A FLUX-Fill based generative framework that synthesizes four coherent fashion-model views from dual-view garment references in a single inpainting forward pass.
DualViewFashion formulates multi-view fashion model generation as a structured inpainting problem. Given paired garment views, typically the front and back of a clothing item, we construct a canonical 7-region canvas and use FLUX Fill to complete the masked model-view regions. The system generates a synchronized set of model images, including reference-pose, front, side, and back views, without editing an existing person image.
This is not a virtual try-on or person-inpainting pipeline. DualViewFashion creates a new fashion model presentation conditioned by garment references.
-
Dual-View Garment Conditioning. Instead of relying on a single product image, DualViewFashion conditions generation on paired garment views. The front view provides appearance cues such as color, print, neckline, and silhouette; the back view preserves reverse-side structure and rear-view design details.
-
Single-Pass Four-View Synthesis. We cast multi-view generation into a single FLUX-Fill inpainting canvas. One forward pass produces four model views: reference-pose, front, side, and back. This avoids independently sampling each viewpoint and reduces cross-view drift.
-
Layout-Constrained Inpainting. The model receives a 7-region layout: two garment reference slots, one identity placeholder slot, and four masked model-view slots. The inpainting objective encourages the model to respect the garment evidence while completing all target views jointly.
-
Identity Placeholder for Future Control. The gray slot in the reference row is intentionally reserved. In the current release it is filled with a neutral gray placeholder and annotated in the examples; future versions will replace this slot with a model identity reference, enabling generation of a fixed model identity wearing the target garment.
-
Image-First, Video-Ready. The core contribution is multi-view image generation. We also provide an optional Wan2.2 multi-frame reference video inference script, where multiple reference images are injected as sparse temporal conditions.
- Image inference code release
- Wan2.2 multi-frame reference video inference code release
- Example gallery release
- Low-resolution full-model checkpoint release (512 x 384)
- High-resolution full-model checkpoint ready (1024 x 768)
- Image training code release
- Wan2.2 video checkpoint release
- Wan2.2 video training code release
- Dataset release
- Model-identity conditioned generation
The examples below are selected from the released full-parameter image checkpoints. The output canvas contains the dual-view garment references in the first row and four generated model views in the second row. The top-right gray region is labeled Future Model ID to match the reserved model-identity placeholder style.
| Example | Generated multi-view canvas |
|---|---|
| Example 1 (512 full model) |
 |
| Example 2 (512 full model) |
 |
| Example 3 (1024 full model, step 1000, actively improving) |
 |
The optional video extension uses multiple reference frames as sparse temporal conditions. The example below shows three reference frames placed at relative positions 0.25, 0.5, and 0.75.
| Reference Frames | Generated Video |
|---|---|
   |
 Download MP4 |
The current video demo is a 10-second, 161-frame generation. We plan to release a 10-second, 241-frame version in a future checkpoint release. The README uses a compact animated preview for reliable inline playback; the original MP4 is kept as a downloadable asset.
conda create -n search-train python=3.10 -y
conda activate search-train
pip install -r requirements.txtDualViewFashion image inference was tested with diffusers 0.39.0.dev0 and a FluxFillPipeline-compatible FLUX Fill implementation.
python inference_full_caption.py \
--garment_front path/to/garment_front.jpg \
--garment_back path/to/garment_back.jpg \
--model_path path/to/FLUX.1-Fill-dev \
--transformer_path path/to/trained-full-multiview-caption-stage1 \
--output outputs/dualviewfashion_grid.png \
--cloth_type "dress" \
--gender "female" \
--cloth_size 512 \
--steps 50 \
--guidance_scale 30.0 \
--seed 42The default command uses the 512 x 384 full-model checkpoint. For the 1024 x 768 full-model checkpoint, use --transformer_path path/to/trained-full-multiview-caption-stage2/checkpoint-1000 and --cloth_size 1024. The 1024 version is initialized from the 512 version and trained for another 1000 steps.
The generated image is a single 7-region canvas. The bottom row contains four synchronized model renderings produced by one inpainting pass.
python video_inference.py \
--model_dir path/to/Wan2.2-I2V-A14B \
--ref_images ref_000.png ref_025.png ref_050.png ref_075.png \
--frame_positions 0 0.25 0.5 0.75 \
--high_lora path/to/dualviewfashion-wan22-high-noise.safetensors \
--low_lora path/to/dualviewfashion-wan22-low-noise.safetensors \
--prompt "a fashion model walks in, poses, turns around, and walks out, full body, studio lighting" \
--output outputs/dualviewfashion_video.mp4The video LoRA checkpoints are available at ShineChen1024/DualViewFashion. Current video inference examples use 161 frames for 10-second generation; a 241-frame, 10-second version is planned.
