无法复现demo中的效果

[Z-Z188]

作者你好，这是一篇很好的工作。但是我用了你们提供的prompt和视频，生成的效果不如你们提供的，请问这是合理的吗？

8_edited.mp4

9_edited.mp4

[Dai-Wenxun]

哦哦这个是用了PE的，我来试试。

[Z-Z188]

哦哦这个是用了PE的，我来试试。

指的是Prompt Enhancement吗，请问能直接给出最终的prompt吗？

[Dai-Wenxun]

Enhanced prompt (Claude-Fable-5-hq)Replace the girl's white short-sleeved sailor-style school uniform top with a brown quilted down puffer jacket, its glossy padded horizontal baffles catching the warm sunset light, the fabric wrinkling and shifting naturally with her arm and torso movements as she walks, with the black backpack straps resting realistically over the puffy jacket and compressing its padding where they press down. Add a blue baseball cap on her head, its curved brim casting a soft natural shadow across her upper face, the cap staying firmly tracked to her head throughout her walking motion while her long dark hair flows out from beneath it and continues to move in the breeze. Her dark pleated skirt, black backpack, walking pose, and gait remain exactly as in the original footage, and the entire original background — the calm ocean, the white metal guardrail, the roadside grass, and the warm golden sunset sky — remains completely unchanged, with the new garments lit consistently by the same low-angle sunset illumination and casting temporally coherent shadows frame to frame.

[Dai-Wenxun]

哦哦这个是用了PE的，我来试试。

指的是Prompt Enhancement吗，请问能直接给出最终的prompt吗？

https://github.com/jd-opensource/JoyAI-Video-Edit/blob/main/deploy/xvideo/serving/pe.py

是的，我们现在用的是Fable-5，跑上面的视频的时候用的qwen2.5-vl-72b（指令遵循比前者差很多）。

[Z-Z188]

Enhanced prompt (Claude-Fable-5-hq)Replace the girl's white short-sleeved sailor-style school uniform top with a brown quilted down puffer jacket, its glossy padded horizontal baffles catching the warm sunset light, the fabric wrinkling and shifting naturally with her arm and torso movements as she walks, with the black backpack straps resting realistically over the puffy jacket and compressing its padding where they press down. Add a blue baseball cap on her head, its curved brim casting a soft natural shadow across her upper face, the cap staying firmly tracked to her head throughout her walking motion while her long dark hair flows out from beneath it and continues to move in the breeze. Her dark pleated skirt, black backpack, walking pose, and gait remain exactly as in the original footage, and the entire original background — the calm ocean, the white metal guardrail, the roadside grass, and the warm golden sunset sky — remains completely unchanged, with the new garments lit consistently by the same low-angle sunset illumination and casting temporally coherent shadows frame to frame.

我用了这个prompt跑了一下，感觉效果还是不如demo，脸部还是模糊

8_edited.1.mp4

[NancyFyong]

+1，我也是差不多这个效果

[Dai-Wenxun]

确实有运动模糊这个问题，这个demo的case我们挑的时间比较早，不太确定是哪个版本的模型了，我们得检查一下。

[yunpeng1998]

如果出现运动模糊可以试一下我的工作，用类似ditto一样的用wan2.2再做一次enhancement：https://github.com/yunpeng1998/Qwen-Video-Edit

[liangxiaowei00]

确实，运动模糊非常明显，+1

[Dai-Wenxun]

我们其实也发现了这个问题，感觉是因为T2V和VAE训练的问题，这个vae的T压缩倍率是8（目前大家采用的都是4）进一步导致了运动模糊，有点难绷。

[liangxiaowei00]

是有其它可用的checkpoint吗，或者如何调整下？

[liangxiaowei00]

https://huggingface.co/spaces/hugging-apps/jdopensource-joyai-video-edit

[Z-Z188]

我们其实也发现了这个问题，感觉是因为T2V和VAE训练的问题，这个vae的T压缩倍率是8（目前大家采用的都是4）进一步导致了运动模糊，有点难绷。

请问这个VAE是自己训练的吗，我看LTX的VAE的T压缩倍率也是8，但感觉模糊可能好一点？

[Dai-Wenxun]

是的，AR的话比双向diffusion还是少了很多算力的尤其还做了DMD加速