Upload folder using huggingface_hub

backup_sessao_2026-08-06/apex_edit.json

@@ -0,0 +1 @@
+{"id": "krea2-identity-edit-unified", "revision": 0, "last_node_id": 108, "last_link_id": 28, "nodes": [{"id": 29, "type": "SaveImage", "pos": [3180, -160], "size": [760, 860], "flags": {}, "order": 23, "mode": 0, "inputs": [{"name": "images", "type": "IMAGE", "link": 26}], "outputs": [{"name": "images", "type": "IMAGE", "links": null}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "SaveImage"}, "widgets_values": ["krea2_identity_edit"]}, {"id": 53, "type": "KSampler", "pos": [2800, -160], "size": [320, 560], "flags": {}, "order": 21, "mode": 0, "inputs": [{"name": "model", "type": "MODEL", "link": 3}, {"name": "positive", "type": "CONDITIONING", "link": 22}, {"name": "negative", "type": "CONDITIONING", "link": 23}, {"name": "latent_image", "type": "LATENT", "link": 24}], "outputs": [{"name": "LATENT", "type": "LATENT", "links": [25]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "KSampler"}, "widgets_values": [1088049369132323, "randomize", 8, 1, "euler", "simple", 1]}, {"id": 54, "type": "VAEDecode", "pos": [2810, 440], "size": [225, 8], "flags": {"collapsed": true}, "order": 22, "mode": 0, "inputs": [{"name": "samples", "type": "LATENT", "link": 25}, {"name": "vae", "type": "VAE", "link": 6}], "outputs": [{"name": "IMAGE", "type": "IMAGE", "links": [26]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "VAEDecode"}, "widgets_values": []}, {"id": 55, "type": "UNETLoader", "pos": [60, -160], "size": [480, 130], "flags": {}, "order": 0, "mode": 0, "inputs": [], "outputs": [{"name": "MODEL", "type": "MODEL", "links": [1]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "UNETLoader", "models": [{"name": "krea2_turbo_fp8_scaled.safetensors", "url": "https://huggingface.co/Comfy-Org/Krea-2/resolve/main/diffusion_models/krea2_turbo_fp8_scaled.safetensors", "directory": "diffusion_models"}]}, "widgets_values": ["krea2_raw_fp8_scaled.safetensors", "default"]}, {"id": 56, "type": "CLIPLoader", "pos": [60, 30], "size": [480, 140], "flags": {}, "order": 1, "mode": 0, "inputs": [], "outputs": [{"name": "CLIP", "type": "CLIP", "links": [4, 5]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "CLIPLoader", "models": [{"name": "qwen3vl_4b_fp8_scaled.safetensors", "url": "https://huggingface.co/Comfy-Org/Krea-2/resolve/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors", "directory": "text_encoders"}]}, "widgets_values": ["qwen3vl_4b_bf16.safetensors", "krea2", "default"]}, {"id": 57, "type": "VAELoader", "pos": [60, 220], "size": [480, 90], "flags": {}, "order": 2, "mode": 0, "inputs": [], "outputs": [{"name": "VAE", "type": "VAE", "links": [6, 7, 8, 16]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "VAELoader", "models": [{"name": "qwen_image_vae.safetensors", "url": "https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors", "directory": "vae"}]}, "widgets_values": ["qwen_image_vae.safetensors"]}, {"id": 71, "type": "LoraLoaderModelOnly", "pos": [60, 370], "size": [480, 120], "flags": {}, "order": 14, "mode": 0, "inputs": [{"name": "model", "type": "MODEL", "link": 1}], "outputs": [{"name": "MODEL", "type": "MODEL", "links": [28]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "LoraLoaderModelOnly"}, "widgets_values": ["krea2_turbo_lora_rank_64_bf16.safetensors", 1.0]}, {"id": 72, "type": "LoadImage", "pos": [640, -160], "size": [520, 440], "flags": {}, "order": 3, "mode": 0, "inputs": [], "outputs": [{"name": "IMAGE", "type": "IMAGE", "links": [11, 12, 13, 14]}, {"name": "MASK", "type": "MASK", "links": null}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "LoadImage"}, "widgets_values": ["example.png", "image"]}, {"id": 73, "type": "VAEEncode", "pos": [1260, 60], "size": [225, 8], "flags": {"collapsed": true}, "order": 15, "mode": 0, "inputs": [{"name": "pixels", "type": "IMAGE", "link": 11}, {"name": "vae", "type": "VAE", "link": 7}], "outputs": [{"name": "LATENT", "type": "LATENT", "links": [15]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "VAEEncode"}, "widgets_values": []}, {"id": 79, "type": "Krea2EditModelPatch", "pos": [2320, -160], "size": [420, 350], "flags": {}, "order": 20, "mode": 0, "inputs": [{"name": "model", "type": "MODEL", "link": 2}, {"name": "source_latent", "type": "LATENT", "link": 15}, {"name": "source_latent_b", "shape": 7, "type": "LATENT", "link": 21}, {"name": "ref_boost_mask", "shape": 7, "type": "MASK", "link": null}, {"name": "vae", "shape": 7, "type": "VAE", "link": 16}, {"name": "source_image", "shape": 7, "type": "IMAGE", "link": 14}, {"name": "source_image_b", "shape": 7, "type": "IMAGE", "link": 20}, {"name": "target_latent", "shape": 7, "type": "LATENT", "link": 27}], "outputs": [{"name": "MODEL", "type": "MODEL", "links": [3]}], "properties": {"Node name for S&R": "Krea2EditModelPatch"}, "widgets_values": [1.0, 1, "fit"]}, {"id": 82, "type": "EmptySD3LatentImage", "pos": [1260, 410], "size": [270, 150], "flags": {}, "order": 16, "mode": 0, "inputs": [{"name": "width", "type": "INT", "widget": {"name": "width"}, "link": 9}, {"name": "height", "type": "INT", "widget": {"name": "height"}, "link": 10}], "outputs": [{"name": "LATENT", "type": "LATENT", "links": [24, 27]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "EmptySD3LatentImage"}, "widgets_values": [1024, 1024, 1]}, {"id": 83, "type": "ResolutionSelector", "pos": [1250, 150], "size": [270, 190], "flags": {}, "order": 4, "mode": 0, "inputs": [], "outputs": [{"name": "width", "type": "INT", "links": [9]}, {"name": "height", "type": "INT", "links": [10]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "ResolutionSelector"}, "widgets_values": ["1:1 (Square)", 1, 8]}, {"id": 84, "type": "Krea2EditGroundedEncode", "pos": [1620, -160], "size": [620, 330], "flags": {}, "order": 18, "mode": 0, "inputs": [{"name": "clip", "type": "CLIP", "link": 4}, {"name": "image", "shape": 7, "type": "IMAGE", "link": 12}, {"name": "image_b", "shape": 7, "type": "IMAGE", "link": 18}], "outputs": [{"name": "CONDITIONING", "type": "CONDITIONING", "links": [22]}], "properties": {"Node name for S&R": "Krea2EditGroundedEncode"}, "widgets_values": ["Change her outfit to a red raincoat.", 768, ""]}, {"id": 85, "type": "Krea2EditGroundedEncode", "pos": [1620, 250], "size": [225, 8], "flags": {"collapsed": true}, "order": 19, "mode": 0, "inputs": [{"name": "clip", "type": "CLIP", "link": 5}, {"name": "image", "shape": 7, "type": "IMAGE", "link": 13}, {"name": "image_b", "shape": 7, "type": "IMAGE", "link": 19}], "outputs": [{"name": "CONDITIONING", "type": "CONDITIONING", "links": [23]}], "title": "negative (leave empty)", "properties": {"Node name for S&R": "Krea2EditGroundedEncode"}, "widgets_values": ["", 768, ""]}, {"id": 90, "type": "LoadImage", "pos": [640, 640], "size": [520, 440], "flags": {}, "order": 5, "mode": 4, "inputs": [], "outputs": [{"name": "IMAGE", "type": "IMAGE", "links": [17, 18, 19, 20]}, {"name": "MASK", "type": "MASK", "links": null}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "LoadImage"}, "widgets_values": ["example_person.png", "image"]}, {"id": 92, "type": "VAEEncode", "pos": [1250, 970], "size": [228.75, 8], "flags": {"collapsed": true}, "order": 17, "mode": 4, "inputs": [{"name": "pixels", "type": "IMAGE", "link": 17}, {"name": "vae", "type": "VAE", "link": 8}], "outputs": [{"name": "LATENT", "type": "LATENT", "links": [21]}], "properties": {"cnr_id": "comfy-core", "ver": "0.26.0", "Node name for S&R": "VAEEncode"}, "widgets_values": []}, {"id": 100, "type": "Note", "pos": [60, 540], "size": [480, 150], "flags": {}, "order": 6, "mode": 0, "inputs": [], "outputs": [], "title": "Model files", "properties": {"Node name for S&R": "Note"}, "widgets_values": ["Files (auto-downloadable via the manager links):\n- krea2_turbo_fp8_scaled (or krea2_raw_bf16 - see GENERATE note)\n- qwen3vl_4b_fp8_scaled text encoder, type 'krea2'\n- qwen_image_vae\n- krea2_identity_edit_v1_2.safetensors LoRA (this model)"], "color": "#432", "bgcolor": "#653"}, {"id": 101, "type": "Note", "pos": [640, 330], "size": [520, 190], "flags": {}, "order": 7, "mode": 0, "inputs": [], "outputs": [], "title": "How inputs work", "properties": {"Node name for S&R": "Note"}, "widgets_values": ["YOUR IMAGE - the image being edited (your character, the photo to\nrestyle, the scene...).\n\nTwo-image mode (group 2 enabled): THIS slot is always the SCENE\n(image 1). The person always goes in group 2 (image 2).\nInputs around 1MP work best. Two-person images: stay at/below ~1.5MP."], "color": "#432", "bgcolor": "#653"}, {"id": 102, "type": "Note", "pos": [640, 1130], "size": [520, 210], "flags": {}, "order": 8, "mode": 0, "inputs": [], "outputs": [], "title": "Enable for two-image edits", "properties": {"Node name for S&R": "Note"}, "widgets_values": ["OPTIONAL second reference - the PERSON for insert-into-scene edits\n('place this person at the table').\n\nThis group ships DISABLED (purple = bypassed) - the workflow is a\nplain single-image editor until you need it.\nEnable: right-click this group's title bar -> toggle Bypass off.\nOrder rule: scene = image 1 (group 1), person = image 2 (here)."], "color": "#432", "bgcolor": "#653"}, {"id": 103, "type": "Note", "pos": [1620, 360], "size": [620, 300], "flags": {}, "order": 9, "mode": 0, "inputs": [], "outputs": [], "title": "Writing instructions", "properties": {"Node name for S&R": "Note"}, "widgets_values": ["Write the edit in plain English:\n  'Relight the scene with warm golden hour sunlight.'\n  'Place this person at the cafe table, holding a coffee.'\n  'Change her outfit to a red raincoat.'\n\nUnique face coming out generic? Append:\n  'Preserve the exact facial identity - distinguishing features: ...'\nand describe what makes the face distinctive.\n\ngrounding_px: how large the model 'sees' the input while reading your\ninstruction. 384-768 is the trained range. Getting doubled/split\ncompositions? LOWER it. The collapsed node below is the negative -\nonly used at CFG > 1; keep its prompt empty."], "color": "#432", "bgcolor": "#653"}, {"id": 104, "type": "Note", "pos": [2320, 230], "size": [420, 420], "flags": {}, "order": 10, "mode": 0, "inputs": [], "outputs": [], "title": "Fidelity dial", "properties": {"Node name for S&R": "Note"}, "widgets_values": ["THE FIDELITY DIAL (ref_boost) - how hard the model looks at your\nreference. New in this release.\n\n  1.0 = classic v1.1 behavior\n  4.0 = recommended (pre-set): much stronger face + body likeness,\n        more reliable edits\n  >10 = over-copy: removals/replacements start failing\n  <1  = suppress the reference (creative freedom)\n\nref_boost_a: the scene dial (image 1) in two-image mode -\nleave at 1.0 unless you are exploring. ref_boost applies to the\nperson (image 2) when group 2 is enabled, else to your image.\n\nfit_mode: how inputs fit a different output aspect\nratio. fit = resample to the output shape (default, recommended); crop (legacy) = center-crop\nonly for v1/v1.1 weights."], "color": "#432", "bgcolor": "#653"}, {"id": 105, "type": "Note", "pos": [2800, 490], "size": [320, 280], "flags": {}, "order": 11, "mode": 0, "inputs": [], "outputs": [], "title": "Sampling", "properties": {"Node name for S&R": "Note"}, "widgets_values": ["Turbo model: 10 steps, cfg 1.\n\nStep dial:\n  8  = stronger composition/instruction adherence\n  12 = more face detail\n  10 = balanced\n\nUsing krea2_raw instead of turbo:\n40 steps, cfg 3-4 (negative then matters)."], "color": "#432", "bgcolor": "#653"}, {"id": 106, "type": "Note", "pos": [620, -600], "size": [1600, 330], "flags": {}, "order": 12, "mode": 0, "inputs": [], "outputs": [], "title": "README", "properties": {"Node name for S&R": "Note"}, "widgets_values": ["KREA 2 IDENTITY EDIT  -  v1.2 (Fidelity update)\nUse an image like a LoRA: your character/scene goes in as a reference, plain-English\ninstructions do the rest.\n\nQUICK START:  1) load your image in group 1   2) write the instruction in group 3   3) Queue.\nThe Fidelity dial (group 4) ships at 4 = strong likeness. Two-image person-into-scene: enable\ngroup 2 (right-click its title -> toggle Bypass).\n\nRequires the comfyui-krea2edit nodes from this release or newer.\ngithub.com/lbouaraba/comfyui-krea2edit   -   huggingface.co/conradlocke/krea2-identity-edit"], "color": "#232", "bgcolor": "#353"}, {"id": 107, "type": "Note", "pos": [1250, -140], "size": [270, 110], "flags": {}, "order": 13, "mode": 0, "inputs": [], "outputs": [], "title": "Resolution", "properties": {"Node name for S&R": "Note"}, "widgets_values": ["Output resolution. 1MP is the\nsweet spot; go higher only for\nsingle-person edits."], "color": "#432", "bgcolor": "#653"}, {"id": 108, "type": "LoraLoaderModelOnly", "pos": [360, 410], "size": [315, 82], "flags": {}, "order": 99, "mode": 0, "inputs": [{"name": "model", "type": "MODEL", "link": 28}], "outputs": [{"name": "MODEL", "type": "MODEL", "links": [2]}], "properties": {"Node name for S&R": "LoraLoaderModelOnly"}, "widgets_values": ["apex_step2000.safetensors", 1.0]}], "links": [[1, 55, 0, 71, 0, "MODEL"], [2, 108, 0, 79, 0, "MODEL"], [3, 79, 0, 53, 0, "MODEL"], [4, 56, 0, 84, 0, "CLIP"], [5, 56, 0, 85, 0, "CLIP"], [6, 57, 0, 54, 1, "VAE"], [7, 57, 0, 73, 1, "VAE"], [8, 57, 0, 92, 1, "VAE"], [9, 83, 0, 82, 0, "INT"], [10, 83, 1, 82, 1, "INT"], [11, 72, 0, 73, 0, "IMAGE"], [12, 72, 0, 84, 1, "IMAGE"], [13, 72, 0, 85, 1, "IMAGE"], [14, 72, 0, 79, 5, "IMAGE"], [15, 73, 0, 79, 1, "LATENT"], [16, 57, 0, 79, 4, "VAE"], [17, 90, 0, 92, 0, "IMAGE"], [18, 90, 0, 84, 2, "IMAGE"], [19, 90, 0, 85, 2, "IMAGE"], [20, 90, 0, 79, 6, "IMAGE"], [21, 92, 0, 79, 2, "LATENT"], [22, 84, 0, 53, 1, "CONDITIONING"], [23, 85, 0, 53, 2, "CONDITIONING"], [24, 82, 0, 53, 3, "LATENT"], [25, 53, 0, 54, 0, "LATENT"], [26, 54, 0, 29, 0, "IMAGE"], [27, 82, 0, 79, 7, "LATENT"], [28, 71, 0, 108, 0, "MODEL"]], "groups": [{"id": 1, "title": "MODELS", "bounding": [40, -240, 540, 950], "color": "#3f789e", "flags": {}}, {"id": 2, "title": "1 \u00b7 IMAGE TO EDIT", "bounding": [620, -240, 580, 790], "color": "#8A8", "flags": {}}, {"id": 3, "title": "2 \u00b7 SECOND REFERENCE (optional \u2014 bypassed)", "bounding": [620, 560, 580, 810], "color": "#a1309b", "flags": {}}, {"id": 4, "title": "3 \u00b7 INSTRUCTION", "bounding": [1600, -240, 680, 930], "color": "#b58b2a", "flags": {}}, {"id": 5, "title": "4 \u00b7 FIDELITY", "bounding": [2300, -240, 460, 910], "color": "#c15f3c", "flags": {}}, {"id": 6, "title": "5 \u00b7 GENERATE", "bounding": [2780, -240, 1200, 1030], "color": "#3f789e", "flags": {}}], "config": {}, "extra": {"frontendVersion": "1.45.19", "ds": {"scale": 0.8264462809917354, "offset": [69.76703364204093, 382.9680133228499]}, "VHS_latentpreview": false, "VHS_latentpreviewrate": 0, "VHS_MetadataImage": true, "VHS_KeepIntermediate": true}, "version": 0.4}

