williamyang1991 / FRESCO

[CVPR 2024] FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video Translation
https://www.mmlab-ntu.com/project/fresco/
Other
734 stars 71 forks source link

Serious Flickering problem #46

Closed Sonia-kong closed 5 months ago

Sonia-kong commented 5 months ago

Hi, Thank you for your excellent work. I tested a basketball use case, the configuration file is below. I found that the video flickers obviously, is there any way to improve it? Thanks so much.

mininterv: 1 # for keyframe selection maxinterv: 1 # for keyframe selection seed: 0 prompt: 'In the basketball hall, teenagers wearing red and green uniforms are playing an intense basketball game, with a scorer standing in the distance' sd_path: 'stablediffusionapi/flat-2d-animerge' use_controlnet: True controlnet_type: 'hed' # 'depth', 'canny' cond_scale: 0.7 use_freeu: False batch_size: 4 num_inference_steps: 20 num_warmup_steps: 5 end_opt_step: 15 run_ebsynth: False max_process: 4 gmflow_path: './model/gmflow_sintel-0c07dcb3.pth' sod_path: './model/epoch_resnet.pth' use_salinecy: True

https://github.com/williamyang1991/FRESCO/assets/171914194/637ce66d-2771-4af8-8341-98a2cd3b6a30

williamyang1991 commented 5 months ago

Maybe you can combine diffusion with ebsynth. Avoid using diffusion for every frame, and use it for key frames. Then use ebsynth for non-keyframe. That is Avoid using mininterv == maxinterv

mininterv: 1 # for keyframe selection maxinterv: 1 # for keyframe selection

Sonia-kong commented 5 months ago

Maybe you can combine diffusion with ebsynth. Avoid using diffusion for every frame, and use it for key frames. Then use ebsynth for non-keyframe. That is Avoid using mininterv == maxinterv

mininterv: 1 # for keyframe selection maxinterv: 1 # for keyframe selection Got it, thanks. I will try it later. Much thanks to your reply.