@Ada321 Yup. Although I'm using Wan on Pinokio, so that might be the conflict. This is the first time I'm seeing this error, both I2V and T2V presents the same error.

Same issues with WanGP through Pinokio, I'm assuming anyone using WanGP is having the same problem?

@Ada321 Seems to be all good now, thanks for addressing it so quickly, you the goat

yeah its the same for me, seems to be broken for now

Good news, looks like they responded to the thread on HF and fixed the issue with Lora keys. The fixed T2V is posted.

Actually, looks like they did the same fix for the I2V loras as well, and just uploaded new versions for both.

Hmmm it still broken for me

@Ada321 Strange, i'll just use the I2V one for now iguess, that one works fine

Yeah spoke too soon. Still broken for me.
The I2V lora doesn't work either. The keys are not being loaded, so nothing is actually being patched to the from the lora.

@Griphen116 @Ada321 Kijai just extract and drop couple of T2V Loras with different ranks https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Lightx2v

@Ada321 And now Kijay add a lot of new I2v Loras with diff ranks ;]

Do you have any workflow ?

@So6sson For me Basic workflow with Lora works fine. From comfyUi wiki: 'Wan2.1 Video LoRA Workflow'

Nice.. I just used a typical WAN i2v workflow and added this Lora. Super quick and good quality.

I just used my normal workflow with this lora, changed the CFG and sampler to LCM...

https://arxiv.org/pdf/2503.19462

Yea it depends of Loras. But try diff strength of this Lora (for exemple 0.6 Lightx2v) and/or other Loras (from 0.9 and under). Sometimes it helps a lot.

I had great success with .35 on this one, causevid stuff im my experience needs to be set weak compared to everything else

@dcham2310 Why LCM? I've been using Euler Beta, lightx2v strength 1.0, good quality

@compo6628585 you are talking aboy the selfforcing i2v lora? the first option?

@CharlieBrown0115 thats the one!

Ada321 does it matter if i use 480p or 720p version, i mean i couldnt find where to download 128rank 720p version

Ada321 Yes, I'm using cfg=1, shift=8. I loaded Lora using the 'lora manager' node that I used in my existing workflow. I'll keep researching.

ravenerkr841 Lightx2v works different with different other Loras. Try this Lora at lower strength. I got good result with 0.6 but it depends of your other Loras.

Its the new version of the self-forcing lora from lightx2v but with more more ranks, bigger file size, maybe better quality than the one they released directly. I think it's from kijai here: https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Lightx2v

There is also a comparison video between the different ranks in that repo i linked, but i'm not sure if the video is comparing the I2V lora or the T2V lora.

yes

For Wan Vace 14B

We wrote our comments simultaneously it seems. Sounds like we have the same problem. Something is breaking somewhere. I tried purging nodes and unloading ones, nothing working so far.

Bypass teacache and skip layer guidance if you're using them. They're being confounded by the sheer impudence of this LoRa. As would any sane node, one would imagine. They weren't designed for such tomfoolery. Anyway, that's what fixed it for me.

Ponder_Stibbons Well, that fixed it for me as well. Thanks, friend!

Ada321 I just located the issue. TeaCache is resetting too late and trying to reuse old data. Why this would happen I don't... oh crap. Layer skipping. Why didn't I think of that. This makes sense. There's a whole lot of crap in my workflow that is obviated by the LoRa that I hadn't thought to bypass to begin with. Not needed.... ah that took five seconds...bypassing teacache and skip layer guidance resolves the problem. Shouldn't be a problem enabling on the second stage, as there is actually stuff for the nodes to do with the second sampler.

I'm surprised that more people didn't have this same problem immediately as well. Seems like it should happen to anyone who enables this LoRa in a memory-optimized workflow. I should have tried an additional VRAM purge first probably, that might have fixed it too, albeit leaving me with superfluous nodes. It seems like the LoRa does some major tinkering to the model that is contaminating the poor VRAM like virtual variola. In any event, ditching the caching and skipping worked, and keeping them on the second stage is fine for a low denoise t2v as long as a different model is used. Despite the description, I'm keeping my cleanup stages. Minor headscratchers aside, this thing really kicks ass. How restrictive it is remains to be seen... but still, wow.

Wow thats impressive, mind sharing your workflow? The workflows i have, im seeing an improvement but only by a margin. Sounds like your improvement is vastly different than mine. Quite possibly the worflow as I use the fusionX ones which is a bit of a spaghetti monster with custom nodes.

Can you specify which LoRA was used and what these two sets of weights are?

Hey - do you happen to have a good workflow for this? I' really curious - haven't tried any self forcing before

xuadamux373  https://civitai.com/images/89299184
just save my video, you will get the workflow

ColorWolve You’re the best man, gonna check it out when back from work

Appreciated

xuadamux373 Enjoy~

Yo this is wild, my gens went from 30 mins to 5

Thanks man

xuadamux373 hahaha right, now this what i called kling at home~

ColorWolve why i can't get your workflow on my comfyui?

itasky 

https://civitai.com/images/89299184

Save this link video, can’t get the workflow? That impossible.. maybe try update your comfyui to latest version?