backup_sessao_2026-08-06/configs/apex_bat_rand.json

@@ -0,0 +1 @@
+{"samples": [{"name": "imagem000836", "reference": "/workspace/datasets/ctx_original/input_A/imagem000836.jpg", "prompt": "A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem001768", "reference": "/workspace/datasets/ctx_original/input_A/imagem001768.jpg", "prompt": "A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text \"ALIENS! ABDUCT ME!!!\" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000294", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000294.jpg", "prompt": "A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000075", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000075.jpg", "prompt": "A tight close-up shot of a young person with dark hair, their face contorted in an intense expression of distress. Their mouth is wide open in a cry, and large, glistening tears stream from their wide, light-colored eyes. The character is positioned on the left side of the frame, with their head tilted slightly. A large, pale, hand-like appendage with multiple fingers extends from the right side of the frame, reaching toward the character's face. The background is a blurred, abstract blend of deep purple and bright magenta light, suggesting a high-energy or supernatural environment.  Character continuity: same character. Background continuity: new background.", "width": 1024, "height": 1024}, {"name": "imagem000759", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000759.jpg", "prompt": "A medium shot of a white and grey robotic figure with a spherical head and glowing cyan circular eyes, set against a cloudy grey sky. The robot holds a small, translucent blue crab with red-tipped claws on its palm. A seagull is captured in motion, flying directly in front of the robot's face, creating a slight motion blur. The robot has segmented arms and a smooth, rounded torso. The background is an overcast sky with soft, diffused light.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000663", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000663.jpg", "prompt": "A high-angle, close-up shot looking down into an open stone sarcophagus. A man with short, light-colored hair lies peacefully inside with his eyes closed, his hands folded over his chest. He wears a dark blue, high-collared military-style jacket with gold trim and a star-shaped pendant. The interior of the sarcophagus is lined with brown, padded fabric. On either side of the frame, the gloved hands of two individuals are visible, gripping the edges of the sarcophagus. Several small, light-colored flower petals are scattered across the man and the interior of the coffin.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000611", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000611.jpg", "prompt": "A close-up, slightly angled view of an old television screen displaying a live news broadcast of a night protest in a city street. The screen shows a large crowd of people gathered on a dark asphalt road at night, illuminated by streetlights and green traffic signals. A circular traffic sign with an upward arrow is visible on the right side of the screen. A news ticker at the bottom of the screen displays the text \"BIOS AUMENTAN A VELOCIDAD ALARMANTE\" and \"AN MANIFESTACIONES CONTRA LEY MARCIAL\". The television is viewed from the side, with the blurred silhouette of a person's head and shoulder visible in the foreground on the left.  Character continuity: same character. Background continuity: same background view.", "width": 1024, "height": 1024}, {"name": "imagem000374", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000374.jpg", "prompt": "A close-up, rear-view shot of a person wearing a dark, tactical-style hooded garment or backpack with a prominent circular vent or mechanical detail on the back. The person is positioned in a desert environment with orange-toned sand dunes visible in the background. The lighting is warm and dim, casting deep shadows across the scene. The person is oriented away from the camera, with a portion of a strap or equipment visible on the left side of the frame.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000273", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000273.jpg", "prompt": "A close-up shot focusing on a person's hand resting on a green felt gaming table. The person is wearing a white long-sleeved shirt with the cuff visible. Next to the hand, there are several tall stacks of circular casino chips in various colors, including purple, white, green, and orange. The lighting is warm and focused on the table surface.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem001519", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem001519.jpg", "prompt": "A close-up, high-angle shot focusing on a person's torso and arm. The individual wears a white jacket with a dark-colored garment underneath. A large, jagged, dark red bloodstain is visible on the chest area of the white jacket, with droplets of blood splattered across the fabric. The person's arm is partially visible, and the scene is set against a dark, indistinct background.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000001", "reference": "/workspace/datasets/ctx_original/input_A/imagem000001.jpg", "prompt": "A wide shot of a spacious, high-ceilinged hall with light-colored stone walls. A row of black, wheeled display carts with glass cases sits on a red carpet, containing various artifacts including a sword and colorful gemstones. A small, black, winged creature with large yellow eyes and a grey belly stands on the red carpet in the foreground, facing the viewer. Two potted plants flank a set of double doors in the background.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}]}

backup_sessao_2026-08-06/configs/apex_bat_ref.json

@@ -0,0 +1 @@
+{"samples": [{"name": "imagem000001", "reference": "/workspace/datasets/ctx_original/input_A/imagem000001.jpg", "prompt": "A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000836", "reference": "/workspace/datasets/ctx_original/input_A/imagem000836.jpg", "prompt": "A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text \"ALIENS! ABDUCT ME!!!\" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem001768", "reference": "/workspace/datasets/ctx_original/input_A/imagem001768.jpg", "prompt": "A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000294", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000294.jpg", "prompt": "A tight close-up shot of a young person with dark hair, their face contorted in an intense expression of distress. Their mouth is wide open in a cry, and large, glistening tears stream from their wide, light-colored eyes. The character is positioned on the left side of the frame, with their head tilted slightly. A large, pale, hand-like appendage with multiple fingers extends from the right side of the frame, reaching toward the character's face. The background is a blurred, abstract blend of deep purple and bright magenta light, suggesting a high-energy or supernatural environment.  Character continuity: same character. Background continuity: new background.", "width": 1024, "height": 1024}, {"name": "imagem000075", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000075.jpg", "prompt": "A medium shot of a white and grey robotic figure with a spherical head and glowing cyan circular eyes, set against a cloudy grey sky. The robot holds a small, translucent blue crab with red-tipped claws on its palm. A seagull is captured in motion, flying directly in front of the robot's face, creating a slight motion blur. The robot has segmented arms and a smooth, rounded torso. The background is an overcast sky with soft, diffused light.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000759", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000759.jpg", "prompt": "A high-angle, close-up shot looking down into an open stone sarcophagus. A man with short, light-colored hair lies peacefully inside with his eyes closed, his hands folded over his chest. He wears a dark blue, high-collared military-style jacket with gold trim and a star-shaped pendant. The interior of the sarcophagus is lined with brown, padded fabric. On either side of the frame, the gloved hands of two individuals are visible, gripping the edges of the sarcophagus. Several small, light-colored flower petals are scattered across the man and the interior of the coffin.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000663", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000663.jpg", "prompt": "A close-up, slightly angled view of an old television screen displaying a live news broadcast of a night protest in a city street. The screen shows a large crowd of people gathered on a dark asphalt road at night, illuminated by streetlights and green traffic signals. A circular traffic sign with an upward arrow is visible on the right side of the screen. A news ticker at the bottom of the screen displays the text \"BIOS AUMENTAN A VELOCIDAD ALARMANTE\" and \"AN MANIFESTACIONES CONTRA LEY MARCIAL\". The television is viewed from the side, with the blurred silhouette of a person's head and shoulder visible in the foreground on the left.  Character continuity: same character. Background continuity: same background view.", "width": 1024, "height": 1024}, {"name": "imagem000611", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000611.jpg", "prompt": "A close-up, rear-view shot of a person wearing a dark, tactical-style hooded garment or backpack with a prominent circular vent or mechanical detail on the back. The person is positioned in a desert environment with orange-toned sand dunes visible in the background. The lighting is warm and dim, casting deep shadows across the scene. The person is oriented away from the camera, with a portion of a strap or equipment visible on the left side of the frame.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000374", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000374.jpg", "prompt": "A close-up shot focusing on a person's hand resting on a green felt gaming table. The person is wearing a white long-sleeved shirt with the cuff visible. Next to the hand, there are several tall stacks of circular casino chips in various colors, including purple, white, green, and orange. The lighting is warm and focused on the table surface.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem000273", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem000273.jpg", "prompt": "A close-up, high-angle shot focusing on a person's torso and arm. The individual wears a white jacket with a dark-colored garment underneath. A large, jagged, dark red bloodstain is visible on the chest area of the white jacket, with droplets of blood splattered across the fabric. The person's arm is partially visible, and the scene is set against a dark, indistinct background.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}, {"name": "imagem001519", "reference": "/workspace/datasets/ctx_heldout/input_A/imagem001519.jpg", "prompt": "A wide shot of a spacious, high-ceilinged hall with light-colored stone walls. A row of black, wheeled display carts with glass cases sits on a red carpet, containing various artifacts including a sword and colorful gemstones. A small, black, winged creature with large yellow eyes and a grey belly stands on the red carpet in the foreground, facing the viewer. Two potted plants flank a set of double doors in the background.  Character continuity: same character. Background continuity: new view of the same background.", "width": 1024, "height": 1024}]}

backup_sessao_2026-08-06/configs/apex_ctx_512.toml

@@ -0,0 +1,66 @@
+# krea2_apex @512 — CTX (docs/KREA2_APEX_SPEC.md §2.7).
+#
+# Contrato (o que difere de todo braço anterior):
+#   - geometria frame_fit: ref no eixo de frame (pos[...,0]=1) e CENTRADA no
+#     grid do alvo com offsets FRACIONÁRIOS; sem width_shift, sem reescala
+#   - ref preprocessada com o ramo `fit` do comfyui-krea2edit (CROP_TOL 0.08,
+#     /16 floor com cap, crop-to-grid, bicubic+antialias)
+#   - reference_timestep='target' (o node calcula UM tvec para toda a sequência;
+#     um LoRA t=0 degrada lá em silêncio)
+#   - timestep_law='krea2_apex': t = sigmoid(s(N)·z + m(N)), peso da loss ≡ 1,
+#     amostragem estratificada golden-ratio. MUTUAMENTE EXCLUSIVO com flux_shift
+#   - LoRA em blocks + txtfusion + txtmlp (258 módulos)
+#   - live_text_encoding: sem cache de texto; jitter de grounding e clause
+#     dropout re-sorteados POR STEP
+
+output_dir = '/workspace/checkpoints/k2_apex_ctx'
+dataset = '/workspace/configs/apex_ctx_dataset.toml'
+epochs = 1000
+max_steps = 5000
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 250
+save_every_n_steps = 250
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO: o fit dá refs de shapes diferentes
+blocks_to_swap = 0
+
+[model]
+type = 'krea2_apex'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+timestep_law = 'krea2_apex'   # NÃO usar flux_shift/shift junto
+live_text_encoding = true
+
+[krea2_apex]
+reference_timestep = 'target'
+vl_prompt_style = 'plain'
+vl_longest_side = 768
+vl_grounding_jitter = [384, 768]   # agora POR STEP
+caption_dropout = 0.05
+clause_dropout_prob = 0.40
+clause_dropout_rate = 0.60
+keep_clauses = 1
+
+[adapter]
+type = 'lora'
+rank = 64
+# alpha: o trainer força alpha = rank (scaling 1). Não declarar.
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/apex_ctx_dataset.toml

@@ -0,0 +1,23 @@
+# krea2_apex — dataset CTX (1.247 pares A->B).
+#
+# resolutions=[512] com AR buckets: a lei de timestep do apex
+# (m/s dependentes de N) se auto-ajusta por bucket, então não há razão para
+# forçar quadrado. min/max AR 0.5..2.0 cobre o material do CTX.
+#
+# control_path = sources: a referência é pré-processada com o fit do node
+# (AR-preservado, crop-to-grid /16, bicubic+antialias), NÃO com
+# ImageOps.fit. As dimensões A<->B do CTX são iguais em ~100% dos pares, então
+# na prática o fit é no-op aqui — ele existe para comprar o contrato de
+# inferência.
+
+resolutions = [512]
+enable_ar_bucket = true
+min_ar = 0.5
+max_ar = 2.0
+num_ar_buckets = 7
+frame_buckets = [1]
+
+[[directory]]
+path = '/workspace/datasets/konrad_orig/ctx/targets'
+control_path = '/workspace/datasets/konrad_orig/ctx/sources'
+num_repeats = 1

backup_sessao_2026-08-06/configs/apex_ctx_smoke.toml

@@ -0,0 +1,66 @@
+# krea2_apex @512 — CTX (docs/KREA2_APEX_SPEC.md §2.7).
+#
+# Contrato (o que difere de todo braço anterior):
+#   - geometria frame_fit: ref no eixo de frame (pos[...,0]=1) e CENTRADA no
+#     grid do alvo com offsets FRACIONÁRIOS; sem width_shift, sem reescala
+#   - ref preprocessada com o ramo `fit` do comfyui-krea2edit (CROP_TOL 0.08,
+#     /16 floor com cap, crop-to-grid, bicubic+antialias)
+#   - reference_timestep='target' (o node calcula UM tvec para toda a sequência;
+#     um LoRA t=0 degrada lá em silêncio)
+#   - timestep_law='krea2_apex': t = sigmoid(s(N)·z + m(N)), peso da loss ≡ 1,
+#     amostragem estratificada golden-ratio. MUTUAMENTE EXCLUSIVO com flux_shift
+#   - LoRA em blocks + txtfusion + txtmlp (258 módulos)
+#   - live_text_encoding: sem cache de texto; jitter de grounding e clause
+#     dropout re-sorteados POR STEP
+
+output_dir = '/workspace/checkpoints/k2_apex_smoke'
+dataset = '/workspace/configs/apex_ctx_dataset.toml'
+epochs = 1000
+max_steps = 10
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 5
+save_every_n_steps = 10
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO: o fit dá refs de shapes diferentes
+blocks_to_swap = 0
+
+[model]
+type = 'krea2_apex'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+timestep_law = 'krea2_apex'   # NÃO usar flux_shift/shift junto
+live_text_encoding = true
+
+[krea2_apex]
+reference_timestep = 'target'
+vl_prompt_style = 'plain'
+vl_longest_side = 768
+vl_grounding_jitter = [384, 768]   # agora POR STEP
+caption_dropout = 0.05
+clause_dropout_prob = 0.40
+clause_dropout_rate = 0.60
+keep_clauses = 1
+
+[adapter]
+type = 'lora'
+rank = 64
+# alpha: o trainer força alpha = rank (scaling 1). Não declarar.
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/apex_ctx_smoke_resume.toml

@@ -0,0 +1,66 @@
+# krea2_apex @512 — CTX (docs/KREA2_APEX_SPEC.md §2.7).
+#
+# Contrato (o que difere de todo braço anterior):
+#   - geometria frame_fit: ref no eixo de frame (pos[...,0]=1) e CENTRADA no
+#     grid do alvo com offsets FRACIONÁRIOS; sem width_shift, sem reescala
+#   - ref preprocessada com o ramo `fit` do comfyui-krea2edit (CROP_TOL 0.08,
+#     /16 floor com cap, crop-to-grid, bicubic+antialias)
+#   - reference_timestep='target' (o node calcula UM tvec para toda a sequência;
+#     um LoRA t=0 degrada lá em silêncio)
+#   - timestep_law='krea2_apex': t = sigmoid(s(N)·z + m(N)), peso da loss ≡ 1,
+#     amostragem estratificada golden-ratio. MUTUAMENTE EXCLUSIVO com flux_shift
+#   - LoRA em blocks + txtfusion + txtmlp (258 módulos)
+#   - live_text_encoding: sem cache de texto; jitter de grounding e clause
+#     dropout re-sorteados POR STEP
+
+output_dir = '/workspace/checkpoints/k2_apex_smoke'
+dataset = '/workspace/configs/apex_ctx_dataset.toml'
+epochs = 1000
+max_steps = 20
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 5
+save_every_n_steps = 10
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO: o fit dá refs de shapes diferentes
+blocks_to_swap = 0
+
+[model]
+type = 'krea2_apex'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+timestep_law = 'krea2_apex'   # NÃO usar flux_shift/shift junto
+live_text_encoding = true
+
+[krea2_apex]
+reference_timestep = 'target'
+vl_prompt_style = 'plain'
+vl_longest_side = 768
+vl_grounding_jitter = [384, 768]   # agora POR STEP
+caption_dropout = 0.05
+clause_dropout_prob = 0.40
+clause_dropout_rate = 0.60
+keep_clauses = 1
+
+[adapter]
+type = 'lora'
+rank = 64
+# alpha: o trainer força alpha = rank (scaling 1). Não declarar.
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/armA_frame.toml

@@ -0,0 +1,62 @@
+# KREA2 edit saga 2026-07-29 — treino principal.
+#
+# Base: receita validada krea2_multiref_grounded (docs/KREA2_MULTIREF_RECEITA.md),
+# aqui com max_refs=1 (todos os datasets são pares A->B).
+# - width_shift + refs a t=0 + LoRA condition-only routada + grounding Qwen3-VL
+# - caption_dropout=0.1 (uncond do CFG; condition_dropout é proibido no grounded)
+# - txtfusion_rank=128 (canal de leitura reforçado — tarefa é semântica: editar)
+# - rank 64 (usuário pediu >32; dezenas de milhares de imagens)
+
+output_dir = '/workspace/checkpoints/armA_frame'
+dataset = '/workspace/configs/arms_ds4_384_dataset.toml'
+epochs = 1000
+max_steps = 500
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 100
+save_every_n_steps = 250
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO com multi_ref
+blocks_to_swap = 0           # usuário pediu velocidade; VRAM medida 20,1G com swap 8, cabe sem
+
+[model]
+type = 'krea2_multiref_grounded'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+flux_shift = true
+
+[krea2_multiref_grounded]
+max_refs = 1
+slot_axis = 'frame'
+position_mode = 'width_shift'
+reference_position_offset = 1.0
+reference_timestep = 'zero'
+condition_dropout = 0.0      # proibido no caminho grounded
+condition_only_lora = true
+vl_prompt_style = 'picture_n'
+vl_image_label = 'image'
+vl_image_max_pixels = 147456 # 384² — grounding; cache é o gargalo, medir no smoke
+caption_dropout = 0.0
+txtfusion_rank = 128
+
+[adapter]
+type = 'lora'
+rank = 64
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 2e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/armB_global.toml

