WAN 2.1 NOTES

icelouse I think as well that using a lightx lora on the high noise model specifically tends towards slowmo. On the examples I've run where I've removed it (same seed), it's usually produced a normal speed video. Still more testing to do.

Yes, the problem are the lightning loras. The best solution to finish with the slowmo videos i tried is to use 3 Advanced Ksamplers: First, one for High Noise model without the lightning lora and CFG >1, the next one, for High Noise model too but using the lightning lora and CFG 1, and finally, the last one with Low Noise model, CFG 1 and using the lightning lora. Usually i use 6 steps total, with 2 steps for each phase. This works really good for t2v and for i2v.

icelouse Please forgive me because I haven't actually looked at your workflow, but my guess would be that you're still doing 12fps like WAN 2.1 wants, but WAN 2.2 actually outputs animation speed targeting 24fps.

sirdrak for this setup, what is noise config? do you return with extra noise between steps 1 and 2, add noise between steps 1 and 2? I am just getting a noise fest trying 3 advanced, where 2 just works.

psspsspsspssspss Noise config for each Advanced ksampler:

- 1st: Add_Noise: enable. Return_with_leftover_noise: enable
- 2nd: Add_noise: disable. Return_with_leftover_noise: enable
- 3rd: Add_noise: disable. Return_with_leftover_noise: disable

ArtificialSweetener_ only the small model is 24fps. The 14b model is still 16fps

sirdrak The key to start getting good results was to bump the shift up to 5. I was using 3.5 from my two-stage flow, but this was too noisy for the non-lightening stage.

@psspsspsspssspss I'm using shift 7-8, actually 8... And the results are really good.

Nope. Doesn't even have to be close. I should have an example in the gallery below once it clears moderation.

I fine generally describing the motion at the start of the prompt, rather than a bunch of descriptors, is really helpful for getting it to start asap / faster.

wowwwwwwwwwwwww

Changing the weights alone didn't resolve 'slow motion' issues for lightning for me - but using a 3-sampler approach DID solve the issue. See this post for details: https://huggingface.co/lightx2v/Wan2.2-Lightning/discussions/20#6897529022ebf7a1e53df7c7

@kurokuro I have heard that 3ksampler approach a while ago. I wanted to try it myself too. Also I found that this "slow motion" issue has a couple of reasons. For example, lora weights, fps-lenght compatibility and so on. I'll take a look at that 3sample thing again. Thanks man

High noise section is supposed to generally have more strength than low noise section and using a lightx speed lora on the high noise side in a two-sampler workflow is problematic.

Running no lightx on the high noise side and 1.0 strength on the low noise side in a workflow like Fusion-X Ingredients is generally good for almost never having slowmo problems unless it's introduced by a poorly written lora (like the cinematic lensflare one, which kills both inference aesthetic and action.

Lightx lora usage on high noise side can work in a triple sampler setup where the first high noise side gets no lightx and then second high noise and low noise can use strength of 1.0 though in t2v I've found the 2.1 version of t2v lightx used in both high and low spot is better than the "2.2" version with high and low separate models.

@BRai4now Yeah you're right. But for now, I'm using cloud based system. so, I got no problem generally with vram or system ram usage. Sometimes I'm renting a GPU which gives 80gb vram and 100gb System Ram. So, yeah I'm giving money but no vram restrictions and OOM risk. 3Sampler method is a real life saver as you said

@ValarDohaerisTR and I've not tried it yet, but saw my first quad sampler setup, but that may just be silliness and offer no actual benefits.

I second this opinion. While I'm still working out my low-noise refinement due to penis issues, the results are already exceptional. Engaging... you can say that again. I just feel more of a connection with the act using this LoRA.

I agree with this