Publish HQ three-style manifest and validation

README.md

@@ -4,71 +4,64 @@ pipeline_tag: image-to-image
 base_model: neonforestmist/Clover-Image-Tiny-Inpaint
 license: creativeml-openrail-m
 tags:
-  - clover-image
-  - inpainting
   - coreml
+  - ios
+  - inpainting
+  - lora
+  - stable-diffusion
 ---
 
-# Clover Image Tiny Inpaint — Core ML
+# Clover Image Tiny Inpaint HQ — Core ML
 
-Compiled Core ML resources for the v2 SD 1.4-class Clover Image Tiny
-inpainting model. The optional download contains **1,671,581,989 bytes**:
-approximately **1,672 MB** (1.56 GiB). It is installed separately from Regular
-Clover.
+Core ML resources for the high-quality Clover Image Tiny inpainting pipeline.
+The iOS 18 batch-one U-Net accepts `[1, 9, 64, 64]`, runs classifier-free
+guidance as two serial passes, and exposes 144 mutable Core ML states for exact
+runtime composition of up to three Clover styles.
 
-## Runtime resources
-
-```text
-TextEncoder.mlmodelc
-VAEEncoder.mlmodelc
-VAEDecoder.mlmodelc
-SafetyChecker.mlmodelc
-UnetChunk1.mlmodelc
-UnetChunk2.mlmodelc
-vocab.json
-merges.txt
-```
+The HQ U-Net is under `hq-v3/Unet.mlmodelc`. Shared Clover resources remain at
+the repository root so app upgrades only download files whose checksums
+changed. `hq-v3/adapter-schema.json` maps each small `.safetensors` style into
+one of three independent state slots.
 
-The batch-one U-Net input is `[1, 9, 64, 64]`, ordered as noisy latent, mask,
-and masked-image latent. White mask pixels are regenerated; black pixels are
-preserved. The VAE encoder creates the masked-image latent locally.
+## Validation
 
-Recommended settings are DPM-Solver++, 20 steps, CFG 6.0, and a 96-pixel
-context margin around focused mask crops. Composite the result through the
-exact mask to preserve all source pixels outside the edit.
+PyTorch/Core ML parity on deterministic inputs:
 
-![Context-aware cat inpaint](https://huggingface.co/neonforestmist/Clover-Image-Tiny-Inpaint/resolve/main/examples/cat-result.png)
+| Configuration | PSNR |
+|---|---:|
+| Base HQ inpainting U-Net | **78.56 dB** |
+| Monet 0.70 + Pointillism 0.45 + Watercolor Anime 1.10 | **78.47 dB** |
 
-## Validation
+The release gate is 35 dB. The three-style result validates the exact
+block-concatenated LoRA sum rather than a UI-only approximation.
 
-The converted nine-channel U-Net passed Apple conversion parity at 40.8 dB and
-the independent fixed-shape Core ML smoke test at **46.2 dB**, above the 35 dB
-release threshold. The manifest records every runtime file's size and SHA-256.
+Held-out 24-case inpainting quality relative to the previous release:
 
-```bash
-coreml-tools/convert_inpaint.sh \
-  coreml-inpaint-v2 artifacts/clover-image-tiny-inpaint-v2
-coreml-tools/smoke-test-inpaint.sh \
-  artifacts/clover-image-tiny-inpaint-v2 coreml-inpaint-v2
-.venv-coreml/bin/python coreml-tools/manifest_inpaint.py \
-  coreml-inpaint-release-v2 \
-  --model-dir artifacts/clover-image-tiny-inpaint-v2
-```
+| Metric | Previous | HQ |
+|---|---:|---:|
+| Masked prompt CLIP similarity | 0.2642 | **0.2768** |
+| Masked target MAE | 0.2510 | **0.2231** |
+| Changed pixels outside the mask | 0 | **0** |
 
-The package targets iOS 18. The chunked batch-one U-Net lets the Swift runtime
-perform classifier-free guidance as two serial passes to reduce peak memory.
+## Runtime contract
 
-Regular Clover Image Tiny LoRAs target a four-channel U-Net and cannot be
-loaded dynamically into this nine-channel package. An inpainting-specific LoRA
-can be fused before conversion as a separate Core ML bundle.
+- Minimum OS: iOS 18
+- Resolution: 512×512
+- U-Net input: noisy latent (4) + mask (1) + masked image latent (4)
+- Mask: white regenerates, black preserves
+- Maximum simultaneous styles: 3
+- Safety checker: not constructed by the Clover iOS pipeline
 
 ## Citation
 
 ```bibtex
-@software{lozadaperez2026cloverimagetinyinpaint,
+@software{lozadaperez2026cloverimagetinyinpaintcoreml,
   author = {Lukas Lozada Perez},
-  title = {Clover Image Tiny Inpaint: Compact SD 1.4-Class Image Inpainting},
+  title = {Clover Image Tiny Inpaint HQ Core ML},
   year = {2026},
-  url = {https://huggingface.co/neonforestmist/Clover-Image-Tiny-Inpaint}
+  url = {https://huggingface.co/neonforestmist/Clover-Image-Tiny-Inpaint-CoreML}
 }
 ```
+
+Designed and developed independently by Lukas Lozada Perez. Open weights under
+the CreativeML Open RAIL-M license; inference runs completely on device.

