The CogVideoX1.5-5B-I2V of diffusers

[liuxiaoyu1104]

System Info / 系統信息

diffusers 0.32.0.dev0
torch 2.4.1+cu121
python 3.10.14

Information / 问题信息

• The official example scripts / 官方的示例脚本
• My own modified scripts / 我自己修改的脚本和任务

Reproduction / 复现过程

I tried using CogVideoX1.5-5B-I2V and CogVideoX-5B-I2V based on CogVideoXImageToVideoPipeline(diffusers).
For CogVideoX-5B-I2V, width= 720, height = 480, num_frames = 49, num_inference_steps = 50.
For CogVideoX1.5-5B-I2V, width=1360, height=768, num_frames = 77, num_inference_steps = 50.

The generated videos of CogVideoX-5B-I2V are good.
In the generated videos of CogVideoX1.5-5B-I2V, the brightness of the first few frames is inconsistent with the images, and the latter part of the video exhibits blurriness and temporal inconsistency.

The image:

The result of CogVideoX-5B-I2V:

A.man.walking.in.the.road._480_720_49_1.0_output.mp4

The result of CogVideoX1.5-5B-I2V:

A.man.walking.in.the.road._768_1360_77_output.mp4

Expected behavior / 期待表现

The brightness of videos generated byCogVideoX1.5-5B-I2V is consistent with the images.