search

kijai / ComfyUI-CogVideoXWrapper

981 stars 59 forks source link

The videos generated by the CogVideoX-fun models become mosaics #260

Open dys2017 opened 2 days ago

dys2017 commented 2 days ago

Hello, I have encountered a major issue. After updating the version, all videos generated based on the cogvideo X-fun model have turned into mosaics without any errors. However, they have become pure noise. May I ask what's going on?

dys2017 commented 2 days ago

I used these workflows for testing： cogvideox_Fun_I2V_02 cogvideox_Fun_I2V_Tora

kijai commented 1 day ago

On what hardware? And only those? Can't replicate the issue at all with those workflows at least.

kuronico59 commented 1 day ago

hi, same issue here at 49 frames, windows 11, RTX 4060 ti 16gb, 32gb ram, i didnt have that issue at 16 frames. i test at 32 to see. edit: at 32 it's ok. i do a new test at 49 edit2: and strangely enough my 49 frame test is ok now. although my first generation also made mosaics

in anycase, the workflows with 1.5 is ok ! great !!

DmytroSokhach commented 22 hours ago

My outputs are completely "burned". I started 2 days ago, so can't say about previous v CogVideoX_burned.json ersions. @kijai please see attached workflow and sample output. I would do additional tests if asked.

# Configuration:
Platform: Windows Python version: 3.10.11
Total VRAM 24563 MB, total RAM 65346 MB
pytorch version: 2.5.1+cu124
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 3090 Ti : cudaMallocAsync

https://github.com/user-attachments/assets/80a1a0ec-e9d8-49e1-8f3f-ca1485f9a73c

Attaching source image for reference FLUX_Metadata_0025

kijai commented 21 hours ago

My outputs are completely "burned". I started 2 days ago, so can't say about previous v CogVideoX_burned.json ersions. @kijai please see attached workflow and sample output. I would do additional tests if asked.

# Configuration:
Platform: Windows Python version: 3.10.11
Total VRAM 24563 MB, total RAM 65346 MB
pytorch version: 2.5.1+cu124
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 3090 Ti : cudaMallocAsync

24_11_21_130830_00001.mp4 Attaching source image for reference FLUX_Metadata_0025

This is an old workflow, most of the CogVideo nodes there would need to be recreated. Please check the readme for the big update notice.

DmytroSokhach commented 19 hours ago

This is an old workflow, most of the CogVideo nodes there would need to be recreated. Please check the readme for the big update notice.

I understand that on Update 8 was BREAKING ✌️ I've re-created nodes on the old workflow, cause they were giving errors, attached VAE links, etc. And now think the workflow I have attached has every node in latest state. Question to you as a master here: How do I find/spot "outdated" nodes without re-creating whole workflow? I've also checked JSON file, but could not identify version or anything.

kijai commented 19 hours ago

This is an old workflow, most of the CogVideo nodes there would need to be recreated. Please check the readme for the big update notice.

I understand that on Update 8 was BREAKING ✌️ I've re-created nodes on the old workflow, cause they were giving errors, attached VAE links, etc. And now think the workflow I have attached has every node in latest state. Question to you as a master here: How do I find/spot "outdated" nodes without re-creating whole workflow? I've also checked JSON file, but could not identify version or anything.

First I want to know if the example workflows work, if they do then updating old ones would entail basically re-creating all the CogVideo specific nodes (fastest way is the right click "Fix node" option). Depending how things were wired there may be bad links that don't error out but affect wrong values etc.

It's impossible for me to know how any other workflow is wired besides mine, I'm well aware updates like this are annoying but like I explained it became necessary to fight the bloat caused by all the new models/features that have been released.

DmytroSokhach commented 19 hours ago

re-creating all the CogVideo specific nodes (fastest way is the right click "Fix node" option). Depending how things were wired there may be bad links that don't error out but affect wrong values etc.

Thank you for pointing out "bad links that don't error out but affect wrong values". I've carefully re-created only CogVideo* nodes and result is amazing!

https://github.com/user-attachments/assets/ea10cc74-10f1-4653-9f54-67f668b82532