manifest.json

@@ -1,12 +1,14 @@
 {
-  "schema_version": 1,
+  "schema_version": 2,
   "model": "neonforestmist/Clover-Image-Tiny-Inpaint",
   "base_model": "neonforestmist/Clover-Image-Tiny",
   "minimum_ios": "18.0",
+  "requires_base_model": true,
   "resolution": [
     512,
     512
   ],
+  "max_adapter_count": 3,
   "unet_input": {
     "shape": [
       1,
@@ -28,131 +30,43 @@
     "mask_crop_padding": 96,
     "composite_outside_mask": "exact source pixels"
   },
+  "validation": {
+    "samples": 24,
+    "masked_clip_similarity": 0.27676651254296303,
+    "masked_target_mae": 0.2230850402265787,
+    "base_psnr_db": 78.55727510619732,
+    "three_style_psnr_db": 78.4737949968816
+  },
   "resources": [
     {
-      "path": "SafetyChecker.mlmodelc/analytics/coremldata.bin",
-      "size": 243,
-      "sha256": "8211508a22331e6868fa9d207c6532ae5fc11791e8ceac6bf6fa4c7dd89004e8"
-    },
-    {
-      "path": "SafetyChecker.mlmodelc/coremldata.bin",
-      "size": 1549,
-      "sha256": "f57494e9bc01a089c73ba6b1b9554e1a3df155efbf7cf36223ba10ae494f0d58"
-    },
-    {
-      "path": "SafetyChecker.mlmodelc/metadata.json",
-      "size": 4687,
-      "sha256": "4af6e1bd32288fd5e6e6f171d7e63a8a3f3ec776b45c16bb7e3fe352e36158f0"
-    },
-    {
-      "path": "SafetyChecker.mlmodelc/model.mil",
-      "size": 318080,
-      "sha256": "aacdc5f65afa08beec17ab2f0f04d866ce6df14dab7845ebe6465ffd25790ea3"
-    },
-    {
-      "path": "SafetyChecker.mlmodelc/weights/weight.bin",
-      "size": 607990114,
-      "sha256": "ffdfb1a4774feb816dd2ccc3e304d109013bc9591e188dcab840789583fdc1b4"
-    },
-    {
-      "path": "TextEncoder.mlmodelc/analytics/coremldata.bin",
+      "path": "Unet.mlmodelc/analytics/coremldata.bin",
+      "remote_path": "hq-v3/Unet.mlmodelc/analytics/coremldata.bin",
       "size": 243,
-      "sha256": "3ba6d80c5e94f50f17a70a965c2f63d24fa4239672d0fb5dadcf888e69cafc0f"
+      "sha256": "0776280c592bd31e478eecd0cac5918984d48e629d1ffa5a9d6432a698448c9e"
     },
     {
-      "path": "TextEncoder.mlmodelc/coremldata.bin",
-      "size": 952,
-      "sha256": "734dc3fdcdcc31367ffae6a58b1e4b0d713eb3f043b89811b7fcaec6df67a495"
+      "path": "Unet.mlmodelc/coremldata.bin",
+      "remote_path": "hq-v3/Unet.mlmodelc/coremldata.bin",
+      "size": 14009,
+      "sha256": "4cb24ed826e3e090138e99e3eb334e706d6b8bf733771fe034355268456e6834"
     },
     {
-      "path": "TextEncoder.mlmodelc/metadata.json",
-      "size": 2992,
-      "sha256": "108a55b2e93052a093256052dbdfa9a6515c8918fbcd0c75661ffd753401ca25"
+      "path": "Unet.mlmodelc/metadata.json",
+      "remote_path": "hq-v3/Unet.mlmodelc/metadata.json",
+      "size": 50802,
+      "sha256": "54a5d730ce8790f028faacfe44cf9b555fb3109153cb5ffd8016ff5deb0375d8"
     },
     {
-      "path": "TextEncoder.mlmodelc/model.mil",
-      "size": 154795,
-      "sha256": "e3a95541282de8bc236c672d8170bfea3aceb478bd5fc5f96f390857d758be80"
+      "path": "Unet.mlmodelc/model.mil",
+      "remote_path": "hq-v3/Unet.mlmodelc/model.mil",
+      "size": 1206218,
+      "sha256": "2f7238a5825ae16c9dc054536d9a5dd7cf0963af2423979b4f84ae9121c6606a"
     },
     {
-      "path": "TextEncoder.mlmodelc/weights/weight.bin",
-      "size": 246145536,
-      "sha256": "b3e8699e18a386ba82633384e0a6f5c090983f1c66b24a26fad3554fc2a12ddb"
-    },
-    {
-      "path": "UnetChunk1.mlmodelc/analytics/coremldata.bin",
-      "size": 243,
