Demo videos reproduction

[ChemrovKirill]

Hi Showlab team, thanks for open-sourcing your brilliant work!

I attempted to reproduce your demo results using the demo.py and demo_diffusers.py pipelines. However, I noticed that the face transfer isn't as perfect as shown in the demo. The face color shifts slightly, and the expression changes as well. Does the pipeline used for the demo differ from the one on GitHub? Or perhaps you used some post-processing with segmentation?

My result:

Demo:

[linyq17]

Hi, thanks for your interest in our work!

To reproduce our demo results, please use this model, which was trained exclusively on the reference task at Stage 3 and achieves the best reference performance.

Note that in the demo we omitted part of the prompt for brevity. The full prompts and results are provided below.

python demo.py \
    --ckpt_path ./models/ckpt/wan2.2_ti2v_5b_qwen25vl_3b_stage3_refvid_only_720x1280_81f_pad_first/model.safetensors \
    --video_path ./benchmark/RefVIE-Bench/source_videos/38.mp4 \
    --ref_image ./benchmark/RefVIE-Bench/ref_images/subjects/26_flatlay_techwear_jacket_1328_1328_1.png \
    --prompt "Replace the girl's cloth with a matte black techwear tactical jacket. All other parts of the video must remain unchanged." \
    --save_path test_output.mp4 --ref_pad_first true

test_output_concat.mp4

[ChemrovKirill]

Thank you for the quick and detailed response!
The prompt makes a big difference. I managed to reproduce the results with your example.