@@ -0,0 +1,62 @@
+# KREA2 edit saga 2026-07-29 — treino principal.
+#
+# Base: receita validada krea2_multiref_grounded (docs/KREA2_MULTIREF_RECEITA.md),
+# aqui com max_refs=1 (todos os datasets são pares A->B).
+# - width_shift + refs a t=0 + LoRA condition-only routada + grounding Qwen3-VL
+# - caption_dropout=0.1 (uncond do CFG; condition_dropout é proibido no grounded)
+# - txtfusion_rank=128 (canal de leitura reforçado — tarefa é semântica: editar)
+# - rank 64 (usuário pediu >32; dezenas de milhares de imagens)
+
+output_dir = '/workspace/checkpoints/armB_global'
+dataset = '/workspace/configs/arms_ds4_384_dataset.toml'
+epochs = 1000
+max_steps = 500
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 100
+save_every_n_steps = 250
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO com multi_ref
+blocks_to_swap = 0           # usuário pediu velocidade; VRAM medida 20,1G com swap 8, cabe sem
+
+[model]
+type = 'krea2_multiref_grounded'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+flux_shift = true
+
+[krea2_multiref_grounded]
+max_refs = 1
+slot_axis = 'width'
+position_mode = 'width_shift'
+reference_position_offset = 1.0
+reference_timestep = 'zero'
+condition_dropout = 0.0      # proibido no caminho grounded
+condition_only_lora = false
+vl_prompt_style = 'picture_n'
+vl_image_label = 'image'
+vl_image_max_pixels = 147456 # 384² — grounding; cache é o gargalo, medir no smoke
+caption_dropout = 0.0
+txtfusion_rank = 128
+
+[adapter]
+type = 'lora'
+rank = 64
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 2e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/armC_v2.toml

@@ -0,0 +1,62 @@
+# KREA2 edit saga 2026-07-29 — treino principal.
+#
+# Base: receita validada krea2_multiref_grounded (docs/KREA2_MULTIREF_RECEITA.md),
+# aqui com max_refs=1 (todos os datasets são pares A->B).
+# - width_shift + refs a t=0 + LoRA condition-only routada + grounding Qwen3-VL
+# - caption_dropout=0.1 (uncond do CFG; condition_dropout é proibido no grounded)
+# - txtfusion_rank=128 (canal de leitura reforçado — tarefa é semântica: editar)
+# - rank 64 (usuário pediu >32; dezenas de milhares de imagens)
+
+output_dir = '/workspace/checkpoints/armC_v2'
+dataset = '/workspace/configs/arms_ds4_768_dataset.toml'
+epochs = 1000
+max_steps = 500
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 100
+save_every_n_steps = 250
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO com multi_ref
+blocks_to_swap = 0           # usuário pediu velocidade; VRAM medida 20,1G com swap 8, cabe sem
+
+[model]
+type = 'krea2_multiref_grounded'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+flux_shift = true
+
+[krea2_multiref_grounded]
+max_refs = 1
+slot_axis = 'width'
+position_mode = 'width_shift'
+reference_position_offset = 1.0
+reference_timestep = 'zero'
+condition_dropout = 0.0      # proibido no caminho grounded
+condition_only_lora = false
+vl_prompt_style = 'picture_n'
+vl_image_label = 'image'
+vl_longest_side = 768        # grounding do probe original
+caption_dropout = 0.0
+txtfusion_rank = 128
+
+[adapter]
+type = 'lora'
+rank = 64
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 2e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/arms_ds4_384_dataset.toml

@@ -0,0 +1,13 @@
+resolutions = [512]
+enable_ar_bucket = true
+min_ar = 0.5
+max_ar = 2.0
+num_ar_buckets = 7
+frame_buckets = [1]
+
+[[directory]]
+path = '/workspace/datasets/arms_ds4_384/target'
+control_path = '/workspace/datasets/arms_ds4_384/refs'
+multi_ref = true
+max_refs = 1
+num_repeats = 1

backup_sessao_2026-08-06/configs/arms_ds4_768_dataset.toml

@@ -0,0 +1,13 @@
+resolutions = [512]
+enable_ar_bucket = true
+min_ar = 0.5
+max_ar = 2.0
+num_ar_buckets = 7
+frame_buckets = [1]
+
+[[directory]]
+path = '/workspace/datasets/arms_ds4_768/target'
+control_path = '/workspace/datasets/arms_ds4_768/refs'
+multi_ref = true
+max_refs = 1
+num_repeats = 1

backup_sessao_2026-08-06/configs/k2_proximacena_smoke.toml

@@ -0,0 +1,57 @@
+# K2-PROXIMACENA-GROUNDED-V2 — retreino longo do krea2_omini_grounded com
+# captions de descrição pura (o contrato do probe original que funcionou).
+# Desvios do probe validado, com motivo:
+#   - caption_dropout 0.1 (melhoria §9 do OMINI_GROUNDED_SEGREDO: treina o uncond do CFG)
+#   - vl_image_max_pixels 384² (cache de texto: 768 custaria >700 GB p/ ~18k pares;
+#     384² é o desvio já validado no multiref)
+#   - sem vl_grounding_jitter (custo de cache; grounding fixo em 384²)
+#   - gradient_accumulation_steps 1 (pedido do usuário: sem batch fingido)
+#   - resolução 1024 com AR buckets (pedido do usuário)
+output_dir = '/workspace/checkpoints/k2_proximacena_smoke'
+dataset = '/workspace/configs/proximacena_smoke_dataset.toml'
+epochs = 1000
+max_steps = 12
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 2
+save_every_n_steps = 5
+checkpoint_every_n_minutes = 60
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1
+
+[model]
+type = 'krea2_omini_grounded'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+flux_shift = true
+
+[krea2_omini_grounded]
+position_mode = 'width_shift'
+reference_position_offset = 1.0
+reference_timestep = 'zero'
+condition_dropout = 0.0
+condition_only_lora = true
+vl_prompt_style = 'plain'
+vl_image_max_pixels = 147456
+caption_dropout = 0.1
+
+[adapter]
+type = 'lora'
+rank = 64
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/k2_proximacena_v2.toml

@@ -0,0 +1,57 @@
+# K2-PROXIMACENA-GROUNDED-V2 — retreino longo do krea2_omini_grounded com
+# captions de descrição pura (o contrato do probe original que funcionou).
+# Desvios do probe validado, com motivo:
+#   - caption_dropout 0.1 (melhoria §9 do OMINI_GROUNDED_SEGREDO: treina o uncond do CFG)
+#   - vl_image_max_pixels 384² (cache de texto: 768 custaria >700 GB p/ ~18k pares;
+#     384² é o desvio já validado no multiref)
+#   - sem vl_grounding_jitter (custo de cache; grounding fixo em 384²)
+#   - gradient_accumulation_steps 1 (pedido do usuário: sem batch fingido)
+#   - resolução 1024 com AR buckets (pedido do usuário)
+output_dir = '/workspace/checkpoints/k2_proximacena_v2'
+dataset = '/workspace/configs/proximacena_dataset.toml'
+epochs = 1000
+max_steps = 15000
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 100
+save_every_n_steps = 250
+checkpoint_every_n_minutes = 60
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1
+
+[model]
+type = 'krea2_omini_grounded'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+flux_shift = true
+
+[krea2_omini_grounded]
+position_mode = 'width_shift'
+reference_position_offset = 1.0
+reference_timestep = 'zero'
+condition_dropout = 0.0
+condition_only_lora = true
+vl_prompt_style = 'plain'
+vl_image_max_pixels = 147456
+caption_dropout = 0.1
+
+[adapter]
+type = 'lora'
+rank = 64
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/konrad_k2.yaml

@@ -0,0 +1,42 @@
+# konrad-k2-test1real — receita do krea2edit-trainer (conradlocke) COMO ELA É,
+# no dataset recaptionado (10.991 pares + ds4 curado com peso extra).
+# Único desvio da receita: nossas captions são DESCRIÇÕES da imagem B, não
+# instruções de edição (decisão do usuário: mesmo dataset de agora).
+job: extension
+config:
+  name: "konrad_k2_test1real"
+  process:
+    - type: "sd_trainer"
+      training_folder: "/workspace/checkpoints/konrad"
+      device: cuda:0
+      logging: { use_wandb: false }
+      network: { type: "lora", linear: 64, linear_alpha: 64 }
+      save: { dtype: float16, save_every: 250, max_step_saves_to_keep: 8 }
+      datasets:
+        - folder_path: "/workspace/datasets/konrad/targets"
+          control_path: ["/workspace/datasets/konrad/sources"]
+          num_repeats: 1
+          caption_ext: "txt"
+          caption_dropout_rate: 0.0
+          resolution: [512]
+          buckets: true
+          cache_latents_to_disk: true
+          cache_text_embeddings: false
+        - folder_path: "/workspace/datasets/konrad_ds4/targets"
+          control_path: ["/workspace/datasets/konrad_ds4/sources"]
+          num_repeats: 2
+          caption_ext: "txt"
+          caption_dropout_rate: 0.0
+          resolution: [512]
+          buckets: true
+          cache_latents_to_disk: true
+          cache_text_embeddings: false
+      train: { batch_size: 1, steps: 15000, gradient_accumulation: 1, train_unet: true,
+        train_text_encoder: false, gradient_checkpointing: true, disable_sampling: true,
+        noise_scheduler: "flowmatch", timestep_type: "weighted", optimizer: "adamw8bit",
+        lr: 0.0001, lr_scheduler: "constant", dtype: bf16 }
+      model: { name_or_path: "/workspace/models/krea2/raw_oficial/raw.safetensors", arch: "krea2_edit",
+        quantize: true, quantize_te: true,
+        model_kwargs: { text_encoder_path: "Qwen/Qwen3-VL-4B-Instruct", fit_refs: true } }
+      sample: { sample_every: 100000, samples: [] }
+meta: { name: "konrad_k2_test1real", version: "1.0" }

backup_sessao_2026-08-06/configs/konrad_k2_arm2.yaml

@@ -0,0 +1,46 @@
+# konrad-k2-test1real — receita do krea2edit-trainer (conradlocke) COMO ELA É,
+# no dataset recaptionado (10.991 pares + ds4 curado com peso extra).
+# Único desvio da receita: nossas captions são DESCRIÇÕES da imagem B, não
+# instruções de edição (decisão do usuário: mesmo dataset de agora).
+job: extension
+config:
+  name: "konrad_k2_arm2"
+  process:
+    - type: "sd_trainer"
+      training_folder: "/workspace/checkpoints/konrad"
+      device: cuda:0
+      logging: { use_wandb: false }
+      network: { type: "lora", linear: 64, linear_alpha: 64 }
+      save: { dtype: float16, save_every: 250, max_step_saves_to_keep: 8 }
+      datasets:
+        - folder_path: "/workspace/datasets/konrad/targets"
+          control_path: ["/workspace/datasets/konrad/sources"]
+          num_repeats: 1
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05
+          token_dropout_rate: 0.5
+          keep_tokens: 1
+          resolution: [512]
+          buckets: true
+          cache_latents_to_disk: true
+          cache_text_embeddings: false
+        - folder_path: "/workspace/datasets/konrad_ds4/targets"
+          control_path: ["/workspace/datasets/konrad_ds4/sources"]
+          num_repeats: 2
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05
+          token_dropout_rate: 0.5
+          keep_tokens: 1
+          resolution: [512]
+          buckets: true
+          cache_latents_to_disk: true
+          cache_text_embeddings: false
+      train: { batch_size: 1, steps: 15000, gradient_accumulation: 1, train_unet: true,
+        train_text_encoder: false, gradient_checkpointing: true, disable_sampling: true,
+        noise_scheduler: "flowmatch", timestep_type: "weighted", optimizer: "adamw8bit",
+        lr: 0.0001, lr_scheduler: "constant", dtype: bf16 }
+      model: { name_or_path: "/workspace/models/krea2/nk2e/raw_bf16.safetensors", arch: "krea2_edit",
+        quantize: true, quantize_te: true,
+        model_kwargs: { text_encoder_path: "Qwen/Qwen3-VL-4B-Instruct", fit_refs: true } }
+      sample: { sample_every: 100000, samples: [] }
+meta: { name: "konrad_k2_arm2", version: "1.0" }

backup_sessao_2026-08-06/configs/konrad_orig_samples.json

@@ -0,0 +1,47 @@
+{
+ "comment": "6 samples fixos do run konrad_orig (captions ORIGINAIS de instrução): 3 pares ctx + 3 selecionados pelo usuario",
+ "samples": [
+  {
+   "name": "ctx_imagem000001",
+   "reference": "/workspace/datasets/ctx_original/input_A/imagem000001.jpg",
+   "prompt": "A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.",
+   "width": 512,
+   "height": 512
+  },
+  {
+   "name": "ctx_imagem000836",
+   "reference": "/workspace/datasets/ctx_original/input_A/imagem000836.jpg",
+   "prompt": "A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text \"ALIENS! ABDUCT ME!!!\" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.",
+   "width": 704,
+   "height": 384
+  },
+  {
+   "name": "ctx_imagem001768",
+   "reference": "/workspace/datasets/ctx_original/input_A/imagem001768.jpg",
+   "prompt": "A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.",
+   "width": 768,
+   "height": 768
+  },
+  {
+   "name": "goblin_fight",
+   "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image1.jpg",
+   "prompt": "dynamic wide shot from a low angle, a young woman with long wavy blonde hair and amber eyes standing in a fighting stance, both hands gripping the hilt of a raised steel sword angled across her body, knees bent and one foot slid back, furrowed brows and fierce gaze locked on her opponent, wearing a cream off-shoulder blouse with a laced blue corset, facing a hulking muscular goblin with mottled green skin, pointed ears and bared tusks, its clawed fists raised mid-charge, dust kicked up between them, weathered stone wall and a wooden pillar with hanging ivy behind them, warm sunlight raking across the stone",
+   "width": 1152,
+   "height": 896
+  },
+  {
+   "name": "pov_bandaged_hands",
+   "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image2.jpg",
+   "prompt": "medium close-up from a first-person viewpoint, a girl with short choppy blue hair and small braids looking up toward the camera with a soft concerned expression, lips slightly parted, wearing a teal short-sleeved shirt over a pink striped undershirt, seated cross-legged on a patched mattress, in the lower frame two forearms wrapped in stained white bandages reaching toward her, the bandaged hands open with palms up, a dim wooden attic room around them lit by warm orange light from a round window, small trinkets dangling from the beams",
+   "width": 1024,
+   "height": 1024
+  },
+  {
+   "name": "orc_hug_cloud",
+   "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image3.webp",
+   "prompt": "wide shot from the side at eye level, a young woman with a curly pink bob, blue eyes, clear angular goggles pushed up on her face and ornate white and gold armor, on her tiptoes with both arms wrapped tightly around the neck of a towering broad-shouldered orc with weathered green skin and small tusks, her cheek pressed against his chest, eyes closed in a wide open-mouth smile, the orc's thick arms folded gently around her back, both standing on the rounded top of a dense white cloud, more clouds drifting below against a deep blue sky, bright sunlight from above",
+   "width": 896,
+   "height": 512
+  }
+ ]
+}

backup_sessao_2026-08-06/configs/konrad_smoke.yaml

@@ -0,0 +1,42 @@
+# konrad-k2-test1real — receita do krea2edit-trainer (conradlocke) COMO ELA É,
+# no dataset recaptionado (10.991 pares + ds4 curado com peso extra).
+# Único desvio da receita: nossas captions são DESCRIÇÕES da imagem B, não
+# instruções de edição (decisão do usuário: mesmo dataset de agora).
+job: extension
+config:
+  name: "konrad_smoke"
+  process:
+    - type: "sd_trainer"
+      training_folder: "/workspace/checkpoints/konrad"
+      device: cuda:0
+      logging: { use_wandb: false }
+      network: { type: "lora", linear: 64, linear_alpha: 64 }
+      save: { dtype: float16, save_every: 10, max_step_saves_to_keep: 8 }
+      datasets:
+        - folder_path: "/workspace/datasets/konrad/targets"
+          control_path: ["/workspace/datasets/konrad/sources"]
+          num_repeats: 1
+          caption_ext: "txt"
+          caption_dropout_rate: 0.0
+          resolution: [512]
+          buckets: true
+          cache_latents_to_disk: true
+          cache_text_embeddings: false
+        - folder_path: "/workspace/datasets/konrad_ds4/targets"
+          control_path: ["/workspace/datasets/konrad_ds4/sources"]
+          num_repeats: 2
+          caption_ext: "txt"
+          caption_dropout_rate: 0.0
+          resolution: [512]
+          buckets: true
+          cache_latents_to_disk: true
+          cache_text_embeddings: false
+      train: { batch_size: 1, steps: 12, gradient_accumulation: 1, train_unet: true,
+        train_text_encoder: false, gradient_checkpointing: true, disable_sampling: true,
+        noise_scheduler: "flowmatch", timestep_type: "weighted", optimizer: "adamw8bit",
+        lr: 0.0001, lr_scheduler: "constant", dtype: bf16 }
+      model: { name_or_path: "/workspace/models/krea2/nk2e/raw_bf16.safetensors", arch: "krea2_edit",
+        quantize: true, quantize_te: true,
+        model_kwargs: { text_encoder_path: "Qwen/Qwen3-VL-4B-Instruct", fit_refs: true } }
+      sample: { sample_every: 100000, samples: [] }
+meta: { name: "konrad_smoke", version: "1.0" }

backup_sessao_2026-08-06/configs/nk2e_bat2_prompts_json.txt

@@ -0,0 +1,3 @@
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image1.jpg | dynamic wide shot from a low angle, a young woman with long wavy blonde hair and amber eyes standing in a fighting stance, both hands gripping the hilt of a raised steel sword angled across her body, knees bent and one foot slid back, furrowed brows and fierce gaze locked on her opponent, wearing a cream off-shoulder blouse with a laced blue corset, facing a hulking muscular goblin with mottled green skin, pointed ears and bared tusks, its clawed fists raised mid-charge, dust kicked up between them, weathered stone wall and a wooden pillar with hanging ivy behind them, warm sunlight raking across the stone
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image2.jpg | medium close-up from a first-person viewpoint, a girl with short choppy blue hair and small braids looking up toward the camera with a soft concerned expression, lips slightly parted, wearing a teal short-sleeved shirt over a pink striped undershirt, seated cross-legged on a patched mattress, in the lower frame two forearms wrapped in stained white bandages reaching toward her, the bandaged hands open with palms up, a dim wooden attic room around them lit by warm orange light from a round window, small trinkets dangling from the beams
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image3.webp | wide shot from the side at eye level, a young woman with a curly pink bob, blue eyes, clear angular goggles pushed up on her face and ornate white and gold armor, on her tiptoes with both arms wrapped tightly around the neck of a towering broad-shouldered orc with weathered green skin and small tusks, her cheek pressed against his chest, eyes closed in a wide open-mouth smile, the orc's thick arms folded gently around her back, both standing on the rounded top of a dense white cloud, more clouds drifting below against a deep blue sky, bright sunlight from above

backup_sessao_2026-08-06/configs/nk2e_bat2_txt_originais.txt

@@ -0,0 +1,3 @@
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image1.jpg | Essa personagem lutando contra um gobling grande e musculoso, ela ta no mesmo envirorment, segurando uma espada e pronta para luta, num shot maneiro cinematico
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image2.jpg | foco na garota de cabelo azul, medium close-up, podemos ver no pov da garota de cabelo vermelho ou algo assim. podemos ver no pov as mãos enfaixadas de curativos e etc.
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image3.webp | essa garota sorrindo dando um abraço em um ogro/orc, a visão é de lado, podemos ver a cena inteira, eles estão em cima de uma nuvem, felizes como se estivessemos a dias sem se verem, dando um abraço fofo

backup_sessao_2026-08-06/configs/nk2e_ctx_samples.txt

@@ -0,0 +1,3 @@
+/workspace/datasets/ctx_original/input_A/imagem000001.jpg | A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem000836.jpg | A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text "ALIENS! ABDUCT ME!!!" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem001768.jpg | A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.

