Alternate settings for steps/sampler/framerate.

#54
by opasai2 - opened

What are your guys' experience with playing around with the settings? I usually run the I2V model and I've found that using a "higher" step count (6 or 8) usually doesn't change much, but it used to lessen the color distortion in the first frames with earlier models.

I've only tried testing euler_a vs sa_solver and it seems that sa_solver produces a bit more movement?

Also, I get great results with using about 16 frames per second and doing temporal upscaling for 1.5x.

Phr00t, keep it up dude! The v9 version is incredibly good, it rarely produces noise in the first frames and prompt adherence is out of this world.

I always use 16fps. Mostly euler_a/beta, sometimes sa_solver/beta and sometimes lcm/simple
About start noise you can cut couple of first frames in ComfyUi before generation like in this workflow - https://civitai.com/models/1855273/wan22-14b-rapid-allinone-workflow-with-lora-and-color-match?modelVersionId=2115514

But I think v9 don't have noise issue. Same like v1 which I also use many times. Both version give other results which is great, also some Loras work better or worse depending of version.

Anyhow, great Checkpoint, with all this models I forget how to use normal wan2.1 or wan2.2.

This is amazing. I can’t believe I managed to run this on my low-VRAM machine. RTX 3060 Ti with 32GB RAM. The first generation took 10 minutes, while the second and the rest took around 4 minutes at best. Wow… the prompt adherence and the quality are amazing too. amazing job!

DEIS, kl_optimal scheduler. Also gradient_estimation was decent. In another post he said to try 12 steps, which I will do.

Hello, this checkpoint is fantastic but why the charachters move so fast almost frantic? I couldn't correct with the prompt, any suggestion? Thanks!

why the characters move so fast almost frantic?

If you're using the exact workflow from the examples and the v9 version, then maybe your prompt doesn't fit the video time frame: I've found that too descriptive prompts in videos with a lower total frame count (about 41-61) will cause this kind of frantic movement. I would try lowering the video time length.

Also, post your workflow and prompt, that way we can help you better.

Hello, this checkpoint is fantastic but why the charachters move so fast almost frantic? I couldn't correct with the prompt, any suggestion? Thanks!

24fps? Try 16fps (and 81 frames for 5sec).

Thanks for replying, the workflow is the one from the examples, i'm trying the nsfw checkpoint and I prefer not to post the prompt :)
It's very simple though, I'll try different prompts.
A question, someone tried CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors? I did, but I don't think there's an improvement.
Thanks again!

Hello, this checkpoint is fantastic but why the charachters move so fast almost frantic? I couldn't correct with the prompt, any suggestion? Thanks!

24fps? Try 16fps (and 81 frames for 5sec).

Thanks for the suggestion, I'll try!

i'm trying the nsfw checkpoint and I prefer not to post the prompt :)

The nsfw checkpoint for me sometimes produces a lot of unwanted movement (especially face expressions). Another option to fix the frantic movement could be to use the normal checkpoint and add specific nsfw loras if needed.

Yeah, the NSFW merge does have lots of movement in it, which is typically good for certain activities. However, I do wish it followed prompts better when the movement is not desired... some of the LORAs I include overfit for that kinda motion... maybe I can find a better mix in the future to address that.

I would like to ask if I missed something. In my comfyui, I didn't see euler_a/beta and sa_solver/beta.

image.png

Python version: 3.11.0 | packaged by conda-forge | (main, Oct 25 2022, 06:12:32) [MSC v.1929 64 bit (AMD64)]
ComfyUI version: 0.3.52
ComfyUI frontend version: 1.25.10
### Loading: ComfyUI-Manager (V3.36)
### ComfyUI Revision: 3828 [39aa06bd] *DETACHED | Released on '2025-08-26'

I would like to ask if I missed something. In my comfyui, I didn't see euler_a/beta and sa_solver/beta.

image.png

Python version: 3.11.0 | packaged by conda-forge | (main, Oct 25 2022, 06:12:32) [MSC v.1929 64 bit (AMD64)]
ComfyUI version: 0.3.52
ComfyUI frontend version: 1.25.10
### Loading: ComfyUI-Manager (V3.36)
### ComfyUI Revision: 3828 [39aa06bd] *DETACHED | Released on '2025-08-26'

euler_a = euler_ancestral
and 'beta' is below at scheduler.

Sign up or log in to comment