Checkpoints are hosted on Hugging Face: ShineChen1024/DualViewFashion
| Model | Path | Status |
|---|---|---|
| DualViewFashion full model (512 x 384) | trained-full-multiview-caption-stage1/transformer |
Available |
| DualViewFashion full model (1024 x 768) | trained-full-multiview-caption-stage2/checkpoint-1000/transformer |
Available |
| DualViewFashion Wan2.2 video LoRA | - | Available |
| DualViewFashion dataset | - | Coming soon |
+------------------------------------------------+
| DualViewFashion |
| FLUX-Fill Layout Inpainting |
| |
Garment Front -->| [ Garment Front ] [ Garment Back ] [ ID Slot ]|
Garment Back -->| |
Identity Slot -->| [ Ref-Pose ] [ Front ] [ Side ] [ Back Model ]|--> Four model views
Masked Targets ->| |
+------------------------------------------------+
Key difference from virtual try-on pipelines:
| Virtual Try-On / Person Inpainting | DualViewFashion | |
|---|---|---|
| Input person | Required | Not required |
| Target task | Edit clothing on an existing person | Generate new fashion model views |
| Garment evidence | Often single-view | Dual-view garment references |
| View generation | Usually one edited image | Four coordinated model views |
| Generation process | Person-region inpainting | Layout-conditioned multi-view inpainting |
- Canvas Construction. The front and back garment images are placed in the first two reference slots. A neutral gray identity placeholder is placed in the third reference slot.
- Mask Definition. The complete bottom row is masked as the target region for generation.
- FLUX-Fill Inpainting. A LoRA-adapted FLUX Fill model completes the masked region in one pass, producing four synchronized fashion-model views.
- Future Identity Control. The gray placeholder will be replaced by a model identity reference, allowing the same model identity to be rendered wearing different garments.
DualViewFashion adopts a two-stage full-parameter training curriculum for the image model:
- Stage 1 - Low-Resolution Training (512 x 384). The model first learns the canonical 7-region inpainting formulation, dual-view garment conditioning, and coarse cross-view correspondence at 512 x 384 resolution.
- Stage 2 - High-Resolution Training (1024 x 768). Starting from the 512 x 384 full-model checkpoint, we continue full-parameter training at 1024 x 768 resolution for another 1000 steps to improve garment texture fidelity, silhouette quality, view consistency, and back-view detail preservation.
Both released image checkpoints are full FLUX transformer checkpoints, not LoRA adapters. Training code is available under the train/ directory. Dataset release is coming soon.
Future extensions include model-identity conditioning through the reserved identity placeholder and Wan2.2 multi-frame reference video training.
DualViewFashion/
|-- inference.py # FLUX-Fill image inference (fixed prompt)
|-- inference_caption.py # FLUX-Fill image inference (caption prompt, LoRA)
|-- inference_full_caption.py # FLUX-Fill image inference (caption prompt, full model)
|-- video_inference.py # Wan2.2 multi-frame reference video inference
|-- train/
| |-- finetune-mm-stage1.sh # LoRA training stage 1 (fixed prompt)
| |-- finetune-mm-stage2.sh # LoRA training stage 2 (fixed prompt)
| |-- finetune-mm-caption-stage1.sh # LoRA training stage 1 (caption prompt)
| |-- finetune-mm-caption-stage2.sh # LoRA training stage 2 (caption prompt)
| |-- finetune-mm-caption-stage1-full.sh # Full training stage 1 (caption prompt)
| |-- finetune-mm-caption-stage2-full.sh # Full training stage 2 (caption prompt)
| |-- finetune-multiview.py # LoRA training script (fixed prompt)
| |-- finetune-multiview-caption.py # LoRA training script (caption prompt)
| |-- finetune-multiview-caption-full-deepspeed.py # Full training script (caption prompt)
| `-- ds_config_zero2.json # DeepSpeed ZeRO-2 config
|-- assets/examples/ # Generated examples
|-- requirements.txt
|-- README.md
`-- LICENSE
If you find this work useful, please cite:
@article{dualviewfashion2026,
title={DualViewFashion: Dual-View Garment-Conditioned Multi-View Fashion Model Generation},
author={},
year={2026}
}This project builds upon FLUX Fill for layout-conditioned image inpainting and Wan2.2/DiffSynth for optional multi-frame reference video generation.