backup_sessao_2026-08-06/configs/nk2e_ctx_samples_11.txt

@@ -0,0 +1,11 @@
+/workspace/datasets/ctx_original/input_A/imagem000001.jpg | A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem000836.jpg | A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text "ALIENS! ABDUCT ME!!!" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem001768.jpg | A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000075.jpg | A medium shot of a white and grey robotic figure with a spherical head and glowing cyan circular eyes, set against a cloudy grey sky. The robot holds a small, translucent blue crab with red-tipped claws on its palm. A seagull is captured in motion, flying directly in front of the robot's face, creating a slight motion blur. The robot has segmented arms and a smooth, rounded torso. The background is an overcast sky with soft, diffused light.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000273.jpg | A close-up, high-angle shot focusing on a person's torso and arm. The individual wears a white jacket with a dark-colored garment underneath. A large, jagged, dark red bloodstain is visible on the chest area of the white jacket, with droplets of blood splattered across the fabric. The person's arm is partially visible, and the scene is set against a dark, indistinct background.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000294.jpg | A tight close-up shot of a young person with dark hair, their face contorted in an intense expression of distress. Their mouth is wide open in a cry, and large, glistening tears stream from their wide, light-colored eyes. The character is positioned on the left side of the frame, with their head tilted slightly. A large, pale, hand-like appendage with multiple fingers extends from the right side of the frame, reaching toward the character's face. The background is a blurred, abstract blend of deep purple and bright magenta light, suggesting a high-energy or supernatural environment.  Character continuity: same character. Background continuity: new background.
+/workspace/datasets/ctx_heldout/input_A/imagem000374.jpg | A close-up shot focusing on a person's hand resting on a green felt gaming table. The person is wearing a white long-sleeved shirt with the cuff visible. Next to the hand, there are several tall stacks of circular casino chips in various colors, including purple, white, green, and orange. The lighting is warm and focused on the table surface.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000611.jpg | A close-up, rear-view shot of a person wearing a dark, tactical-style hooded garment or backpack with a prominent circular vent or mechanical detail on the back. The person is positioned in a desert environment with orange-toned sand dunes visible in the background. The lighting is warm and dim, casting deep shadows across the scene. The person is oriented away from the camera, with a portion of a strap or equipment visible on the left side of the frame.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000663.jpg | A close-up, slightly angled view of an old television screen displaying a live news broadcast of a night protest in a city street. The screen shows a large crowd of people gathered on a dark asphalt road at night, illuminated by streetlights and green traffic signals. A circular traffic sign with an upward arrow is visible on the right side of the screen. A news ticker at the bottom of the screen displays the text "BIOS AUMENTAN A VELOCIDAD ALARMANTE" and "AN MANIFESTACIONES CONTRA LEY MARCIAL". The television is viewed from the side, with the blurred silhouette of a person's head and shoulder visible in the foreground on the left.  Character continuity: same character. Background continuity: same background view.
+/workspace/datasets/ctx_heldout/input_A/imagem000759.jpg | A high-angle, close-up shot looking down into an open stone sarcophagus. A man with short, light-colored hair lies peacefully inside with his eyes closed, his hands folded over his chest. He wears a dark blue, high-collared military-style jacket with gold trim and a star-shaped pendant. The interior of the sarcophagus is lined with brown, padded fabric. On either side of the frame, the gloved hands of two individuals are visible, gripping the edges of the sarcophagus. Several small, light-colored flower petals are scattered across the man and the interior of the coffin.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem001519.jpg | A wide shot of a spacious, high-ceilinged hall with light-colored stone walls. A row of black, wheeled display carts with glass cases sits on a red carpet, containing various artifacts including a sword and colorful gemstones. A small, black, winged creature with large yellow eyes and a grey belly stands on the red carpet in the foreground, facing the viewer. Two potted plants flank a set of double doors in the background.  Character continuity: same character. Background continuity: new view of the same background.

backup_sessao_2026-08-06/configs/nk2e_ds4novo_samples11.txt

@@ -0,0 +1,11 @@
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000001.jpg | eye-level medium shot at a professional hockey rink during a face-off, the referee in a black and white striped jersey stands in the center ready to drop the puck, flanked by two young players in blue and mint-green jerseys crouching with their hockey sticks on the ice, two smaller teammates wait in the background on either side of the referee, stadium seating filled with spectators is visible in the darkened background under bright spotlights, the foreground is framed by the blurred silhouettes of two other players facing the scene from the sides
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000836.jpg | high-angle long shot looking down at a flat expanse of sand, a small child with a light tan complexion in a green t-shirt, dark shorts, and a white sun hat lying flat on their back in the center, arms and legs spread out, one hand holding a green snack, a small teal container placed near their legs, surrounding the child is a large, intricate geometric design drawn in the sand featuring lines, circles, and arrows, with the words aliens, and abduct me, etched in a circle, bright daylight illuminating the sandy landscape
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem001768.jpg | medium shot at eye-level, a woman with blonde hair in a ponytail wearing a white shirt and green apron sits at a long wooden table, her body angled away from the camera, a person in a bright purple hooded sweatshirt stands opposite her, bowing deeply with their head down and long black hair covering their face, hands held together in a prayer-like gesture gripping a thin stick, a wooden table between them holds a stack of wooden bowls and a goblet, stone walls with a large wooden door and a window provide the background, rustic interior with wooden barrels on the floor
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000294.jpg | close-up from a low angle, a character with messy black hair and wide brown eyes screaming in distress, tears welling and spilling from the eyes, mouth wide open showing teeth, pale skin rendered with purple shadows, positioned on the left, a large hand or claw-like appendage with splayed fingers reaching in from the right, the hand colored in vibrant shades of purple and pink, abstract background filled with motion streaks and light flares, the entire frame dominated by intense facial expression and the approaching hand, vibrant violet and magenta light casting across the figures
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000075.jpg | medium shot at a slightly low angle, a large white and metallic robot with a spherical head and glowing cyan eyes stands against a backdrop of heavy, grey, storm-filled clouds, a motion-blurred seagull flies directly across the robot's face, the robot keeps its left arm bent, hand held open and palm upward to cradle a tiny, bright blue crab-like creature with prominent red pincers, the robot's chest and segmented shoulder plates are smooth and light-colored, the lighting is diffuse and cool-toned, capturing the contrast between the massive, immobile machine and the fleeting movement of the bird
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000759.jpg | medium shot at a straight-on view, a pale man with short white hair and eyes closed resting inside a wooden coffin, wearing a dark green high-collared coat with gold trim over a light blue collared shirt with a star pendant, hands crossed neatly over his chest in white gloves, scattered pink flower petals resting on his coat and the rumpled dark grey fabric lining the coffin, several hands in gloves visible at the far left and right edges of the frame gripping vertical wooden coffin bars, soft ambient lighting reflecting off his face and the wood grain
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000663.jpg | medium shot from behind the shoulder of a person watching a television, a dark silhouette of a figure visible in the left foreground, the television screen showing a grainy night broadcast of a large crowd protesting in a city street, tall illuminated buildings and glowing streetlights in the background, a red circular traffic sign perched on a pole on the right side, wet asphalt reflecting blue and white light, scrolling news text at the bottom of the screen reporting on martial law protests, screen glow illuminating the edges of the room in cool blue tones
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000611.jpg | medium close-up from a low angle, a mysterious figure draped in a heavy, dark, structured cloak sits amidst a vast, barren desert of orange-hued, wind-swept sand dunes, the figure's face obscured by a stiff, angular hood that reveals a recessed mechanical circular grate deep inside, a strap hanging from the side of the garment, warm dim lighting casting long shadows across the sandy floor, the scene bathed in a deep amber glow, the horizon line blurred in the soft, dark background, the overall atmosphere quiet, desolate, and arid
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000374.jpg | close-up, eye-level side view of a bare foot with fair skin standing vertically against a tall stack of purple and white poker chips on a green felt table, the heel of the foot resting directly on the felt surface, a person's forearm clad in a crisp white long-sleeved shirt sleeve entering from the right, the arm resting flat on the green table near the stacks, other stacks of green, orange, and blue poker chips arranged nearby, soft directional lighting casting subtle shadows on the white shirt fabric and the felt table surface
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000273.jpg | extreme close-up on a metallic surface, a thick and viscous splash of bright red blood flows downward over a dark circular gear-like component, bright droplets of blood splattered across adjacent grey armor plating in the dark environment, a sharp white jagged edge visible in the upper left corner partially stained with blood, the lighting is muted and dim emphasizing the high contrast of the vivid red fluid against the cool tones of the dark machine parts, a concentrated focus on the visceral pooling of liquid against the textured industrial surface
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem001519.jpg | wide shot at eye-level, a small black chibi-style creature with large yellow eyes and tiny wings stands on a vibrant red carpet in the foreground, facing the viewer with a neutral expression, positioned in a grand, airy hall with beige paneled walls, a row of black wooden wheeled carts with glass display cases holds a silver sword, multiple glowing jewels, and a framed painting, tall green plants stand near the carts and beside a dark wooden double door in the back, all illuminated by bright, soft indoor lighting, the floor is polished and clean

backup_sessao_2026-08-06/configs/nk2e_ds4novo_samples11_shufref.txt

@@ -0,0 +1,11 @@
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000836.jpg | eye-level medium shot at a professional hockey rink during a face-off, the referee in a black and white striped jersey stands in the center ready to drop the puck, flanked by two young players in blue and mint-green jerseys crouching with their hockey sticks on the ice, two smaller teammates wait in the background on either side of the referee, stadium seating filled with spectators is visible in the darkened background under bright spotlights, the foreground is framed by the blurred silhouettes of two other players facing the scene from the sides
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem001768.jpg | high-angle long shot looking down at a flat expanse of sand, a small child with a light tan complexion in a green t-shirt, dark shorts, and a white sun hat lying flat on their back in the center, arms and legs spread out, one hand holding a green snack, a small teal container placed near their legs, surrounding the child is a large, intricate geometric design drawn in the sand featuring lines, circles, and arrows, with the words aliens, and abduct me, etched in a circle, bright daylight illuminating the sandy landscape
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000294.jpg | medium shot at eye-level, a woman with blonde hair in a ponytail wearing a white shirt and green apron sits at a long wooden table, her body angled away from the camera, a person in a bright purple hooded sweatshirt stands opposite her, bowing deeply with their head down and long black hair covering their face, hands held together in a prayer-like gesture gripping a thin stick, a wooden table between them holds a stack of wooden bowls and a goblet, stone walls with a large wooden door and a window provide the background, rustic interior with wooden barrels on the floor
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000075.jpg | close-up from a low angle, a character with messy black hair and wide brown eyes screaming in distress, tears welling and spilling from the eyes, mouth wide open showing teeth, pale skin rendered with purple shadows, positioned on the left, a large hand or claw-like appendage with splayed fingers reaching in from the right, the hand colored in vibrant shades of purple and pink, abstract background filled with motion streaks and light flares, the entire frame dominated by intense facial expression and the approaching hand, vibrant violet and magenta light casting across the figures
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000759.jpg | medium shot at a slightly low angle, a large white and metallic robot with a spherical head and glowing cyan eyes stands against a backdrop of heavy, grey, storm-filled clouds, a motion-blurred seagull flies directly across the robot's face, the robot keeps its left arm bent, hand held open and palm upward to cradle a tiny, bright blue crab-like creature with prominent red pincers, the robot's chest and segmented shoulder plates are smooth and light-colored, the lighting is diffuse and cool-toned, capturing the contrast between the massive, immobile machine and the fleeting movement of the bird
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000663.jpg | medium shot at a straight-on view, a pale man with short white hair and eyes closed resting inside a wooden coffin, wearing a dark green high-collared coat with gold trim over a light blue collared shirt with a star pendant, hands crossed neatly over his chest in white gloves, scattered pink flower petals resting on his coat and the rumpled dark grey fabric lining the coffin, several hands in gloves visible at the far left and right edges of the frame gripping vertical wooden coffin bars, soft ambient lighting reflecting off his face and the wood grain
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000611.jpg | medium shot from behind the shoulder of a person watching a television, a dark silhouette of a figure visible in the left foreground, the television screen showing a grainy night broadcast of a large crowd protesting in a city street, tall illuminated buildings and glowing streetlights in the background, a red circular traffic sign perched on a pole on the right side, wet asphalt reflecting blue and white light, scrolling news text at the bottom of the screen reporting on martial law protests, screen glow illuminating the edges of the room in cool blue tones
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000374.jpg | medium close-up from a low angle, a mysterious figure draped in a heavy, dark, structured cloak sits amidst a vast, barren desert of orange-hued, wind-swept sand dunes, the figure's face obscured by a stiff, angular hood that reveals a recessed mechanical circular grate deep inside, a strap hanging from the side of the garment, warm dim lighting casting long shadows across the sandy floor, the scene bathed in a deep amber glow, the horizon line blurred in the soft, dark background, the overall atmosphere quiet, desolate, and arid
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000273.jpg | close-up, eye-level side view of a bare foot with fair skin standing vertically against a tall stack of purple and white poker chips on a green felt table, the heel of the foot resting directly on the felt surface, a person's forearm clad in a crisp white long-sleeved shirt sleeve entering from the right, the arm resting flat on the green table near the stacks, other stacks of green, orange, and blue poker chips arranged nearby, soft directional lighting casting subtle shadows on the white shirt fabric and the felt table surface
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem001519.jpg | extreme close-up on a metallic surface, a thick and viscous splash of bright red blood flows downward over a dark circular gear-like component, bright droplets of blood splattered across adjacent grey armor plating in the dark environment, a sharp white jagged edge visible in the upper left corner partially stained with blood, the lighting is muted and dim emphasizing the high contrast of the vivid red fluid against the cool tones of the dark machine parts, a concentrated focus on the visceral pooling of liquid against the textured industrial surface
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000001.jpg | wide shot at eye-level, a small black chibi-style creature with large yellow eyes and tiny wings stands on a vibrant red carpet in the foreground, facing the viewer with a neutral expression, positioned in a grand, airy hall with beige paneled walls, a row of black wooden wheeled carts with glass display cases holds a silver sword, multiple glowing jewels, and a framed painting, tall green plants stand near the carts and beside a dark wooden double door in the back, all illuminated by bright, soft indoor lighting, the floor is polished and clean

backup_sessao_2026-08-06/configs/nk2e_ds4novo_samples3.txt

@@ -0,0 +1,3 @@
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000001.jpg | eye-level medium shot at a professional hockey rink during a face-off, the referee in a black and white striped jersey stands in the center ready to drop the puck, flanked by two young players in blue and mint-green jerseys crouching with their hockey sticks on the ice, two smaller teammates wait in the background on either side of the referee, stadium seating filled with spectators is visible in the darkened background under bright spotlights, the foreground is framed by the blurred silhouettes of two other players facing the scene from the sides
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000836.jpg | high-angle long shot looking down at a flat expanse of sand, a small child with a light tan complexion in a green t-shirt, dark shorts, and a white sun hat lying flat on their back in the center, arms and legs spread out, one hand holding a green snack, a small teal container placed near their legs, surrounding the child is a large, intricate geometric design drawn in the sand featuring lines, circles, and arrows, with the words aliens, and abduct me, etched in a circle, bright daylight illuminating the sandy landscape
+/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem001768.jpg | medium shot at eye-level, a woman with blonde hair in a ponytail wearing a white shirt and green apron sits at a long wooden table, her body angled away from the camera, a person in a bright purple hooded sweatshirt stands opposite her, bowing deeply with their head down and long black hair covering their face, hands held together in a prayer-like gesture gripping a thin stick, a wooden table between them holds a stack of wooden bowls and a goblet, stone walls with a large wooden door and a window provide the background, rustic interior with wooden barrels on the floor

