes.to(device=hidden_states.device, dtype=torch.long)
+
+        if not torch.is_tensor(timestep):
+            timestep = torch.tensor([timestep], device=hidden_states.device, dtype=hidden_states.dtype)
+        timestep = timestep.to(device=hidden_states.device, dtype=hidden_states.dtype).flatten()
+        if timestep.numel() == 1:
+            timestep = timestep.repeat(image_sizes.shape[0])
+        class_labels = class_labels.to(device=hidden_states.device, dtype=torch.long).flatten()
+
+        hidden_states = self.x_embedder(hidden_states).squeeze(1)
+        freqs_cos, freqs_sin = self.rope(image_sizes, hidden_states.device)
+
+        seqlens = image_sizes[:, 0] * image_sizes[:, 1]
+        cu_seqlens = torch.cat(
+            [torch.zeros(1, device=hidden_states.device, dtype=torch.int32), torch.cumsum(seqlens, dim=0).int()]
+        )
+
+        conditioning = self.t_embedder(timestep) + self.y_embedder(class_labels)
+        conditioning = torch.cat([conditioning[i].repeat(int(seqlens[i]), 1) for i in range(image_sizes.shape[0])], dim=0)
+
+        projection_states = []
+        for index, block in enumerate(self.blocks):
+            if self.use_checkpoint and self.training:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    block, hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin, use_reentrant=False
+                )
+            else:
+                hidden_states = block(hidden_states, conditioning, cu_seqlens, freqs_cos, freqs_sin)
+            if output_projection_states and (index + 1) == self.encoder_depth:
+                projection_states.append(self.projector(hidden_states))
+
+        hidden_states = self.final_layer(hidden_states, conditioning)
+        hidden_states = hidden_states.reshape(hidden_states.shape[0], self.out_channels, self.patch_size, self.patch_size)
+        if input_was_image:
+            hidden_states = self._unpack_latents(hidden_states, image_sizes)
+
+        if not return_dict:
+            output = (hidden_states,)
+            if output_projection_states:
+                output = output + (tuple(projection_states),)
+            return output
+        return NiTTransformer2DModelOutput(
+            sample=hidden_states,
+            projection_states=tuple(projection_states) if output_projection_states else None,
+        )

NiT-S/vae/config.json

@@ -0,0 +1,88 @@
+{
+  "_class_name": "AutoencoderDC",
+  "_diffusers_version": "0.32.2",
+  "attention_head_dim": 32,
+  "decoder_act_fns": "silu",
+  "decoder_block_out_channels": [
+    128,
+    256,
+    512,
+    512,
+    1024,
+    1024
+  ],
+  "decoder_block_types": [
+    "ResBlock",
+    "ResBlock",
+    "ResBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock"
+  ],
+  "decoder_layers_per_block": [
+    3,
+    3,
+    3,
+    3,
+    3,
+    3
+  ],
+  "decoder_norm_types": "rms_norm",
+  "decoder_qkv_multiscales": [
+    [],
+    [],
+    [],
+    [
+      5
+    ],
+    [
+      5
+    ],
+    [
+      5
+    ]
+  ],
+  "downsample_block_type": "Conv",
+  "encoder_block_out_channels": [
+    128,
+    256,
+    512,
+    512,
+    1024,
+    1024
+  ],
+  "encoder_block_types": [
+    "ResBlock",
+    "ResBlock",
+    "ResBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock",
+    "EfficientViTBlock"
+  ],
+  "encoder_layers_per_block": [
+    2,
+    2,
+    2,
+    3,
+    3,
+    3
+  ],
+  "encoder_qkv_multiscales": [
+    [],
+    [],
+    [],
+    [
+      5
+    ],
+    [
+      5
+    ],
+    [
+      5
+    ]
+  ],
+  "in_channels": 3,
+  "latent_channels": 32,
+  "scaling_factor": 0.41407,
+  "upsample_block_type": "interpolate"
+}

NiT-S/vae/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfd991d1b54ffabf22745c5885589d8f2a7bc59930d95d92bd741c4fc64454bb
+size 1249044836

README.md

@@ -29,6 +29,9 @@ No separate `NiT-diffusers` package at inference time; only PyPI `diffusers` plu
 
 | Checkpoint | Path | Resolution | Recommended settings |
 | --- | --- | --- | --- |
+| NiT-S | `./NiT-S` | 256×256 | 250 steps, CFG 2.25, interval (0.0, 0.7) |
+| NiT-B | `./NiT-B` | 256×256 | 250 steps, CFG 2.25, interval (0.0, 0.7) |
+| NiT-L | `./NiT-L` | 512×512 | 250 steps, CFG 2.05, interval (0.0, 0.7) |
 | NiT-XL | `./NiT-XL` | 512×512 | 250 steps, CFG 2.05, interval (0.0, 0.7) |
 
 ## ImageNet class labels
@@ -81,7 +84,31 @@ image = pipe(
 image.save("demo.png")
 ```
 
-Load a **variant subfolder** (e.g. `./NiT-XL`), not the repo root.
+Load a **variant subfolder** (e.g. `./NiT-XL`, `./NiT-L`, `./NiT-B`, or `./NiT-S`), not the repo root.
+
+For NiT-S / NiT-B at 256×256 (official defaults):
+
+```python
+model_dir = Path("./NiT-S").resolve()  # or ./NiT-B
+pipe = DiffusionPipeline.from_pretrained(
+    str(model_dir),
+    local_files_only=True,
+    custom_pipeline=str(model_dir / "pipeline.py"),
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+)
+pipe.to("cuda")
+
+image = pipe(
+    class_labels="golden retriever",
+    height=256,
+    width=256,
+    num_inference_steps=250,
+    guidance_scale=2.25,
+    guidance_interval=(0.0, 0.7),
+    generator=torch.Generator(device="cuda").manual_seed(42),
+).images[0]
+```
 
 Hub usage follows Hugging Face model-id style (`UserID/RepoID`):