Thanks to the Big Brother of Fish_Bbbbbb from Bilibili
Almost based on this flow~
With the help of 2 creative loras for High noise and Low noise, it generates much quicker than my wan2.1 flows while better performance.
【2 key LORAs】
Wan2.2-I2V-A14B-low_noise_-4steps-lora-model.safetensors
Wan21_T2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors
Description
FAQ
Comments (7)
The 'trick' to low VRAM is ensuring you have at least 64GB of system RAM (a very cheap option), and then stream the models from RAM.
Sadly a massive number of people who could easily upgrade their RAM are told this would make no difference because all that 'matters' is VRAM. WRONG! Only with LLMs is it essential the model reside in VRAM.
If you enjoy video generation, for heaven's sake upgrade your system RAM!
When I run this workflow, the usage of RAM is about 40GB for system, which is not that much on my 64GB RAM computer. If I got a RTX 4060ti 16G, the longer video would be generated once, and there would not be a OOM exception.
So I think VRAM is more important than the system RAM.
My vram is only 12 gb on a 4070 rtx. I have 64 GB ram though and I can run maybe 80% of models that are claimed need 32 GB vram and all sorts. Yes it eats up nearly all my ram, and it's definitely slower and slows things down, but I still get 720p videos at around 4 minutes every 5s. Which is OK for me. But yeah I do get OOM sometimes, but then again, at least I don't have to have a frigging 1,000 + dollars gfx card.
@MadCat2k bingo
@MadCat2k Yeah，sometimes OOM，but click the run button to continue, then it works. Unnecessary to buy a 5090, because it donot improve the performance that much i thought, since i've tested that on a PaaS platform. The gap between my 4060 and the RICH card can be narrowed.
Also you guys should add more SWAP memory. Even with 64gb of RAM, having +32gb (or even +64gb) of PAGING file could get rid of most (if not all) OOMs. Comfyui is not memory optimized yet.
as an user with 12gb vram and 32gb ram, there is no hope to run wan 2.2? thank you for your reply