backup_sessao_2026-08-06/configs/nk2e_orig_samples11.txt

@@ -0,0 +1,11 @@
+/workspace/datasets/ctx_original/input_A/imagem000001.jpg | A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem000836.jpg | A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text "ALIENS! ABDUCT ME!!!" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem001768.jpg | A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000294.jpg | A tight close-up shot of a young person with dark hair, their face contorted in an intense expression of distress. Their mouth is wide open in a cry, and large, glistening tears stream from their wide, light-colored eyes. The character is positioned on the left side of the frame, with their head tilted slightly. A large, pale, hand-like appendage with multiple fingers extends from the right side of the frame, reaching toward the character's face. The background is a blurred, abstract blend of deep purple and bright magenta light, suggesting a high-energy or supernatural environment.  Character continuity: same character. Background continuity: new background.
+/workspace/datasets/ctx_heldout/input_A/imagem000075.jpg | A medium shot of a white and grey robotic figure with a spherical head and glowing cyan circular eyes, set against a cloudy grey sky. The robot holds a small, translucent blue crab with red-tipped claws on its palm. A seagull is captured in motion, flying directly in front of the robot's face, creating a slight motion blur. The robot has segmented arms and a smooth, rounded torso. The background is an overcast sky with soft, diffused light.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000759.jpg | A high-angle, close-up shot looking down into an open stone sarcophagus. A man with short, light-colored hair lies peacefully inside with his eyes closed, his hands folded over his chest. He wears a dark blue, high-collared military-style jacket with gold trim and a star-shaped pendant. The interior of the sarcophagus is lined with brown, padded fabric. On either side of the frame, the gloved hands of two individuals are visible, gripping the edges of the sarcophagus. Several small, light-colored flower petals are scattered across the man and the interior of the coffin.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000663.jpg | A close-up, slightly angled view of an old television screen displaying a live news broadcast of a night protest in a city street. The screen shows a large crowd of people gathered on a dark asphalt road at night, illuminated by streetlights and green traffic signals. A circular traffic sign with an upward arrow is visible on the right side of the screen. A news ticker at the bottom of the screen displays the text "BIOS AUMENTAN A VELOCIDAD ALARMANTE" and "AN MANIFESTACIONES CONTRA LEY MARCIAL". The television is viewed from the side, with the blurred silhouette of a person's head and shoulder visible in the foreground on the left.  Character continuity: same character. Background continuity: same background view.
+/workspace/datasets/ctx_heldout/input_A/imagem000611.jpg | A close-up, rear-view shot of a person wearing a dark, tactical-style hooded garment or backpack with a prominent circular vent or mechanical detail on the back. The person is positioned in a desert environment with orange-toned sand dunes visible in the background. The lighting is warm and dim, casting deep shadows across the scene. The person is oriented away from the camera, with a portion of a strap or equipment visible on the left side of the frame.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000374.jpg | A close-up shot focusing on a person's hand resting on a green felt gaming table. The person is wearing a white long-sleeved shirt with the cuff visible. Next to the hand, there are several tall stacks of circular casino chips in various colors, including purple, white, green, and orange. The lighting is warm and focused on the table surface.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000273.jpg | A close-up, high-angle shot focusing on a person's torso and arm. The individual wears a white jacket with a dark-colored garment underneath. A large, jagged, dark red bloodstain is visible on the chest area of the white jacket, with droplets of blood splattered across the fabric. The person's arm is partially visible, and the scene is set against a dark, indistinct background.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem001519.jpg | A wide shot of a spacious, high-ceilinged hall with light-colored stone walls. A row of black, wheeled display carts with glass cases sits on a red carpet, containing various artifacts including a sword and colorful gemstones. A small, black, winged creature with large yellow eyes and a grey belly stands on the red carpet in the foreground, facing the viewer. Two potted plants flank a set of double doors in the background.  Character continuity: same character. Background continuity: new view of the same background.

backup_sessao_2026-08-06/configs/nk2e_orig_samples11_shufref.txt

@@ -0,0 +1,11 @@
+/workspace/datasets/ctx_original/input_A/imagem000836.jpg | A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem001768.jpg | A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text "ALIENS! ABDUCT ME!!!" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000294.jpg | A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000075.jpg | A tight close-up shot of a young person with dark hair, their face contorted in an intense expression of distress. Their mouth is wide open in a cry, and large, glistening tears stream from their wide, light-colored eyes. The character is positioned on the left side of the frame, with their head tilted slightly. A large, pale, hand-like appendage with multiple fingers extends from the right side of the frame, reaching toward the character's face. The background is a blurred, abstract blend of deep purple and bright magenta light, suggesting a high-energy or supernatural environment.  Character continuity: same character. Background continuity: new background.
+/workspace/datasets/ctx_heldout/input_A/imagem000759.jpg | A medium shot of a white and grey robotic figure with a spherical head and glowing cyan circular eyes, set against a cloudy grey sky. The robot holds a small, translucent blue crab with red-tipped claws on its palm. A seagull is captured in motion, flying directly in front of the robot's face, creating a slight motion blur. The robot has segmented arms and a smooth, rounded torso. The background is an overcast sky with soft, diffused light.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000663.jpg | A high-angle, close-up shot looking down into an open stone sarcophagus. A man with short, light-colored hair lies peacefully inside with his eyes closed, his hands folded over his chest. He wears a dark blue, high-collared military-style jacket with gold trim and a star-shaped pendant. The interior of the sarcophagus is lined with brown, padded fabric. On either side of the frame, the gloved hands of two individuals are visible, gripping the edges of the sarcophagus. Several small, light-colored flower petals are scattered across the man and the interior of the coffin.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000611.jpg | A close-up, slightly angled view of an old television screen displaying a live news broadcast of a night protest in a city street. The screen shows a large crowd of people gathered on a dark asphalt road at night, illuminated by streetlights and green traffic signals. A circular traffic sign with an upward arrow is visible on the right side of the screen. A news ticker at the bottom of the screen displays the text "BIOS AUMENTAN A VELOCIDAD ALARMANTE" and "AN MANIFESTACIONES CONTRA LEY MARCIAL". The television is viewed from the side, with the blurred silhouette of a person's head and shoulder visible in the foreground on the left.  Character continuity: same character. Background continuity: same background view.
+/workspace/datasets/ctx_heldout/input_A/imagem000374.jpg | A close-up, rear-view shot of a person wearing a dark, tactical-style hooded garment or backpack with a prominent circular vent or mechanical detail on the back. The person is positioned in a desert environment with orange-toned sand dunes visible in the background. The lighting is warm and dim, casting deep shadows across the scene. The person is oriented away from the camera, with a portion of a strap or equipment visible on the left side of the frame.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000273.jpg | A close-up shot focusing on a person's hand resting on a green felt gaming table. The person is wearing a white long-sleeved shirt with the cuff visible. Next to the hand, there are several tall stacks of circular casino chips in various colors, including purple, white, green, and orange. The lighting is warm and focused on the table surface.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem001519.jpg | A close-up, high-angle shot focusing on a person's torso and arm. The individual wears a white jacket with a dark-colored garment underneath. A large, jagged, dark red bloodstain is visible on the chest area of the white jacket, with droplets of blood splattered across the fabric. The person's arm is partially visible, and the scene is set against a dark, indistinct background.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem000001.jpg | A wide shot of a spacious, high-ceilinged hall with light-colored stone walls. A row of black, wheeled display carts with glass cases sits on a red carpet, containing various artifacts including a sword and colorful gemstones. A small, black, winged creature with large yellow eyes and a grey belly stands on the red carpet in the foreground, facing the viewer. Two potted plants flank a set of double doors in the background.  Character continuity: same character. Background continuity: new view of the same background.

backup_sessao_2026-08-06/configs/nk2e_orig_samples14.txt

@@ -0,0 +1,14 @@
+/workspace/datasets/ctx_original/input_A/imagem000001.jpg | A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem000836.jpg | A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text "ALIENS! ABDUCT ME!!!" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem001768.jpg | A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000294.jpg | A tight close-up shot of a young person with dark hair, their face contorted in an intense expression of distress. Their mouth is wide open in a cry, and large, glistening tears stream from their wide, light-colored eyes. The character is positioned on the left side of the frame, with their head tilted slightly. A large, pale, hand-like appendage with multiple fingers extends from the right side of the frame, reaching toward the character's face. The background is a blurred, abstract blend of deep purple and bright magenta light, suggesting a high-energy or supernatural environment.  Character continuity: same character. Background continuity: new background.
+/workspace/datasets/ctx_heldout/input_A/imagem000075.jpg | A medium shot of a white and grey robotic figure with a spherical head and glowing cyan circular eyes, set against a cloudy grey sky. The robot holds a small, translucent blue crab with red-tipped claws on its palm. A seagull is captured in motion, flying directly in front of the robot's face, creating a slight motion blur. The robot has segmented arms and a smooth, rounded torso. The background is an overcast sky with soft, diffused light.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000759.jpg | A high-angle, close-up shot looking down into an open stone sarcophagus. A man with short, light-colored hair lies peacefully inside with his eyes closed, his hands folded over his chest. He wears a dark blue, high-collared military-style jacket with gold trim and a star-shaped pendant. The interior of the sarcophagus is lined with brown, padded fabric. On either side of the frame, the gloved hands of two individuals are visible, gripping the edges of the sarcophagus. Several small, light-colored flower petals are scattered across the man and the interior of the coffin.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000663.jpg | A close-up, slightly angled view of an old television screen displaying a live news broadcast of a night protest in a city street. The screen shows a large crowd of people gathered on a dark asphalt road at night, illuminated by streetlights and green traffic signals. A circular traffic sign with an upward arrow is visible on the right side of the screen. A news ticker at the bottom of the screen displays the text "BIOS AUMENTAN A VELOCIDAD ALARMANTE" and "AN MANIFESTACIONES CONTRA LEY MARCIAL". The television is viewed from the side, with the blurred silhouette of a person's head and shoulder visible in the foreground on the left.  Character continuity: same character. Background continuity: same background view.
+/workspace/datasets/ctx_heldout/input_A/imagem000611.jpg | A close-up, rear-view shot of a person wearing a dark, tactical-style hooded garment or backpack with a prominent circular vent or mechanical detail on the back. The person is positioned in a desert environment with orange-toned sand dunes visible in the background. The lighting is warm and dim, casting deep shadows across the scene. The person is oriented away from the camera, with a portion of a strap or equipment visible on the left side of the frame.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000374.jpg | A close-up shot focusing on a person's hand resting on a green felt gaming table. The person is wearing a white long-sleeved shirt with the cuff visible. Next to the hand, there are several tall stacks of circular casino chips in various colors, including purple, white, green, and orange. The lighting is warm and focused on the table surface.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem000273.jpg | A close-up, high-angle shot focusing on a person's torso and arm. The individual wears a white jacket with a dark-colored garment underneath. A large, jagged, dark red bloodstain is visible on the chest area of the white jacket, with droplets of blood splattered across the fabric. The person's arm is partially visible, and the scene is set against a dark, indistinct background.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_heldout/input_A/imagem001519.jpg | A wide shot of a spacious, high-ceilinged hall with light-colored stone walls. A row of black, wheeled display carts with glass cases sits on a red carpet, containing various artifacts including a sword and colorful gemstones. A small, black, winged creature with large yellow eyes and a grey belly stands on the red carpet in the foreground, facing the viewer. Two potted plants flank a set of double doors in the background.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image1.jpg | dynamic wide shot from a low angle, a young woman with long wavy blonde hair and amber eyes standing in a fighting stance, both hands gripping the hilt of a raised steel sword angled across her body, knees bent and one foot slid back, furrowed brows and fierce gaze locked on her opponent, wearing a cream off-shoulder blouse with a laced blue corset, facing a hulking muscular goblin with mottled green skin, pointed ears and bared tusks, its clawed fists raised mid-charge, dust kicked up between them, weathered stone wall and a wooden pillar with hanging ivy behind them, warm sunlight raking across the stone
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image2.jpg | medium close-up from a first-person viewpoint, a girl with short choppy blue hair and small braids looking up toward the camera with a soft concerned expression, lips slightly parted, wearing a teal short-sleeved shirt over a pink striped undershirt, seated cross-legged on a patched mattress, in the lower frame two forearms wrapped in stained white bandages reaching toward her, the bandaged hands open with palms up, a dim wooden attic room around them lit by warm orange light from a round window, small trinkets dangling from the beams
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image3.webp | wide shot from the side at eye level, a young woman with a curly pink bob, blue eyes, clear angular goggles pushed up on her face and ornate white and gold armor, on her tiptoes with both arms wrapped tightly around the neck of a towering broad-shouldered orc with weathered green skin and small tusks, her cheek pressed against his chest, eyes closed in a wide open-mouth smile, the orc's thick arms folded gently around her back, both standing on the rounded top of a dense white cloud, more clouds drifting below against a deep blue sky, bright sunlight from above

backup_sessao_2026-08-06/configs/nk2e_orig_samples3.txt

@@ -0,0 +1,3 @@
+/workspace/datasets/ctx_original/input_A/imagem000001.jpg | A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem000836.jpg | A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text "ALIENS! ABDUCT ME!!!" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem001768.jpg | A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.

backup_sessao_2026-08-06/configs/nk2e_orig_samples6.txt

@@ -0,0 +1,6 @@
+/workspace/datasets/ctx_original/input_A/imagem000001.jpg | A wide shot of a youth ice hockey face-off at center ice. A referee in a black and white striped jersey stands in the middle, holding a puck. To the left, a player in a blue jersey with a team logo and a blue helmet leans over their stick. To the right, a player in a mint green jersey with a team logo and a green helmet leans over their stick. Two other players in matching team jerseys are positioned slightly behind them. The background shows a crowded arena with spectators in the stands. The ice surface features a red center line and a red face-off circle.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem000836.jpg | A high-angle, wide shot of a young boy lying on his back in the middle of a large, circular design etched into the sand. The boy wears a green shirt, shorts, and a metallic colander on his head. The sand art features a large circle surrounding him with the text "ALIENS! ABDUCT ME!!!" written inside, accompanied by various geometric patterns, concentric circles, and arrows pointing toward the center. A small green box sits on the sand next to the boy. The surrounding area is vast, flat, and sandy.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/ctx_original/input_A/imagem001768.jpg | A medium shot inside a rustic tavern. A person with long, straight black hair wearing a purple hooded sweatshirt stands on the right, holding a black and gold rectangular object with both hands near their chest. Facing them on the left, a person with blonde hair tied in a low ponytail, wearing a white shirt and a dark green apron, sits at a long wooden table. On the table between them are several wooden bowls and a small goblet. The background features a stone wall, a wooden door, and a window with a grid pattern.  Character continuity: same character. Background continuity: new view of the same background.
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image1.jpg | dynamic wide shot from a low angle, a young woman with long wavy blonde hair and amber eyes standing in a fighting stance, both hands gripping the hilt of a raised steel sword angled across her body, knees bent and one foot slid back, furrowed brows and fierce gaze locked on her opponent, wearing a cream off-shoulder blouse with a laced blue corset, facing a hulking muscular goblin with mottled green skin, pointed ears and bared tusks, its clawed fists raised mid-charge, dust kicked up between them, weathered stone wall and a wooden pillar with hanging ivy behind them, warm sunlight raking across the stone
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image2.jpg | medium close-up from a first-person viewpoint, a girl with short choppy blue hair and small braids looking up toward the camera with a soft concerned expression, lips slightly parted, wearing a teal short-sleeved shirt over a pink striped undershirt, seated cross-legged on a patched mattress, in the lower frame two forearms wrapped in stained white bandages reaching toward her, the bandaged hands open with palms up, a dim wooden attic room around them lit by warm orange light from a round window, small trinkets dangling from the beams
+/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image3.webp | wide shot from the side at eye level, a young woman with a curly pink bob, blue eyes, clear angular goggles pushed up on her face and ornate white and gold armor, on her tiptoes with both arms wrapped tightly around the neck of a towering broad-shouldered orc with weathered green skin and small tusks, her cheek pressed against his chest, eyes closed in a wide open-mouth smile, the orc's thick arms folded gently around her back, both standing on the rounded top of a dense white cloud, more clouds drifting below against a deep blue sky, bright sunlight from above

backup_sessao_2026-08-06/configs/omini_sauce_1.toml

@@ -0,0 +1,55 @@
+# GROUNDEDSECRET — treino sério do krea2_omini_grounded (7000 steps).
+# Método validado no probe 250 + melhorias mapeadas em docs/OMINI_GROUNDED_SEGREDO.md:
+#   - caption_dropout 0.1 (treina o uncond grounded do CFG)
+#   - vl_grounding_jitter 384-768 (dial de inferência robusto)
+#   - multi-resolução 512/768/1024 (dataset é 1080p nativo)
+output_dir = '/workspace/checkpoints/omini_sauce_1'
+dataset = '/workspace/configs/omini_sauce_dataset.toml'
+epochs = 1000
+max_steps = 20000
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 100
+save_every_n_steps = 250
+checkpoint_every_n_minutes = 60
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1
+blocks_to_swap = 8
+
+[model]
+type = 'krea2_omini_grounded'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+flux_shift = true
+
+[krea2_omini_grounded]
+position_mode = 'width_shift'
+reference_position_offset = 1.0
+reference_timestep = 'zero'
+condition_dropout = 0.0
+condition_only_lora = true
+vl_prompt_style = 'plain'
+vl_longest_side = 768
+vl_grounding_jitter = [384, 768]
+caption_dropout = 0.1
+
+[adapter]
+type = 'lora'
+rank = 64
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/omini_sauce_dataset.toml

