Instructions to use Mike0021/qwen-image-edit-2511-pbr-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use Mike0021/qwen-image-edit-2511-pbr-lora with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline from diffusers.utils import load_image # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2511", dtype=torch.bfloat16, device_map="cuda") pipe.load_lora_weights("Mike0021/qwen-image-edit-2511-pbr-lora") prompt = "Turn this cat into a dog" input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png") image = pipe(image=input_image, prompt=prompt).images[0] - Inference
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
PBR maps LoRA for Qwen-Image-Edit-2511
Turns a lit photo of an object into one of its PBR maps, one instruction each:
Turn this photo into its albedo map: the base colour of every surface, with all lighting, shadows and reflections removed.Turn this photo into its roughness map: grayscale, black where the surface is glossy, white where it is rough.Turn this photo into its metallic map: white where the surface is bare metal, black everywhere else.
Made for the PBR option of the projection-texturing Space: each generated view of a model is decomposed, and the
maps are projected and baked like the colour. Best with the Lightning 4-step LoRA
(lightx2v/Qwen-Image-Edit-2511-Lightning), at 768 px.
Training
- Data: 2,820 synthetic views rendered in Blender Cycles: primitives, assemblies and 74 untextured meshes of
everyday objects, with MatSynth materials (
gvecchio/MatSynth, CC0 / CC-BY; metals three times as likely) and plain PBR materials (paint, plastic, ceramic, bare metal), under Poly Haven (CC0) and Blender studio HDRIs. For each view the exact base colour, roughness and metallic of every pixel are rendered with the same camera. - v1 (
pbr/step-4230): DiffSynth-Studio, rank 32, learning rate 1e-4, 2 epochs over 8,460 pairs (4x H200, 2 h). - v2 (
pbr2/step-4215, recommended): 2,800 more views in which a quarter of the materials are glossy paint, prints and lacquer that are never metal; 1 epoch over all 16,860 pairs. It stops reading glossy paint as metal.
Benchmark
42 HDRI-lit views of 14 Khronos glTF sample assets (not in the training data) with exact ground truth:
| Method | albedo si-RMSE | roughness MAE | metallic MAE | metallic acc. |
|---|---|---|---|---|
| naive (photo, roughness 0.5, no metal) | 0.171 | 0.257 | 0.283 | 0.72 |
| Marigold-IID appearance v1-1 | 0.133 | 0.192 | 0.444 | 0.50 |
| RGB<->X | 0.143 | 0.247 | 0.442 | 0.56 |
| v1 + Lightning, 4 steps (DiffSynth) | 0.108 | 0.191 | 0.251 | 0.75 |
| v1 + Lightning, 4 steps (diffusers) | 0.114 | 0.196 | 0.268 | 0.72 |
| v2 + Lightning, 4 steps (diffusers) | 0.108 | 0.218 | 0.255 | 0.74 |
With diffusers, load the file's keys under transformer. without .default (DiffSynth saves PEFT keys of the
bare transformer): {"transformer." + k.replace(".default", ""): v}.
- Downloads last month
- -
Model tree for Mike0021/qwen-image-edit-2511-pbr-lora
Base model
Qwen/Qwen-Image-Edit-2511