Video generation, often, requires a lot of VRAM. I spent a lot of time researching how to run video generation on my RTX 3060 Ti, and I want to share with you how I manage this.
All the workflows are in images or in attachments.
COG VIDEO
Cog video was the hardest one for me to work. From this source that CogVideo5bI2V requires 9GB of VRAM (I2V version requires only 5GB, but I want to focus on I2V generation). After some testing, I found out that enabling sequensial_ cpu_ offload this model started working for me.
Doing this, makes CogVideo to work, but it takes 39min to generate 49 frames.
Lowering steps to 20 redused generation time to 14/18 min (generation time is really inconsistent for some reason)
CogVideo2b (i used this version to use img2vid) don't require that much gpu but output isn't great. It is usefull only for people that want to get very quick results)
CogVideo2b with 50 steps (8 min)
CogVideo2b with 20 steps (3,5 min)
I also tried GGUF models, but they didn't work for me. 5b models don't work because
sequensial_ cpu_offload doesn't work for me, and 2b don't have GGUF models.
All of these results aren't great. I noticed that people are getting really good results, so i you know if I do something wrong, you can tell me in the comments.
Hunyan Video
The best video generation model that worked for me.
I remake this workflow, and used GGUF model instead of normal one.
While testing, hunyuan-video-t2v-720p-Q3_K_M worked for me the fastest.
Hunyan Video doesn't have i2v support, but it has v2v, so I added it to my workflow for people to want to use it.
Hunyan video (12 min)
BONUS - LTX VIDEO
Is a bonus becouse for now LTX Video works for me, but it generates VERY BAD videos.
I using the same paramiters as in this workflow (i just removed prompt enhancer), but i don't get even close to generating good videos. If someone knows how to fix it, please tell me in the comments. If i manage to fix this, i will update this article.