@@ -0,0 +1,26 @@
+resolutions = [1024]
+enable_ar_bucket = true
+min_ar = 0.5
+max_ar = 2.0
+num_ar_buckets = 9
+frame_buckets = [1]
+
+[[directory]]
+path = '/workspace/datasets/sauce9k/ds1_recortados/images_B'
+control_path = '/workspace/datasets/sauce9k/ds1_recortados/images_A'
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/sauce9k/ds2_poxima_v2/images_B'
+control_path = '/workspace/datasets/sauce9k/ds2_poxima_v2/images_A'
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/sauce9k/ds3_comikontext/images_B'
+control_path = '/workspace/datasets/sauce9k/ds3_comikontext/images_A'
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/sauce9k/ds4_contexto_curado/images_B'
+control_path = '/workspace/datasets/sauce9k/ds4_contexto_curado/images_A'
+num_repeats = 3

backup_sessao_2026-08-06/configs/proximacena_dataset.toml

@@ -0,0 +1,29 @@
+# Dataset do run k2_proximacena_v2 — 4 fontes, pareadas por stem.
+# Captions: descrição pura e detalhada da imagem B (vidcap preset grounded-scene).
+# ds4 = contexto_rush curado a mão pelo usuário -> ênfase 3x via num_repeats.
+resolutions = [1024]
+enable_ar_bucket = true
+min_ar = 0.5
+max_ar = 2.0
+num_ar_buckets = 9
+frame_buckets = [1]
+
+[[directory]]
+path = '/workspace/datasets/proxima_cena/ds1_recortados/images_B'
+control_path = '/workspace/datasets/proxima_cena/ds1_recortados/images_A'
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/proxima_cena/ds2_poxima_v2/images_B'
+control_path = '/workspace/datasets/proxima_cena/ds2_poxima_v2/images_A'
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/proxima_cena/ds3_comikontext/images_B'
+control_path = '/workspace/datasets/proxima_cena/ds3_comikontext/images_A'
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/proxima_cena/ds4_contexto_curado/images_B'
+control_path = '/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A'
+num_repeats = 3

backup_sessao_2026-08-06/configs/proximacena_samples.json

@@ -0,0 +1,26 @@
+{
+  "comment": "3 fixed sampling examples for the k2_proximacena_v2 run. Prompts written in the grounded-scene caption dialect (pure target description, framing first, no style words). Only 'reference', 'prompt', 'width' and 'height' are consumed by the supervisor.",
+  "samples": [
+    {
+      "name": "goblin_fight",
+      "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image1.jpg",
+      "prompt": "dynamic wide shot from a low angle, a young woman with long wavy blonde hair and amber eyes standing in a fighting stance, both hands gripping the hilt of a raised steel sword angled across her body, knees bent and one foot slid back, furrowed brows and fierce gaze locked on her opponent, wearing a cream off-shoulder blouse with a laced blue corset, facing a hulking muscular goblin with mottled green skin, pointed ears and bared tusks, its clawed fists raised mid-charge, dust kicked up between them, weathered stone wall and a wooden pillar with hanging ivy behind them, warm sunlight raking across the stone",
+      "width": 1152,
+      "height": 896
+    },
+    {
+      "name": "pov_bandaged_hands",
+      "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image2.jpg",
+      "prompt": "medium close-up from a first-person viewpoint, a girl with short choppy blue hair and small braids looking up toward the camera with a soft concerned expression, lips slightly parted, wearing a teal short-sleeved shirt over a pink striped undershirt, seated cross-legged on a patched mattress, in the lower frame two forearms wrapped in stained white bandages reaching toward her, the bandaged hands open with palms up, a dim wooden attic room around them lit by warm orange light from a round window, small trinkets dangling from the beams",
+      "width": 1152,
+      "height": 896
+    },
+    {
+      "name": "orc_hug_cloud",
+      "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image3.webp",
+      "prompt": "wide shot from the side at eye level, a young woman with a curly pink bob, blue eyes, clear angular goggles pushed up on her face and ornate white and gold armor, on her tiptoes with both arms wrapped tightly around the neck of a towering broad-shouldered orc with weathered green skin and small tusks, her cheek pressed against his chest, eyes closed in a wide open-mouth smile, the orc's thick arms folded gently around her back, both standing on the rounded top of a dense white cloud, more clouds drifting below against a deep blue sky, bright sunlight from above",
+      "width": 1152,
+      "height": 896
+    }
+  ]
+}

backup_sessao_2026-08-06/configs/proximacena_smoke_dataset.toml

@@ -0,0 +1,11 @@
+resolutions = [1024]
+enable_ar_bucket = true
+min_ar = 0.5
+max_ar = 2.0
+num_ar_buckets = 9
+frame_buckets = [1]
+
+[[directory]]
+path = '/workspace/datasets/proxima_cena_smoke/images_B'
+control_path = '/workspace/datasets/proxima_cena_smoke/images_A'
+num_repeats = 1

backup_sessao_2026-08-06/configs/saga512_dataset.toml

@@ -0,0 +1,36 @@
+# saga512 — dataset completo recaptionado (Gemini grounded-scene), corte p/ disco
+# autorizado: 10.991 pares (ds1 2172, ds2 5400 integral, ds3 2172, ds4 1247 s/ held-out).
+resolutions = [512]
+enable_ar_bucket = true
+min_ar = 0.5
+max_ar = 2.0
+num_ar_buckets = 7
+frame_buckets = [1]
+
+[[directory]]
+path = '/workspace/datasets/saga512/ds1_recortados/target'
+control_path = '/workspace/datasets/saga512/ds1_recortados/refs'
+multi_ref = true
+max_refs = 1
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/saga512/ds2_poxima_v2/target'
+control_path = '/workspace/datasets/saga512/ds2_poxima_v2/refs'
+multi_ref = true
+max_refs = 1
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/saga512/ds3_comikontext/target'
+control_path = '/workspace/datasets/saga512/ds3_comikontext/refs'
+multi_ref = true
+max_refs = 1
+num_repeats = 1
+
+[[directory]]
+path = '/workspace/datasets/saga512/ds4_contexto_curado/target'
+control_path = '/workspace/datasets/saga512/ds4_contexto_curado/refs'
+multi_ref = true
+max_refs = 1
+num_repeats = 3

backup_sessao_2026-08-06/configs/saga512_samples.json

@@ -0,0 +1,47 @@
+{
+ "comment": "6 samples fixos do run saga512: 3 pares do dataset (captions novas) + 3 selecionados pelo usuario",
+ "samples": [
+  {
+   "name": "ds4_imagem000001",
+   "reference": "/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000001.jpg",
+   "prompt": "eye-level medium shot at a professional hockey rink during a face-off, the referee in a black and white striped jersey stands in the center ready to drop the puck, flanked by two young players in blue and mint-green jerseys crouching with their hockey sticks on the ice, two smaller teammates wait in the background on either side of the referee, stadium seating filled with spectators is visible in the darkened background under bright spotlights, the foreground is framed by the blurred silhouettes of two other players facing the scene from the sides",
+   "width": 704,
+   "height": 384
+  },
+  {
+   "name": "ds4_imagem000836",
+   "reference": "/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem000836.jpg",
+   "prompt": "high-angle long shot looking down at a flat expanse of sand, a small child with a light tan complexion in a green t-shirt, dark shorts, and a white sun hat lying flat on their back in the center, arms and legs spread out, one hand holding a green snack, a small teal container placed near their legs, surrounding the child is a large, intricate geometric design drawn in the sand featuring lines, circles, and arrows, with the words aliens, and abduct me, etched in a circle, bright daylight illuminating the sandy landscape",
+   "width": 704,
+   "height": 384
+  },
+  {
+   "name": "ds4_imagem001768",
+   "reference": "/workspace/datasets/proxima_cena/ds4_contexto_curado/images_A/ds4_imagem001768.jpg",
+   "prompt": "medium shot at eye-level, a woman with blonde hair in a ponytail wearing a white shirt and green apron sits at a long wooden table, her body angled away from the camera, a person in a bright purple hooded sweatshirt stands opposite her, bowing deeply with their head down and long black hair covering their face, hands held together in a prayer-like gesture gripping a thin stick, a wooden table between them holds a stack of wooden bowls and a goblet, stone walls with a large wooden door and a window provide the background, rustic interior with wooden barrels on the floor",
+   "width": 704,
+   "height": 384
+  },
+  {
+   "name": "goblin_fight",
+   "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image1.jpg",
+   "prompt": "dynamic wide shot from a low angle, a young woman with long wavy blonde hair and amber eyes standing in a fighting stance, both hands gripping the hilt of a raised steel sword angled across her body, knees bent and one foot slid back, furrowed brows and fierce gaze locked on her opponent, wearing a cream off-shoulder blouse with a laced blue corset, facing a hulking muscular goblin with mottled green skin, pointed ears and bared tusks, its clawed fists raised mid-charge, dust kicked up between them, weathered stone wall and a wooden pillar with hanging ivy behind them, warm sunlight raking across the stone",
+   "width": 1152,
+   "height": 896
+  },
+  {
+   "name": "pov_bandaged_hands",
+   "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image2.jpg",
+   "prompt": "medium close-up from a first-person viewpoint, a girl with short choppy blue hair and small braids looking up toward the camera with a soft concerned expression, lips slightly parted, wearing a teal short-sleeved shirt over a pink striped undershirt, seated cross-legged on a patched mattress, in the lower frame two forearms wrapped in stained white bandages reaching toward her, the bandaged hands open with palms up, a dim wooden attic room around them lit by warm orange light from a round window, small trinkets dangling from the beams",
+   "width": 1152,
+   "height": 896
+  },
+  {
+   "name": "orc_hug_cloud",
+   "reference": "/workspace/datasets/raw/sampling_ref/imagens_geracao_samples/image3.webp",
+   "prompt": "wide shot from the side at eye level, a young woman with a curly pink bob, blue eyes, clear angular goggles pushed up on her face and ornate white and gold armor, on her tiptoes with both arms wrapped tightly around the neck of a towering broad-shouldered orc with weathered green skin and small tusks, her cheek pressed against his chest, eyes closed in a wide open-mouth smile, the orc's thick arms folded gently around her back, both standing on the rounded top of a dense white cloud, more clouds drifting below against a deep blue sky, bright sunlight from above",
+   "width": 1152,
+   "height": 896
+  }
+ ]
+}

backup_sessao_2026-08-06/configs/saga512_smoke.toml

@@ -0,0 +1,62 @@
+# KREA2 edit saga 2026-07-29 — treino principal.
+#
+# Base: receita validada krea2_multiref_grounded (docs/KREA2_MULTIREF_RECEITA.md),
+# aqui com max_refs=1 (todos os datasets são pares A->B).
+# - width_shift + refs a t=0 + LoRA condition-only routada + grounding Qwen3-VL
+# - caption_dropout=0.1 (uncond do CFG; condition_dropout é proibido no grounded)
+# - txtfusion_rank=128 (canal de leitura reforçado — tarefa é semântica: editar)
+# - rank 64 (usuário pediu >32; dezenas de milhares de imagens)
+
+output_dir = '/workspace/checkpoints/saga512_smoke'
+dataset = '/workspace/configs/saga512_dataset.toml'
+epochs = 1000
+max_steps = 10
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 100
+save_every_n_steps = 10
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO com multi_ref
+blocks_to_swap = 0           # usuário pediu velocidade; VRAM medida 20,1G com swap 8, cabe sem
+
+[model]
+type = 'krea2_multiref_grounded'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+flux_shift = true
+
+[krea2_multiref_grounded]
+max_refs = 1
+slot_axis = 'width'
+position_mode = 'width_shift'
+reference_position_offset = 1.0
+reference_timestep = 'zero'
+condition_dropout = 0.0      # proibido no caminho grounded
+condition_only_lora = true
+vl_prompt_style = 'picture_n'
+vl_image_label = 'image'
+vl_image_max_pixels = 147456 # 384² — grounding; cache é o gargalo, medir no smoke
+caption_dropout = 0.1
+txtfusion_rank = 128
+
+[adapter]
+type = 'lora'
+rank = 64
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 1e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/configs/saga512_train.toml

@@ -0,0 +1,62 @@
+# KREA2 edit saga 2026-07-29 — treino principal.
+#
+# Base: receita validada krea2_multiref_grounded (docs/KREA2_MULTIREF_RECEITA.md),
+# aqui com max_refs=1 (todos os datasets são pares A->B).
+# - width_shift + refs a t=0 + LoRA condition-only routada + grounding Qwen3-VL
+# - caption_dropout=0.1 (uncond do CFG; condition_dropout é proibido no grounded)
+# - txtfusion_rank=128 (canal de leitura reforçado — tarefa é semântica: editar)
+# - rank 64 (usuário pediu >32; dezenas de milhares de imagens)
+
+output_dir = '/workspace/checkpoints/saga512'
+dataset = '/workspace/configs/saga512_dataset.toml'
+epochs = 1000
+max_steps = 15000
+micro_batch_size_per_gpu = 1
+gradient_accumulation_steps = 1
+pipeline_stages = 1
+gradient_clipping = 1.0
+warmup_steps = 100
+save_every_n_steps = 250
+checkpoint_every_n_minutes = 120
+activation_checkpointing = true
+save_dtype = 'bfloat16'
+caching_batch_size = 1       # OBRIGATÓRIO com multi_ref
+blocks_to_swap = 0           # usuário pediu velocidade; VRAM medida 20,1G com swap 8, cabe sem
+
+[model]
+type = 'krea2_multiref_grounded'
+diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
+vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
+text_encoders = [
+    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
+]
+dtype = 'bfloat16'
+diffusion_model_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+flux_shift = true
+
+[krea2_multiref_grounded]
+max_refs = 1
+slot_axis = 'width'
+position_mode = 'width_shift'
+reference_position_offset = 1.0
+reference_timestep = 'zero'
+condition_dropout = 0.0      # proibido no caminho grounded
+condition_only_lora = true
+vl_prompt_style = 'picture_n'
+vl_image_label = 'image'
+vl_image_max_pixels = 147456 # 384² — grounding; cache é o gargalo, medir no smoke
+caption_dropout = 0.1
+txtfusion_rank = 128
+
+[adapter]
+type = 'lora'
+rank = 64
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'AdamW8bitKahan'
+lr = 2e-4
+betas = [0.9, 0.99]
+weight_decay = 0.01
+stabilize = false

backup_sessao_2026-08-06/nk2e_arquivos_novos.tgz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05213e2e79e311ac7c2523961d03e35ce71e3a782ced24aab78e4f0573a5ac06
+size 20953

