Holy crap, this actually works. It turns a 20 minute creation process into 3 minutes. No visible quality loss. Incredible.

By finetunes, are you referring to vace / fun? I was unaware there were any community ones.

How's the lora combination support? Fast / lightning have always given me grief with those.

And the samples don't really seem to be doing much for wan vids, not the best presentation frankly; if you could near replicate some existing published prompts it might be more enticing.

Nice thanks, can work with skyreels v2 I suppose?

Does anyone know if there are differences between the Kajai models here: https://huggingface.co/Kijai/WanVideo_comfy/tree/main

and the ones provided by Comfyui here?: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models

This also works with native comfy nodes/gguf by using the beta scheduler. So you don't even need to use the wrapper. (originally said AYS, but beta is even better and works with 2 steps)

For those with jumbled messy generations on subsequent tries from the 2nd or 3rd i2v, try disabling skip layer and teacache optimizations if that helps, it worked for me

Definitely seems to speed up generation but I can't seem to mix other lora's in with it and there's a noticeable drop in quality (I use GGUF though). It seems very likely my workflow just doesn't work well with it anyway.

what is this i am using wan 2.1 i2v is this gona helpful for me ?

There's a discussion on reddit of people's test results settings - good as 2nd reference:
https://www.reddit.com/r/StableDiffusion/comments/1knuafk/causvid_lora_massive_speedup_for_wan21_made_by/

I've seen noticeable degradation in subject's motion/movement quality with this LoRA. Other than that, it's a godsend.

Good for 'live wallpaper' style (limited) motion so far, quite impressive in fact, and fast thanks to the vastly reduced steps requirements, I will test more classical movement later, and I feel it will not be the same deal, but still, pretty good

Hopefully this will be updated, to help with the movement. If/when it does, it will be a massive game-changer imo, as it literally cut the gen times for me by easily 75%, while keeping video quality to a high level. I cant use this as is, its almost I2I for me 🤣

I tested using LORA, and it was possible to generate videos of the same quality as without using LORA, with about half the number of generation steps.

If you can reduce the number of generation steps, not only will you be able to shorten the generation time, but you can also reduce the amount of VRAM used, so many people can benefit from LORA.

Some observations after some short tests in 1 hour:

1. On native comfyui wan workflow it works with unipc sampler/beta scheduler, seems also works with gradient estimation sampler.

2. It can be used with other lora! Splendid!

3. Noticable speed up on fp8 safetensor for both 480p and 720p. For 480P, a 30-50 frame video can be generatee under 1 minutes. For 720p, a bit longer but the effect is still huge compared to 10mins per generation without the lora.

4. Does not quite work well with GGUF, at least in my test.

5. Quality loss is acceptable, but it depends on the subject I guess. For video with more static object/limited subtle motion/very linear motion/object barely change these kind, the quality loss is not that obvious. But for larger motions it does have more impact.

5. Really see great potential in this lora and technique......This can bring video generation speed to image generation level.

Get bad result with i2v, probably my workflow, is it possible to have a link to a good i2v workflow ?

this changes everything for me

Tested around and it works great with 25 frames (1 sec) and is very fast (needed 43 sec for generation)but as soon I use 97 frames its very slow (running more than 40 minutes).
Triton is also activated and I have a 4090.

Some numbers using an RTX 5070ti, 720x480, 8 second video duration.

Using this workflow https://civitai.com/articles/13328 the low-vram gguf version 3b,

Models: WAN2.1 i2v Q5_K_M.gguf, Clip umt5-xxl-encoder-Q5_K_M.gguf, with automatic prompt of florence2 enabled.

Enabled optimizations: speed regulation, CFGZeroStar, temporal attention, skip Layer, TeaCache (0.19), long video patch (5s+), TorchCompile, sageattention (native, without the node in the workflow, using "--use-sage-attention" launcher flag. Using triton-for-windows (haven't got triton native windows to run yet))

Settings without causvid lora: no lora, CFG 4, steps 20

Duration total: 490,67 seconds, second run 480,99s

Causvid:
Disabled: skip Layer (will create very noisy output if enabled together wit causvid lora)

Settings with lora: lora weight 0.3, CFG 1, steps 15

Duration total: 271,68 seconds

I only did a few test runs, so I can't conclude a reduce of motion yet, as it might as well be seed variations. I did use 2 lora in my testing for motion as well, so keep that in mind, you might render even faster without lora and notice motion reduction. This comment is mostly about speed comparison :)

can it run with GGUF model? :)

Absolute game changer. SageAttention + CausVid got me 5 second videos made in 240 seconds.
Setup was the following:

4090RTX
Wan 2.1 480 14b FP16 (comfyOrg)
SageAttention
CausVid 0.3 strength

CFG 1
Steps 14
3 Other Loras

(No Teacache, it fucks quality up bad for me)

So how exactly does this lora work?
How is this even possible?
It barely even has quality loss.

It will lead to an increase in sharpness and saturation. How to solve this problem?

I get better/more natural movements and better promt adherence with higher causvid values ~0.7.

0.1 absolutely destroys movement. 0.3 is mainly lora induced movement, but 0.7 is the sweet spot for me.

Seems weird, since others report 0.3 as good value..