-      "sha256": "7ee3483e4486c1bdc21299013aa9d3a25fcafb3d064750294c822cf8d6708151"
-    },
-    {
-      "path": "UnetChunk1.mlmodelc/coremldata.bin",
-      "size": 532,
-      "sha256": "f1ebbe33a5324d71a517efe9d5f5b8f9bbcb9bd2e991a57171840161b3a63430"
-    },
-    {
-      "path": "UnetChunk1.mlmodelc/metadata.json",
-      "size": 5242,
-      "sha256": "54c3b23ff2d46f00cbe5e3b03e79abf4e3fbb2bc811fd440219107c79dde2ef1"
-    },
-    {
-      "path": "UnetChunk1.mlmodelc/model.mil",
-      "size": 359479,
-      "sha256": "6067c2cbbb7f35ca3f3811ded8191fab58ddd1c27ee3de5e8249fa6e9c0d7581"
-    },
-    {
-      "path": "UnetChunk1.mlmodelc/weights/weight.bin",
-      "size": 324651584,
-      "sha256": "091529bc9e2b2bdda0508614e223570f7a8a4fe288bc510ec844635ba2f01046"
-    },
-    {
-      "path": "UnetChunk2.mlmodelc/analytics/coremldata.bin",
-      "size": 243,
-      "sha256": "1f7d7cba85c9edd629d133c0224baf71e989ad4bfe1da89f37706f1d6cab45ea"
-    },
-    {
-      "path": "UnetChunk2.mlmodelc/coremldata.bin",
-      "size": 513,
-      "sha256": "94ea7e153dab4f55ad8a40d12f9853422f8bdce240f5b608f3ae72ac3229fb35"
-    },
-    {
-      "path": "UnetChunk2.mlmodelc/metadata.json",
-      "size": 4926,
-      "sha256": "4c3b0ad1527a1b457d22a70385ad3d8424d72d4cecb825a12e996539de649a16"
-    },
-    {
-      "path": "UnetChunk2.mlmodelc/model.mil",
-      "size": 460764,
-      "sha256": "82541a2df3d6b955984b151f66dc7b48f44773c1c4a3ad81b64c474bea56ba9e"
-    },
-    {
-      "path": "UnetChunk2.mlmodelc/weights/weight.bin",
-      "size": 322211328,
-      "sha256": "04910ba6d4ca88b34dd1202c848aebfc33e339010d203573ebc4138a4e9f78ba"
-    },
-    {
-      "path": "VAEDecoder.mlmodelc/analytics/coremldata.bin",
-      "size": 243,
-      "sha256": "ff20b000564aa02c66b7e254258a5c8ba7d34fdb99c54be3c85a42d8eea45a62"
-    },
-    {
-      "path": "VAEDecoder.mlmodelc/coremldata.bin",
-      "size": 879,
-      "sha256": "7bb1f9ddd844d6c8e14c6e52a1e05d7c493e71091d788798d065141bbe3e131b"
-    },
-    {
-      "path": "VAEDecoder.mlmodelc/metadata.json",
-      "size": 2690,
-      "sha256": "72a20fd0a1db9e6d3395c6be9ab3c91b70fb771c138d8b7fef5ccc6b9153539c"
-    },
-    {
-      "path": "VAEDecoder.mlmodelc/model.mil",
-      "size": 173366,
-      "sha256": "db77ff85f06e97d687f09b9e1ba478f7573a3e0e344137d71cb57f1f4f28c25b"
-    },
-    {
-      "path": "VAEDecoder.mlmodelc/weights/weight.bin",
-      "size": 98993280,
-      "sha256": "ca45a093443d3a9719c5f7c61010315e94ffc8019eff27a0eeb52075a05ac752"
+      "path": "Unet.mlmodelc/weights/weight.bin",
+      "remote_path": "hq-v3/Unet.mlmodelc/weights/weight.bin",
+      "size": 1719146496,
+      "sha256": "0f6a88bcd1ae4f36c076d46d404373c2f5c44bf85975ed2737b23643b9040826"
     },
     {
       "path": "VAEEncoder.mlmodelc/analytics/coremldata.bin",
@@ -180,14 +94,10 @@
       "sha256": "71bf826e5ee1b455462aff0b34e3a469f5f58955ae715708962b1d81db92dc64"
     },
     {
-      "path": "merges.txt",
-      "size": 524619,
-      "sha256": "9fd691f7c8039210e0fced15865466c65820d09b63988b0174bfe25de299051a"
-    },
-    {
-      "path": "vocab.json",
-      "size": 1098502,
-      "sha256": "4e6a66b6903c12d273a1a6645f761a2de07e239facfc52fd797b83a75014ca78"
+      "path": "adapter-schema.json",
+      "remote_path": "hq-v3/adapter-schema.json",
+      "size": 55773,
+      "sha256": "b356bae467a4e60cb26151c5b2cad2d2dc298b057f44b71479096036ca53c1d9"
     }
   ]
 }