backup_sessao_2026-08-06/nk2e_patches.diff

@@ -0,0 +1,395 @@
+diff --git a/kreaedit/data.py b/kreaedit/data.py
+index 778445b..c25b3bb 100644
+--- a/kreaedit/data.py
++++ b/kreaedit/data.py
+@@ -92,16 +92,22 @@ class EditTripletDataset(Dataset):
+ 
+     def __getitem__(self, i):
+         d = torch.load(self.files[i], map_location="cpu", weights_only=False)
+-        return {
++        item = {
+             "source_lat": d["source_lat"].float(),
+             "target_lat": d["target_lat"].float(),
+             "instruction": d["instruction"],
+         }
++        if "ref_vl" in d:
++            item["ref_vl"] = d["ref_vl"]  # uint8 (H, W, 3), for grounded encoding
++        return item
+ 
+ 
+ def collate(batch):
+-    return {
++    out = {
+         "source_lat": torch.stack([b["source_lat"] for b in batch]),
+         "target_lat": torch.stack([b["target_lat"] for b in batch]),
+         "instruction": [b["instruction"] for b in batch],
+     }
++    if "ref_vl" in batch[0]:
++        out["ref_vl"] = [b["ref_vl"] for b in batch]  # variable sizes -> list
++    return out
+diff --git a/kreaedit/incontext.py b/kreaedit/incontext.py
+index 40dcfc2..d94c840 100644
+--- a/kreaedit/incontext.py
++++ b/kreaedit/incontext.py
+@@ -69,7 +69,7 @@ def sample_logitnorm_t(b: int, device, mean: float, std: float) -> torch.Tensor:
+ 
+ 
+ def flow_loss_incontext(model, target_lat, ref_lat, txt, txtmask, patch,
+-                        t_mean=0.0, t_std=1.0):
++                        t_mean=0.0, t_std=1.0, ref_t_zero=False):
+     """One in-context flow-matching training step. Returns scalar MSE loss."""
+     b = target_lat.shape[0]
+     t = sample_logitnorm_t(b, target_lat.device, t_mean, t_std).to(target_lat.dtype)
+@@ -79,7 +79,9 @@ def flow_loss_incontext(model, target_lat, ref_lat, txt, txtmask, patch,
+     v_target = noise - target_lat
+ 
+     img, pos, mask, n_target = build_incontext_tokens(x_t, ref_lat, txtmask, patch)
+-    pred = model(img=img, context=txt, t=t, pos=pos, mask=mask)
++    n_ref = img.shape[1] - n_target
++    extra = {"ref_len": n_ref, "ref_timestep": 0.0} if ref_t_zero else {}
++    pred = model(img=img, context=txt, t=t, pos=pos, mask=mask, **extra)
+     pred_target = pred[:, :n_target]
+     v_tok, _ = patchify(v_target, patch)
+     return F.mse_loss(pred_target.float(), v_tok.float())
+@@ -112,7 +114,7 @@ def flow_loss_diptych(model, source_lat, target_lat, txt, txtmask, patch,
+ 
+ @torch.no_grad()
+ def edit_incontext(model, ref_lat, txt, txtmask, ts, patch, guidance=0.0,
+-                   untxt=None, unmask=None, seed=0):
++                   untxt=None, unmask=None, seed=0, ref_t_zero=False):
+     """Sample an edited target latent given a reference latent + instruction.
+ 
+     ts: list of timesteps from sampling.timesteps (1 -> 0). Returns (B,C,H,W).
+@@ -125,11 +127,12 @@ def edit_incontext(model, ref_lat, txt, txtmask, ts, patch, guidance=0.0,
+ 
+     for tcurr, tprev in zip(ts[:-1], ts[1:]):
+         img, pos, mask, n_target = build_incontext_tokens(x, ref_lat, txtmask, patch)
++        extra = {"ref_len": img.shape[1] - n_target, "ref_timestep": 0.0} if ref_t_zero else {}
+         t = torch.full((b,), tcurr, device=device, dtype=x.dtype)
+-        v = model(img=img, context=txt, t=t, pos=pos, mask=mask)[:, :n_target]
++        v = model(img=img, context=txt, t=t, pos=pos, mask=mask, **extra)[:, :n_target]
+         if cfg:
+             imgu, posu, masku, _ = build_incontext_tokens(x, ref_lat, unmask, patch)
+-            vu = model(img=imgu, context=untxt, t=t, pos=posu, mask=masku)[:, :n_target]
++            vu = model(img=imgu, context=untxt, t=t, pos=posu, mask=masku, **extra)[:, :n_target]
+             v = vu + guidance * (v - vu)
+         h_, w_ = h // patch, w // patch
+         v_lat = unpatchify(v, patch, h_, w_, c)
+diff --git a/scripts/edit_repl.py b/scripts/edit_repl.py
+index 52580d2..6cef39e 100644
+--- a/scripts/edit_repl.py
++++ b/scripts/edit_repl.py
+@@ -75,7 +75,12 @@ def detect_rank(lora_path):
+               help="turbo=OSS_TURBO (8 steps, cfg 0, mu 1.15); raw=OSS_RAW (28 steps + cfg)")
+ @click.option("--mu", default=None, type=float, help="timestep-shift; default 1.15(turbo)/auto(raw)")
+ @click.option("--manifest", default=None, type=click.Path(exists=True), help="read lines from a file instead of stdin")
+-def main(lora_path, out_dir, side, keep_ar, steps, cfg, rank, alpha, precision, seed, strength, checkpoint, mu, manifest):
++@click.option("--template", default=None, help="instruction template with {instruction}; must match training")
++@click.option("--width", default=None, type=int, help="force output width (with --height); overrides --side/--keep-ar")
++@click.option("--height", default=None, type=int, help="force output height (with --width)")
++@click.option("--grounded", is_flag=True, help="encode the instruction TOGETHER with the reference image (must match training)")
++@click.option("--ref-t-zero", is_flag=True, help="modulate reference rows at t=0 (must match training)")
++def main(lora_path, out_dir, side, keep_ar, steps, cfg, rank, alpha, precision, seed, strength, checkpoint, mu, manifest, template, width, height, grounded, ref_t_zero):
+     os.makedirs(out_dir, exist_ok=True)
+     device = "cuda"
+     patch = single_mmdit_large_wide.patch
+@@ -91,8 +96,13 @@ def main(lora_path, out_dir, side, keep_ar, steps, cfg, rank, alpha, precision,
+     print(f"[edit] LoRA rank={rank} alpha={alpha}", flush=True)
+ 
+     print("[edit] loading encoder + backbone (~2 min)…", flush=True)
+-    encoder = Qwen3VLConditioner(qwen3_vl_4b.model_id, qwen3_vl_4b.max_length,
+-                                 select_layers=qwen3_vl_4b.select_layers)
++    if grounded:
++        from kreaedit.grounding import GroundedQwen3VLConditioner
++        encoder = GroundedQwen3VLConditioner(qwen3_vl_4b.model_id, qwen3_vl_4b.max_length,
++                                             select_layers=qwen3_vl_4b.select_layers)
++    else:
++        encoder = Qwen3VLConditioner(qwen3_vl_4b.model_id, qwen3_vl_4b.max_length,
++                                     select_layers=qwen3_vl_4b.select_layers)
+     encoder = encoder.to(device, torch.bfloat16).eval().requires_grad_(False)
+ 
+     with torch.device("meta"):
+@@ -116,7 +126,11 @@ def main(lora_path, out_dir, side, keep_ar, steps, cfg, rank, alpha, precision,
+     @torch.no_grad()
+     def run_one(img_path, instruction, guidance, lora_strength):
+         lora.set_lora_strength(model, lora_strength)
+-        if keep_ar:
++        if width and height:
++            im = Image.open(img_path).convert("RGB").resize((width, height), Image.BICUBIC)
++            ref = torch.from_numpy(np.asarray(im).copy()).permute(2, 0, 1).float() / 255.0 * 2 - 1
++            tw, th = width, height
++        elif keep_ar:
+             ref, (tw, th) = load_image_ar(img_path, side)
+         else:
+             ref, (tw, th) = load_image(img_path, side), (side, side)
+@@ -124,15 +138,36 @@ def main(lora_path, out_dir, side, keep_ar, steps, cfg, rank, alpha, precision,
+         ref_lat = vae.encode(ref).to(torch.bfloat16)
+         n_tokens = (th // (8 * patch)) * (tw // (8 * patch))
+         ts = sampling.timesteps(n_tokens, steps, x1, x2, mu=mu)
+-        txt, txtmask = encoder([wrap_instruction(instruction)])
++        wrapped = (wrap_instruction(instruction, template) if template
++                   else wrap_instruction(instruction))
++        if grounded:
++            ref_pil = Image.open(img_path).convert("RGB")
++            txt, txtmask = encoder([wrapped], [ref_pil])
++        else:
++            txt, txtmask = encoder([wrapped])
++        # Drop padded text tokens (batch 1): text RoPE positions are all zero and
++        # masked tokens contribute nothing, so this is exact — and with an
++        # all-ones mask SDPA can take the flash path instead of materializing
++        # the full attention matrix (13.5 GiB at 1024px).
++        keep = txtmask[0].bool()
++        txt, txtmask = txt[:, keep], txtmask[:, keep]
+         untxt = unmask = None
+         if guidance > 0:
+-            untxt, unmask = encoder([""])
++            untxt, unmask = encoder([""], [ref_pil]) if grounded else encoder([""])
++            ukeep = unmask[0].bool()
++            untxt, unmask = untxt[:, ukeep], unmask[:, ukeep]
+             untxt = untxt.to(torch.bfloat16)
+-        out_lat = incontext.edit_incontext(model, ref_lat, txt.to(torch.bfloat16), txtmask,
+-                                           ts, patch, guidance=guidance, untxt=untxt,
+-                                           unmask=unmask, seed=seed)
+-        edited = to_pil(vae.decode(out_lat), tw, th)
++        # Offload the resident encoder during sampling: at 1024px the joint
++        # attention needs >13 GiB and does not fit alongside the 8 GiB encoder.
++        encoder.to("cpu")
++        torch.cuda.empty_cache()
++        try:
++            out_lat = incontext.edit_incontext(model, ref_lat, txt.to(torch.bfloat16), txtmask,
++                                               ts, patch, guidance=guidance, untxt=untxt,
++                                               unmask=unmask, seed=seed, ref_t_zero=ref_t_zero)
++            edited = to_pil(vae.decode(out_lat), tw, th)
++        finally:
++            encoder.to(device)
+         src_pil = to_pil(ref.to(device), tw, th)
+         stem = os.path.splitext(os.path.basename(img_path))[0]
+         base = os.path.join(out_dir, f"{stem}__{slug(instruction)}_cfg{guidance}_s{lora_strength}")
+diff --git a/scripts/pairs_to_cache.py b/scripts/pairs_to_cache.py
+index ec4bca0..3892f3e 100644
+--- a/scripts/pairs_to_cache.py
++++ b/scripts/pairs_to_cache.py
+@@ -77,8 +77,10 @@ def _iter_manifest(manifest):
+ @click.option("--filter", "do_filter", is_flag=True, help="drop pairs that changed too little / too much")
+ @click.option("--min-change", default=0.004, show_default=True, help="min mean pixel diff (0..1)")
+ @click.option("--max-change", default=0.6, show_default=True, help="max mean pixel diff (0..1)")
++@click.option("--with-vl", is_flag=True, help="also store the source image (aspect-preserved, <=147456 px, uint8) for grounded training")
++@click.option("--keep-ar", is_flag=True, help="preserve aspect ratio (area ~= size^2) instead of square squash")
+ def main(out_dir, source_dir, target_dir, instruction, instruction_from_txt, instruction_dir,
+-         manifest, size, do_filter, min_change, max_change):
++         manifest, size, do_filter, min_change, max_change, with_vl, keep_ar):
+     out_dir.mkdir(parents=True, exist_ok=True)
+     device = "cuda" if torch.cuda.is_available() else "cpu"
+     vae = VaeWrapper(QwenAutoencoder(), device=device, dtype=torch.bfloat16)
+@@ -103,8 +105,16 @@ def main(out_dir, source_dir, target_dir, instruction, instruction_from_txt, ins
+             dropped += 1
+             continue
+         try:
+-            s = load_image(src, size).unsqueeze(0)
+-            t = load_image(tgt, size).unsqueeze(0)
++            if keep_ar:
++                from kreaedit.data import load_image_ar
++                s = load_image_ar(src, size)[0].unsqueeze(0)
++                t = load_image_ar(tgt, size)[0].unsqueeze(0)
++                if s.shape != t.shape:  # pairs must share dims for the filter/concat
++                    dropped += 1
++                    continue
++            else:
++                s = load_image(src, size).unsqueeze(0)
++                t = load_image(tgt, size).unsqueeze(0)
+         except Exception:  # unreadable image
+             dropped += 1
+             continue
+@@ -115,7 +125,14 @@ def main(out_dir, source_dir, target_dir, instruction, instruction_from_txt, ins
+                 continue
+         s_lat = vae.encode(s)[0].to(torch.float16).cpu()
+         t_lat = vae.encode(t)[0].to(torch.float16).cpu()
+-        torch.save({"source_lat": s_lat, "target_lat": t_lat, "instruction": instr}, out_path)
++        item = {"source_lat": s_lat, "target_lat": t_lat, "instruction": instr}
++        if with_vl:
++            import numpy as np
++            from PIL import Image as PILImage
++            from kreaedit.grounding import fit_vl_image
++            im = fit_vl_image(PILImage.open(src).convert("RGB"))
++            item["ref_vl"] = torch.from_numpy(np.asarray(im).copy())
++        torch.save(item, out_path)
+         kept += 1
+ 
+     total = len(list(out_dir.glob("*.pt")))
+diff --git a/scripts/train.py b/scripts/train.py
+index a32a52d..6adab8b 100644
+--- a/scripts/train.py
++++ b/scripts/train.py
+@@ -31,7 +31,7 @@ from safetensors.torch import load_file
+ compat.relax_attention()
+ 
+ from kreaedit import incontext, lora
+-from kreaedit.data import EditTripletDataset, collate, wrap_instruction
++from kreaedit.data import DEFAULT_TEMPLATE, EditTripletDataset, collate, wrap_instruction
+ 
+ 
+ # Set by SIGINT/SIGTERM so we can save a clean resume point and stop at a step boundary.
+@@ -190,9 +190,18 @@ def main(config_path, resume, resume_state, fresh):
+     print(f"[train] LoRA on {n} linears; trainable params {trainable/1e6:.1f}M")
+ 
+     # --- resident text encoder (frozen) ---
+-    print("[train] loading Qwen3-VL text encoder")
+-    encoder = Qwen3VLConditioner(qwen3_vl_4b.model_id, qwen3_vl_4b.max_length,
+-                                 select_layers=qwen3_vl_4b.select_layers)
++    grounded = bool(cfg.get("grounded", False))
++    if grounded:
++        from kreaedit.grounding import GroundedQwen3VLConditioner
++        print("[train] loading Qwen3-VL GROUNDED conditioner (vision tower active)")
++        encoder = GroundedQwen3VLConditioner(
++            qwen3_vl_4b.model_id, qwen3_vl_4b.max_length,
++            select_layers=qwen3_vl_4b.select_layers,
++            max_pixels=int(cfg.get("vl_max_pixels", 147456)))
++    else:
++        print("[train] loading Qwen3-VL text encoder")
++        encoder = Qwen3VLConditioner(qwen3_vl_4b.model_id, qwen3_vl_4b.max_length,
++                                     select_layers=qwen3_vl_4b.select_layers)
+     encoder = encoder.to(device=device, dtype=torch.bfloat16).eval().requires_grad_(False)
+ 
+     # --- data ---
+@@ -215,7 +224,11 @@ def main(config_path, resume, resume_state, fresh):
+     step = 0
+     if state_path:
+         step = load_state(state_path, model, opt)
+-        print(f"[train] RESUMED full state from {state_path} at step {step}")
++        # load_state restores the saved optimizer param_groups (including lr),
++        # silently ignoring the config value — force the config lr back.
++        for g in opt.param_groups:
++            g["lr"] = cfg["lr"]
++        print(f"[train] RESUMED full state from {state_path} at step {step} (lr={cfg['lr']})")
+     else:
+         print("[train] starting fresh at step 0")
+ 
+@@ -229,19 +242,29 @@ def main(config_path, resume, resume_state, fresh):
+             # Instruction dropout -> teaches an unconditional (empty-instruction)
+             # prediction so inference-time CFG is in-distribution and clean.
+             drop = cfg.get("prompt_dropout", 0.0)
++            tmpl = cfg.get("instruction_template", DEFAULT_TEMPLATE)
+             with torch.no_grad():
+                 prompts = [
+                     "" if (drop > 0 and torch.rand(1).item() < drop)
+-                    else wrap_instruction(i)
++                    else wrap_instruction(i, tmpl)
+                     for i in batch["instruction"]
+                 ]
+-                txt, txtmask = encoder(prompts)
++                if grounded:
++                    # grounded contract: dropout empties the caption but the
++                    # reference image stays in BOTH branches (fork convention)
++                    from PIL import Image as _Image
++                    imgs = [_Image.fromarray(v.numpy()) for v in batch["ref_vl"]]
++                    txt, txtmask = encoder(prompts, imgs)
++                else:
++                    txt, txtmask = encoder(prompts)
+             txt = txt.to(torch.bfloat16)
+ 
++            ref_t_zero = bool(cfg.get("ref_timestep_zero", False))
+             if method == "diptych":
+                 loss = incontext.flow_loss_diptych(model, src, tgt, txt, txtmask, patch, t_mean, t_std)
+             else:
+-                loss = incontext.flow_loss_incontext(model, tgt, src, txt, txtmask, patch, t_mean, t_std)
++                loss = incontext.flow_loss_incontext(model, tgt, src, txt, txtmask, patch,
++                                                     t_mean, t_std, ref_t_zero=ref_t_zero)
+             (loss / accum).backward()
+ 
+             if (step + 1) % accum == 0:
+diff --git a/vendor/krea2/mmdit.py b/vendor/krea2/mmdit.py
+index 6ba9609..59375ec 100644
+--- a/vendor/krea2/mmdit.py
++++ b/vendor/krea2/mmdit.py
+@@ -309,6 +309,26 @@ class SingleStreamBlock(nn.Module):
+     def forward(
+         self, x: Tensor, vec: Tensor, freqs: Tensor, mask: Tensor | None = None
+     ) -> Tensor:
++        if isinstance(vec, tuple):
++            # Per-token timestep with two segments: rows [:split] modulated by
++            # vec_main (sampled t), rows [split:] (the clean reference) by
++            # vec_ref (t=0). Exactly equivalent to a per-token tvec that takes
++            # two distinct values; attention itself stays joint.
++            vec_main, vec_ref, split = vec
++            m = self.mod(vec_main)
++            r = self.mod(vec_ref)
++            h = self.prenorm(x)
++            pre = torch.cat([(1 + m[0]) * h[:, :split] + m[1],
++                             (1 + r[0]) * h[:, split:] + r[1]], dim=1)
++            a = self.attn(pre, freqs, mask)
++            x = x + torch.cat([m[2] * a[:, :split], r[2] * a[:, split:]], dim=1)
++            h2 = self.postnorm(x)
++            post = torch.cat([(1 + m[3]) * h2[:, :split] + m[4],
++                              (1 + r[3]) * h2[:, split:] + r[4]], dim=1)
++            mo = self.mlp(post)
++            x = x + torch.cat([m[5] * mo[:, :split], r[5] * mo[:, split:]], dim=1)
++            return x
++
+         prescale, preshift, pregate, postscale, postshift, postgate = self.mod(vec)
+         x = x + pregate * self.attn(
+             (1 + prescale) * self.prenorm(x) + preshift, freqs, mask
+@@ -383,10 +403,18 @@ class SingleStreamDiT(nn.Module):
+         t: Tensor,
+         pos: Tensor,
+         mask: Tensor | None = None,
++        ref_len: int = 0,
++        ref_timestep: float | None = None,
+     ) -> Tensor:
+         img = self.first(img)
+         t = self.tmlp(temb(t, self.config.tdim, device=img.device, dtype=img.dtype))
+         tvec = self.tproj(t)
++        tvec_ref = None
++        if ref_len > 0 and ref_timestep is not None:
++            t_ref = torch.full((img.shape[0],), float(ref_timestep),
++                               device=img.device, dtype=img.dtype)
++            t_ref = self.tmlp(temb(t_ref, self.config.tdim, device=img.device, dtype=img.dtype))
++            tvec_ref = self.tproj(t_ref)
+ 
+         txtmask = _mask(mask[:, : context.shape[1]])
+ 
+@@ -395,21 +423,34 @@ class SingleStreamDiT(nn.Module):
+ 
+         txtlen, imglen = context.shape[1], img.shape[1]
+         combined = torch.cat((context, img), dim=1)
+-
+-        # Pad combined sequence to a multiple of 256 to stabilize compiled kernel shapes.
+-        fulllen = combined.shape[1]
+-        _padlen = (-fulllen) % 256
+-        if _padlen > 0:
+-            combined = F.pad(combined, (0, 0, 0, _padlen))
+-            mask = F.pad(mask, (0, _padlen), value=False)
+-            pos = F.pad(pos, (0, 0, 0, _padlen))
+-
+-        mask = _mask(mask)
++        ref_split = combined.shape[1] - ref_len  # before any pad rows are appended
++
++        if bool(mask.all()):
++            # No padded tokens: skip the compile-shape padding (compile is
++            # disabled here) and hand SDPA a None mask so it can use the
++            # flash path instead of materializing the (H, L, L) matrix.
++            mask = None
++        else:
++            # Pad combined sequence to a multiple of 256 to stabilize compiled kernel shapes.
++            fulllen = combined.shape[1]
++            _padlen = (-fulllen) % 256
++            if _padlen > 0:
++                combined = F.pad(combined, (0, 0, 0, _padlen))
++                mask = F.pad(mask, (0, _padlen), value=False)
++                pos = F.pad(pos, (0, 0, 0, _padlen))
++
++            mask = _mask(mask)
+ 
+         freqs = self.posemb(pos)
+ 
++        blockvec = tvec
++        if tvec_ref is not None:
++            # reference rows sit at the tail of the sequence (before any pad,
++            # which only exists in the masked branch where pad rows are inert)
++            blockvec = (tvec, tvec_ref, ref_split)
++
+         for block in self.blocks:
+-            combined = block(combined, tvec, freqs, mask)
++            combined = block(combined, blockvec, freqs, mask)
+ 
+         final = self.last(combined, t)
+         output = final[:, txtlen : txtlen + imglen, :]

