jzhang-POP/d4966-texfix-lora

D-4966 texture-fix LoRA (private)

Edit LoRAs for Qwen-Image-2.1 that restore the real D-4966 upholstery (dense short-pile ochre plush) on a generated sofa image, leaving shape, seams, tufting, lighting and the room unchanged. They are pass 2 of the D-4966 pipeline: clay render -> Qwen pass 1 -> upscale to 3840 -> tiled, masked fix with this LoRA.

Licence. v1 to v4 are LoRAs for Qwen-Image-2.1 (Qwen Research License: R&D and evaluation only): not for client deliverables without a commercial licence from Qwen. e2511/ is the licence-clean version on Qwen-Image-Edit-2511 (Apache-2.0), trained without any Qwen-2.1 output (legal sign-off still pending).

Versions

Folder	Rank	Steps	Training data	Recommended checkpoint
v1/	32	3,000	base-model regenerations + blur/JPEG/colour drift	d4966_texfix_v1.safetensors (3000)
v2/	64	4,000	+ no colour drift, identity pairs, joint relighting	d4966_texfix_v2.safetensors (4000)
v3/	64	4,000	+ wrong-fabric and CG-smooth inputs	d4966_texfix_v3_000002500.safetensors
v4/	64	4,000	+ Qwen's own plush ("self" pairs)	d4966_texfix_v4_000002500.safetensors

Every saved checkpoint is included (every 250 or 500 steps); the file without a step number is the final step. Do not use v4 final (4000): on unseen views it redraws the tufting.

Use v4 step 2500 (v4/d4966_texfix_v4_000002500.safetensors).

Held-out score (share of the texture gap between pass 1 and the real photo that the LoRA closes, 40 crops, higher is better): v1 36%, v2 36%, v3 75%, v4 step 2500 78%. It also keeps the geometry (tufts and seams in place) and, on real pass-1 images from camera views never used in training, keeps colour and shading better than v3 (v3 adds a lighter, golden, patchy cast).

e2511: Qwen-Image-Edit-2511 (Apache-2.0)

Folder	Base	Rank	Steps	Recommended checkpoint
e2511/	Qwen/Qwen-Image-Edit-2511	64	3,000	d4966_texfix_e2511_000002500.safetensors

Trained with ai-toolkit (arch: qwen_image_edit_plus) on photo-only pairs (CG-smooth, blur/JPEG, identity) plus wrong-fabric and own-plush inputs generated by 2511 itself (with the Apache-2.0 Lightning LoRA), so no Qwen-2.1 output is in its training data. Same prompt as v2 to v4. Held-out score: 75% of the texture gap closed (v3 75%, v4 78%); 2511 without the LoRA closes 43% and turns the fabric olive. Colour error dE 1.8 to 5.5, geometry kept (0.94 to 1.00). Scores and sheets: e2511/eval/. Evaluated at true CFG 3, 30 steps, full strength, masked; ~1 MP input. Files are 1.18 GB (rank 64 on the 20B model). Not yet tested on real pass-1 frames.

Prompt (one fixed instruction, must match training)

• v1: d4966: restore the sofa upholstery to the real fabric, a dense short-pile ochre plush with a soft, slightly curly nap and a gentle sheen. Keep the shape, seams, tufting, lighting, colours and everything outside the upholstery unchanged.
• v2, v3, v4: same without slightly curly (it pulled the model toward bouclé): d4966: restore the sofa upholstery to the real fabric, a dense short-pile ochre plush with a soft nap and a gentle sheen. Keep the shape, seams, tufting, lighting, colours and everything outside the upholstery unchanged.

Settings used

Full strength (regenerate from noise), true CFG 3 with negative prompt " ", 30 steps, ~1 MP tiles (1024 px, 256 px overlap) at the final resolution, masked denoising (latents outside the dilated sofa mask pinned to the input), composite inside the mask, decode with madebyollin/texture-fix-vae-for-qwen-image-2.1. Adaptive mode skips tiles whose fabric is already sharp (detail ratio >= 0.48).

ComfyUI example workflows (comfyui/)

Drag a .json into ComfyUI (needs a recent ComfyUI with the Qwen-Image-2.1 nodes, Oct 2026). The LoRA files load as they are with LoraLoaderModelOnly. *_api.json = the same graphs in API format; build_*.py regenerate them.

Workflow	Model	Input
d4966_texfix_qwen21.json	Qwen-Image-2.1 base + v4, 30 steps, CFG 3	one ~1 MP tile + sofa mask
d4966_texfix_qwen21_turbo.json	Qwen-Image-2.1 Turbo + v4, 8 steps, CFG 1 (LoRA trained on base: untested on Turbo)	one tile
d4966_texfix_qwen21_tiled_turbo.json	same as Turbo, full frame: 3840x2176, 15 tiles, each tile its own edit reference	frame + sofa mask
d4966_texfix_e2511_tile.json	Qwen-Image-Edit-2511 + e2511, 30 steps, CFG 3	one tile
d4966_texfix_e2511_tiled.json	same, full frame, 15 tiles	frame + sofa mask

All of them: the training prompt, masked denoising inside the grown sofa mask, paste back inside the mask; the 2.1 workflows decode with the texture-fix VAE. Example inputs in comfyui/example_inputs/ (put them in ComfyUI/input/d4966/): a pass-1 tile and full frame with masks, and a held-out crop with its real photo for comparison. Built and node-checked against ComfyUI 0.39; not yet run end to end.

Loading in diffusers

Trained with ai-toolkit (arch: qwen_image_2), which saves the image MLP as one fused img_mlp.gate_up LoRA. Diffusers has two separate Linears and silently skips the fused keys, so split them first (both halves share lora_A; lora_B is cut in half along its rows into img_mlp.gate_layer and img_mlp.proj). The project's 04_infer.py has split_fused_gate_up() and check_lora_coverage() for this.

ComfyUI loads the files as they are with the stock LoraLoaderModelOnly node: all 192 layers the LoRA targets match the key names of Comfy-Org/Qwen-Image-2.1 (qwen_image_2.1_bf16.safetensors), fused gate_up included. Use ~1 MP tiles (1024 px), the exact training prompt, CFG 3, 30 steps, denoise 1.0, a grown sofa mask as latent noise mask, and the texture-fix VAE (texture_fix_vae_for_qwen_image_2.1_bf16.safetensors) for decoding.