Fantastic work!
Do you think it would be possible to somehow merge this LoRA into the Wan base model?
Or am I totally oversimplifying things here? 😄

workflow for native too?

god 100sec 65length 960*544 on 3060 lol (1.3B 0.5 lora, 6 steps , 1cfg). Native workflow

i found, what low cfg washes out liquid projectiles, only way to reduce that is increase cfg to 2, what increases generation time twice (61 frame from ~1:30 to ~3:30 with the parameters below).

For me optimal parameters for 14B:

- lora weight = 0.3-0.5

- Steps = 8-10 (later I'm doing a second pass with a 1.3B model)

- Cfg = 2

- Sampler = uni_pc

- Scheduler = normal\simple (normal I think slightly better).

P.S. additionaly interesting schedulers for cfg = 1:

-beta - creative

-ddim_uniform - even more creative

(Not usable with high cfg)

hmm i'm currently having an issue in which I have flashes at the begining of my videos with my generated output and idk how to remove that correctly
video with embedded workflow here: https://files.catbox.moe/kkdzs0.mp4

以下配置为我测试的3步最佳设置：

steps = 3

sampler = uni_pc

scheduler = kl_optimal

lora weight = 0.7

checkpoint = Vace_14b_Q4_K_S.gguf

注：使用KJNodes的Get Image or Mask Range From Batch去掉最开始的10帧（当lora weight >0.5时最开始几帧会有闪烁），但是即时浪费掉10帧，这个设置的速度依然足够快！

Do I save your custom node as a .json file? I tried that and dropped it in custom_nodes but it doesn't detect anything. does it need to be in a subfolder in custom_nodes?

I'm testing the brand new workflow and it works fine, but I'm seeing very slow speeds for having a mobile 5090. With Sage Attention enabled it takes about 4 1/2 minutes to render 5 seconds of video using the default settings of the workflow other than me changing the resolution to 416x608 (Half of the image size). Without Sage Attention it's over 5 minutes. I know the mobile 5090 is quite a bit slower than the desktop version, but roughly the same speed using another workflow without causvid but utilizing teacache and to my eyes the quality looks better. Anyone else faced with this? I feel like I'm missing something.

So I really like this, I've been using it in T2V 14b and after a while I realized it is actually doing something that I can't figure out how to reduce.

So when you run teacache into skiplayerguidance and turn that skip up high it creates a kind of stronger outline around things/ contrast. Now I have the model loader running straight into the shift and have bypassed all that and this lora causes that exact same thing except to a more severe degree. I've noticed it here too with peoples generations, a lot of my videos seem to put the subject ontop of the environment too, as in they feel like two separate things, it will generate the background env and then the person and they wont really fit together as well so I'm not sure how to remedy this. I've tried shift 1~3~5. cfg is 1 ofc, different samplers, multiple generations at different steps but it seems to happen no matter what.

I just thought id report my experience if it may help you in some way. Still think its really cool!

Its also making a lot of my subjects stay still sometimes too. strange

I've been trying to troubleshoot this on my own, but I'm honestly out of ideas at this point.

I'm using a Vace14b + Wan T2V setup, following a workflow (strongly similar to this https://www.youtube.com/watch?v=3tu-sTY0k6M) that involves masking objects with SAM and using a reference image along with some VACE magic to replace those objects. While the overall workflow is pretty solid, and causvid speed gains are incredible, I'm running into a recurring issue when using it. I sometimes need higher causvid strength to "help" in replacing the object, however higher strength casues some issues.

Specifically, the video quality drops noticeably—to the point where faces, both in the background and sometimes even in the foreground, get heavily distorted. Details like teeth just become a white blur, and the overall video looks like it's been hit with a bad compression filter. I don't run into this problem when using something like 20 steps with DPM++, the non-masked part of the video looks exactly as in the original, but any CausVid attempts, regardless of sampler or step count, consistently introduce these artifacts.

Has anyone else experienced this issue with CausVid? Is this just an inherent limitation of using it, or is there a workaround (maybe some scheduler/steps combination I didn't try yet) to preserve quality, especially in non-masked areas of the video?

There was no issue when I used this LORA at 0.3, but when I use it at 0.5 to 0.75, the video shows strong contrast. I used the usual I2V workflow — could something be wrong?

Very nice lora! It even improved my training lora.

But for some reason 1-2 frames will blur for a while.

Used your workflow and everything works well. However, when I try to add other loras, seems they just not working, no matter how I adjust the step and lora weight. Is there anything I missing?

been using i2v model, 3~4steps, causvid 0.3~0.4, cfg 6~7 unipc simple, and then 2steps causvid 0.4~0.5, cfg 1. It preserves lora motion and sharpness.
kl_optimal doesn't work for me. didn't use model sampling sd3.

Can't find CausVidControl.

ComfyUI-WanVideoWrapper is already on nightly :/

In theory this is great, in practice, not so much. It kills the video part of the video lol. I'll keep trying it with different Loras and settings, but Im not impressed. I don't want "videos" with little to no motion. That's a picture.

is there any 2 sampler workflow example with wrapper, not native nodes?

Am I crazy or does the workflow below contain only one sampler instead of 2 like the text above indicates?

https://files.catbox.moe/tnkcoz.json

what a mess. how about a workflow without all the python coding nonsense that doesnt work anyway.