backup_sessao_2026-08-06/tools_konrad/comfy_sample.py

@@ -0,0 +1,89 @@
+"""Generate the 6 fixed samples through headless ComfyUI with the official
+comfyui-krea2edit nodes (v1.2.4+ geometry): turbo fp8, grounded 768, fit,
+ref_boost 4, 10 steps cfg 1, seed fixed.
+
+Usage: comfy_sample.py --lora konrad/atual.safetensors --out <dir> [--tag step2000]
+"""
+import argparse
+import json
+import shutil
+import time
+import urllib.request
+from pathlib import Path
+
+API = 'http://127.0.0.1:8288'
+SAMPLES = '/workspace/configs/saga512_samples.json'
+COMFY_IN = Path('/workspace/comfy/ComfyUI/input')
+COMFY_OUT = Path('/workspace/comfy/ComfyUI/output')
+
+
+def post(path, payload):
+    req = urllib.request.Request(API + path, json.dumps(payload).encode(),
+                                 {'Content-Type': 'application/json'})
+    return json.load(urllib.request.urlopen(req))
+
+
+def graph(ref_name, caption, w, h, lora, seed, prefix):
+    return {
+        "1": {"class_type": "UNETLoader", "inputs": {"unet_name": "krea2_raw_fp8_scaled.safetensors", "weight_dtype": "default"}},
+        "1t": {"class_type": "LoraLoaderModelOnly", "inputs": {"model": ["1", 0], "lora_name": "krea2_turbo_lora_rank_64_bf16.safetensors", "strength_model": 1.0}},
+        "2": {"class_type": "LoraLoaderModelOnly", "inputs": {"model": ["1t", 0], "lora_name": lora, "strength_model": 1.0}},
+        "3": {"class_type": "CLIPLoader", "inputs": {"clip_name": "qwen3vl_4b_bf16.safetensors", "type": "krea2", "device": "default"}},
+        "4": {"class_type": "VAELoader", "inputs": {"vae_name": "qwen_image_vae.safetensors"}},
+        "5": {"class_type": "LoadImage", "inputs": {"image": ref_name}},
+        "6": {"class_type": "VAEEncode", "inputs": {"pixels": ["5", 0], "vae": ["4", 0]}},
+        "7": {"class_type": "EmptySD3LatentImage", "inputs": {"width": w, "height": h, "batch_size": 1}},
+        "8": {"class_type": "Krea2EditGroundedEncode", "inputs": {"clip": ["3", 0], "prompt": caption, "image": ["5", 0], "grounding_px": 768, "system_prompt": ""}},
+        "9": {"class_type": "Krea2EditGroundedEncode", "inputs": {"clip": ["3", 0], "prompt": "", "image": ["5", 0], "grounding_px": 768, "system_prompt": ""}},
+        "10": {"class_type": "Krea2EditModelPatch", "inputs": {"model": ["2", 0], "source_latent": ["6", 0], "ref_boost": 4.0, "ref_boost_a": 1.0, "fit_mode": "fit", "vae": ["4", 0], "source_image": ["5", 0], "target_latent": ["7", 0]}},
+        "11": {"class_type": "KSampler", "inputs": {"model": ["10", 0], "positive": ["8", 0], "negative": ["9", 0], "latent_image": ["7", 0], "seed": seed, "steps": 10, "cfg": 1.0, "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}},
+        "12": {"class_type": "VAEDecode", "inputs": {"samples": ["11", 0], "vae": ["4", 0]}},
+        "13": {"class_type": "SaveImage", "inputs": {"images": ["12", 0], "filename_prefix": prefix}},
+    }
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument('--lora', required=True)
+    ap.add_argument('--out', required=True, type=Path)
+    ap.add_argument('--tag', default='sample')
+    ap.add_argument('--seed', type=int, default=76)
+    ap.add_argument('--ref-boost', type=float, default=4.0)
+    ap.add_argument('--samples', default=SAMPLES)
+    ap.add_argument('--area', type=int, default=0,
+                    help='rescale each sample W/H to this area (AR kept, /16); 0 = manifest dims')
+    args = ap.parse_args()
+    args.out.mkdir(parents=True, exist_ok=True)
+    samples = json.load(open(args.samples))['samples']
+    ok = 0
+    for s in samples:
+        ref = Path(s['reference'])
+        dest = COMFY_IN / ref.name
+        if not dest.exists():
+            shutil.copy2(ref, dest)
+        prefix = f"{args.tag}_{s.get('name', ref.stem)}"
+        w, h = s['width'], s['height']
+        if args.area:
+            scale = (args.area / (w * h)) ** 0.5
+            w = max(16, round(w * scale / 16) * 16)
+            h = max(16, round(h * scale / 16) * 16)
+        g = graph(ref.name, s['prompt'], w, h, args.lora, args.seed, prefix)
+        g["10"]["inputs"]["ref_boost"] = args.ref_boost
+        pid = post('/prompt', {'prompt': g})['prompt_id']
+        while True:
+            h = json.load(urllib.request.urlopen(f'{API}/history/{pid}'))
+            if pid in h and h[pid].get('status', {}).get('completed'):
+                break
+            if pid in h and h[pid].get('status', {}).get('status_str') == 'error':
+                print(f'[comfy_sample] ERRO em {prefix}:', json.dumps(h[pid]['status'])[:400])
+                break
+            time.sleep(2)
+        for f in sorted(COMFY_OUT.glob(f'{prefix}_*.png')):
+            shutil.move(str(f), args.out / f.name)
+            ok += 1
+        print(f'[comfy_sample] {prefix} done')
+    print(f'[comfy_sample] {ok} imagens -> {args.out}')
+
+
+if __name__ == '__main__':
+    main()

backup_sessao_2026-08-06/workflow_13k.json

@@ -0,0 +1,913 @@
+{
+  "id": "43b76607-370d-49b8-ac78-edf8c807a51c",
+  "revision": 0,
+  "last_node_id": 40,
+  "last_link_id": 69,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "UNETLoader",
+      "pos": [
+        100,
+        130
+      ],
+      "size": [
+        287.1021484375,
+        82
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "MODEL",
+          "type": "MODEL",
+          "links": [
+            25
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "UNETLoader"
+      },
+      "widgets_values": [
+        "krea2_raw_fp8_scaled.safetensors",
+        "default"
+      ]
+    },
+    {
+      "id": 2,
+      "type": "CLIPLoader",
+      "pos": [
+        100,
+        342
+      ],
+      "size": [
+        270,
+        106
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "CLIP",
+          "type": "CLIP",
+          "links": [
+            38
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "CLIPLoader"
+      },
+      "widgets_values": [
+        "qwen3vl_4b_bf16.safetensors",
+        "krea2",
+        "default"
+      ]
+    },
+    {
+      "id": 3,
+      "type": "VAELoader",
+      "pos": [
+        100,
+        578
+      ],
+      "size": [
+        270,
+        58
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "VAE",
+          "type": "VAE",
+          "links": [
+            23,
+            39
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "VAELoader"
+      },
+      "widgets_values": [
+        "qwen_image_vae.safetensors"
+      ]
+    },
+    {
+      "id": 12,
+      "type": "VAEDecode",
+      "pos": [
+        2268.833203125,
+        130
+      ],
+      "size": [
+        147.073828125,
+        46
+      ],
+      "flags": {},
+      "order": 13,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 51
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 23
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            24
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "VAEDecode"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 20,
+      "type": "K2TrainingBase",
+      "pos": [
+        428.8625507857853,
+        126.49836759784185
+      ],
+      "size": [
+        363.1177734375,
+        130
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 25
+        }
+      ],
+      "outputs": [
+        {
+          "name": "MODEL",
+          "type": "MODEL",
+          "links": [
+            31
+          ]
+        }
+      ],
+      "properties": {
+        "aux_id": "adbrasi/ctxrush-edit",
+        "ver": "a84d6ed8cbc7d8d8af0042b4548f5cf450adca65",
+        "Node name for S&R": "K2TrainingBase"
+      },
+      "widgets_values": [
+        true,
+        "krea2_turbo_lora_rank_64_bf16.safetensors",
+        1,
+        true
+      ]
+    },
+    {
+      "id": 31,
+      "type": "KSamplerSelect",
+      "pos": [
+        2024.4547149072016,
+        -357.1668585163058
+      ],
+      "size": [
+        270,
+        58
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SAMPLER",
+          "type": "SAMPLER",
+          "links": [
+            46
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "KSamplerSelect"
+      },
+      "widgets_values": [
+        "euler"
+      ]
+    },
+    {
+      "id": 32,
+      "type": "BasicScheduler",
+      "pos": [
+        2025.918814907203,
+        -238.5747585163061
+      ],
+      "size": [
+        270,
+        106
+      ],
+      "flags": {},
+      "order": 11,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 48
+        },
+        {
+          "name": "steps",
+          "type": "INT",
+          "widget": {
+            "name": "steps"
+          },
+          "link": 49
+        }
+      ],
+      "outputs": [
+        {
+          "name": "SIGMAS",
+          "type": "SIGMAS",
+          "links": [
+            47
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "BasicScheduler"
+      },
+      "widgets_values": [
+        "simple",
+        20,
+        1
+      ]
+    },
+    {
+      "id": 29,
+      "type": "SamplerCustomAdvanced",
+      "pos": [
+        2406.852931313454,
+        -605.3318085163054
+      ],
+      "size": [
+        412.01386718750007,
+        550.9390000000001
+      ],
+      "flags": {},
+      "order": 12,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "noise",
+          "type": "NOISE",
+          "link": 45
+        },
+        {
+          "name": "guider",
+          "type": "GUIDER",
+          "link": 44
+        },
+        {
+          "name": "sampler",
+          "type": "SAMPLER",
+          "link": 46
+        },
+        {
+          "name": "sigmas",
+          "type": "SIGMAS",
+          "link": 47
+        },
+        {
+          "name": "latent_image",
+          "type": "LATENT",
+          "link": 50
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output",
+          "type": "LATENT",
+          "links": [
+            51
+          ]
+        },
+        {
+          "name": "denoised_output",
+          "type": "LATENT",
+          "links": null
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "SamplerCustomAdvanced"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 28,
+      "type": "K2ReferenceGuider",
+      "pos": [
+        2002.9657430322022,
+        -613.4074585163063
+      ],
+      "size": [
+        333.55234375,
+        170
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 43
+        },
+        {
+          "name": "reference_conditioning",
+          "type": "K2_REF_CONDITIONING",
+          "link": 42
+        },
+        {
+          "name": "reference_mask",
+          "shape": 7,
+          "type": "K2_REF_MASK",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "GUIDER",
+          "type": "GUIDER",
+          "links": [
+            44
+          ]
+        }
+      ],
+      "properties": {
+        "aux_id": "adbrasi/ctxrush-edit",
+        "ver": "a84d6ed8cbc7d8d8af0042b4548f5cf450adca65",
+        "Node name for S&R": "K2ReferenceGuider"
+      },
+      "widgets_values": [
+        1,
+        1,
+        "full_off",
+        false
+      ]
+    },
+    {
+      "id": 30,
+      "type": "RandomNoise",
+      "pos": [
+        1993.120312907202,
+        -908.7572165163064
+      ],
+      "size": [
+        270,
+        82
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "NOISE",
+          "type": "NOISE",
+          "links": [
+            45
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "RandomNoise"
+      },
+      "widgets_values": [
+        76524150451214,
+        "randomize"
+      ]
+    },
+    {
+      "id": 13,
+      "type": "SaveImage",
+      "pos": [
+        2865.2437165706397,
+        -566.3339474665232
+      ],
+      "size": [
+        638.4859374999996,
+        470.1999999999998
+      ],
+      "flags": {},
+      "order": 14,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 24
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": null
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "SaveImage"
+      },
+      "widgets_values": [
+        "asdasdasd"
+      ]
+    },
+    {
+      "id": 26,
+      "type": "ResolutionSelector",
+      "pos": [
+        3255.244359849422,
+        -779.8743729490525
+      ],
+      "size": [
+        270,
+        126
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "width",
+          "type": "INT",
+          "links": [
+            55
+          ]
+        },
+        {
+          "name": "height",
+          "type": "INT",
+          "links": [
+            56
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.29.0",
+        "Node name for S&R": "ResolutionSelector"
+      },
+      "widgets_values": [
+        "16:9 (Widescreen)",
+        0.7,
+        16
+      ]
+    },
+    {
+      "id": 21,
+      "type": "K2LoadReferencePIL",
+      "pos": [
+        2857.767272385095,
+        -950.0849861754249
+      ],
+      "size": [
+        365.7154296875,
+        318
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            54
+          ]
+        }
+      ],
+      "properties": {
+        "aux_id": "adbrasi/ctxrush-edit",
+        "ver": "a84d6ed8cbc7d8d8af0042b4548f5cf450adca65",
+        "Node name for S&R": "K2LoadReferencePIL"
+      },
+      "widgets_values": [
+        "image1.jpg",
+        "",
+        "image"
+      ]
+    },
+    {
+      "id": 27,
+      "type": "CtxRushKrea2MultiRefControlledApply",
+      "pos": [
+        1541.6502914526575,
+        -843.364572152671
+      ],
+      "size": [
+        400,
+        686
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 69
+        },
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "link": 38
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 39
+        },
+        {
+          "name": "image_1",
+          "type": "IMAGE",
+          "link": 54
+        },
+        {
+          "name": "image_2",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "name": "image_3",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "name": "image_4",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "name": "width",
+          "type": "INT",
+          "widget": {
+            "name": "width"
+          },
+          "link": 55
+        },
+        {
+          "name": "height",
+          "type": "INT",
+          "widget": {
+            "name": "height"
+          },
+          "link": 56
+        }
+      ],
+      "outputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "links": [
+            43,
+            48
+          ]
+        },
+        {
+          "name": "reference_conditioning",
+          "type": "K2_REF_CONDITIONING",
+          "links": [
+            42
+          ]
+        },
+        {
+          "name": "positive_with_reference",
+          "type": "CONDITIONING",
+          "links": null
+        },
+        {
+          "name": "negative_with_reference",
+          "type": "CONDITIONING",
+          "links": null
+        },
+        {
+          "name": "positive_no_reference",
+          "type": "CONDITIONING",
+          "links": null
+        },
+        {
+          "name": "negative_no_reference",
+          "type": "CONDITIONING",
+          "links": null
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            50
+          ]
+        },
+        {
+          "name": "steps",
+          "type": "INT",
+          "links": [
+            49
+          ]
+        },
+        {
+          "name": "cfg",
+          "type": "FLOAT",
+          "links": null
+        }
+      ],
+      "properties": {
+        "aux_id": "adbrasi/ctxrush-edit",
+        "ver": "a84d6ed8cbc7d8d8af0042b4548f5cf450adca65",
+        "Node name for S&R": "CtxRushKrea2MultiRefControlledApply"
+      },
+      "widgets_values": [
+        "extreme close-up on the lower face of a new person, tongue extended outward from parted lips, thick yellow liquid coating the mouth and dripping down the chin, neutral flat gray background",
+        "",
+        "krea2_multiref_grounded_13250.safetensors",
+        1,
+        1,
+        "turbo",
+        768,
+        640,
+        1,
+        0,
+        "auto",
+        "auto",
+        "all",
+        "flat",
+        "runner_pil",
+        true,
+        true,
+        false
+      ]
+    },
+    {
+      "id": 25,
+      "type": "PathchSageAttentionKJ",
+      "pos": [
+        928.0390858213289,
+        108.19882801359584
+      ],
+      "size": [
+        270,
+        82
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 31
+        }
+      ],
+      "outputs": [
+        {
+          "name": "MODEL",
+          "type": "MODEL",
+          "links": [
+            69
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "4d46ac107c33ed8a3d181b8776ede66498583380",
+        "Node name for S&R": "PathchSageAttentionKJ"
+      },
+      "widgets_values": [
+        "auto",
+        false
+      ]
+    }
+  ],
+  "links": [
+    [
+      23,
+      3,
+      0,
+      12,
+      1,
+      "VAE"
+    ],
+    [
+      24,
+      12,
+      0,
+      13,
+      0,
+      "IMAGE"
+    ],
+    [
+      25,
+      1,
+      0,
+      20,
+      0,
+      "MODEL"
+    ],
+    [
+      31,
+      20,
+      0,
+      25,
+      0,
+      "MODEL"
+    ],
+    [
+      38,
+      2,
+      0,
+      27,
+      1,
+      "CLIP"
+    ],
+    [
+      39,
+      3,
+      0,
+      27,
+      2,
+      "VAE"
+    ],
+    [
+      42,
+      27,
+      1,
+      28,
+      1,
+      "K2_REF_CONDITIONING"
+    ],
+    [
+      43,
+      27,
+      0,
+      28,
+      0,
+      "MODEL"
+    ],
+    [
+      44,
+      28,
+      0,
+      29,
+      1,
+      "GUIDER"
+    ],
+    [
+      45,
+      30,
+      0,
+      29,
+      0,
+      "NOISE"
+    ],
+    [
+      46,
+      31,
+      0,
+      29,
+      2,
+      "SAMPLER"
+    ],
+    [
+      47,
+      32,
+      0,
+      29,
+      3,
+      "SIGMAS"
+    ],
+    [
+      48,
+      27,
+      0,
+      32,
+      0,
+      "MODEL"
+    ],
+    [
+      49,
+      27,
+      7,
+      32,
+      1,
+      "INT"
+    ],
+    [
+      50,
+      27,
+      6,
+      29,
+      4,
+      "LATENT"
+    ],
+    [
+      51,
+      29,
+      0,
+      12,
+      0,
+      "LATENT"
+    ],
+    [
+      54,
+      21,
+      0,
+      27,
+      3,
+      "IMAGE"
+    ],
+    [
+      55,
+      26,
+      0,
+      27,
+      7,
+      "INT"
+    ],
+    [
+      56,
+      26,
+      1,
+      27,
+      8,
+      "INT"
+    ],
+    [
+      69,
+      25,
+      0,
+      27,
+      0,
+      "MODEL"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.38554328942953314,
+      "offset": [
+        263.56295908444093,
+        1200.5992511100708
+      ]
+    },
+    "frontendVersion": "1.47.12",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true
+  },
+  "version": 0.4
+}