ColorWolve yep i move your video into comfyui latest version and nothing happens. i tried this external workflow app and doesn't show any wf when i put your video into it. https://comfyui-embedded-workflow-editor.vercel.app/

itasky i have no idea, i try this website you mention and none of the workflow is working, even the workflow from loras sample... i am using portable comfyui

itasky maybe your setting can only read json, so try this https://drive.google.com/file/d/1rdWzKxAsUdDe8nKF6dXNE8FDbsk7RABB/view?usp=sharing

ColorWolve oh now your video it's working. i don't know how i fixed maybe loading a json i donwloaded in other comment. By the way i am missing the finalframeselector node :/

itasky i think comfyui manager will showing what missing, so you can download from there

ColorWolve yep i am installing from Mediamixer pack. thanks for your support :)

itasky you are welcome =)

ColorWolve wow i reinstalled comfyui on Ubuntu and with your workflow i get the video in 100%|█████████████████████████████████████████████| 6/6 [01:09<00:00, 11.52s/it] :) so much faster than windows (without triton/sage)

itasky hahaha i see, but i am not fans of linux, so for me, windows speed is enough

For using Kijai WanVideo Wrapper I would say lcm and lcm/beta is pretty good at 6-8 steps flowmatch_causvid is a good one at 6-8 steps. flowmatch_distill is limited to 4 steps so it doesn't like any fast movement without being blurry. Test the various dpm++ schedulers and steps, they seem to do great.

which GGUF are you using if I don't mind asking

not the OP but personally I just use City96's Q3_K_M on a 3080 and 4090 rig. Quality is absolutely fine, and the generations are fast. umt5_xxl_fp16.safetensors for the clip.

DJLegends Depends on what works best for your setup.

Q6 for instance can be more resource intensive than a 14b FP8 model. If you can Run Q6 then switch to FP8 with quantization and you will be better off with speed, stability, VRAM usage.

GGUF model have to be unpacked as they generate which cause them to take longer, but if the Q model is small enough you can negate the extra time it takes. I use FP8 on 12GB VRAM card no issues, just can't do 720p 81 frames. Can do 576p(1024x576) 117 frames no problem and the quality/speed with LightX2V is great. (using Kijai WanVideo Wrapper)

On i2v I notice if you go high frames around 117 WAN will use your image as a last frame as attempt to loop another generation beyond its limit.

NAG and it will follow negative prompts even with cfg-1 (at negative write what you don't want at video)

flo11ok874 do you have a workflow or a guide for using NAG?

BinaryBottleBake It's easy - add 'WanVideoNAG' node with default setting. Put model dot to shift and sampler and conditioning dot to both prompt nodes. Or look at https://civitai.com/models/1736052?modelVersionId=1964792 there is NAG included. It's great workflow btw, just replace old Lightx2v Lora with this new one.

You can also keep faces more stable using this Lora: https://civitai.com/models/1755105/wanfusionxfacenaturalizer

Ada321 yeah but nag increases inference time, im trying to reduce it as much as possible.
i actually just did something simpler, just put more weight in the parts of the prompt that describes movement, it helps a lot

Try out the new Pusa Lora in addition to this at 1.4 strength. Helps with motion according to multiple people and i can confirm that after a few first tests.

marqs89 where is the lora，can you provide a link？

dwiwork0123561 pusa_v1.safetensors
https://huggingface.co/RaphaelLiu/PusaV1/tree/main

marqs89 i think it doesn't work.
i get this message in console
lora key not loaded: blocks.0.cross_attn.k.lora_A.default.weight

lora key not loaded: blocks.0.cross_attn.k.lora_B.default.weight

lora key not loaded: blocks.0.cross_attn.o.lora_A.default.weight

lora key not loaded: blocks.0.cross_attn.o.lora_B.default.weight

lora key not loaded: blocks.0.cross_attn.q.lora_A.default.weight

lora key not loaded: blocks.0.cross_attn.q.lora_B.default.weight

lora key not loaded: blocks.0.cross_attn.v.lora_A.default.weight

lora key not loaded: blocks.0.cross_attn.v.lora_B.default.weight

lora key not loaded: blocks.0.ffn.0.lora_A.default.weight

lora key not loaded: blocks.0.ffn.0.lora_B.default.weight

lora key not loaded: blocks.0.ffn.2.lora_A.default.weight

lora key not loaded: blocks.0.ffn.2.lora_B.default.weight

lora key not loaded: blocks.0.self_attn.k.lora_A.default.weight

lora key not loaded: blocks.0.self_attn.k.lora_B.default.weight

lora key not loaded: blocks.0.self_attn.o.lora_A.default.weight

lora key not loaded: blocks.0.self_attn.o.lora_B.default.weight

lora key not loaded: blocks.0.self_attn.q.lora_A.default.weight

lora key not loaded: blocks.0.self_attn.q.lora_B.default.weight

lora key not loaded: blocks.0.self_attn.v.lora_A.default.weight

lora key not loaded: blocks.0.self_attn.v.lora_B.default.weight

the complete message is too long, it goes from 0 to 39 and the generations looks the same, i guess it's not loading it at all.
i'm using wan i2v 480p GGUF Q8

Because that's how the "accelerator" LoRAs work, they achieve the speed up (well most of them at least, eg. CausVid, Lightx2v) by not needing to use CFG (aka it needs to be set at 1). Setting the CFG to 1 will instantly cut the gen time by half (and vice versa, hence you get doubles the gen time). There's one drawback to this tho, which is you can't use Negative Prompt if the CFG is 1, but there's a workaround to this, which is by using NAG (Normalized Attention Guidance). So, there's basically no drawbacks of not using CFG now.

When CFG > 1, the negative prompt is applied. When 1, the negative prompt is ignored, so the speed is doubled. So how do you put the negative prompt? Inject it into the model itself via the NAG node (like Lora). That's the magic.

FYI, my experience is that for AI, Linux is actually slower than Windows. Not much, but noticeable. I tested with Linux Mint, same version of the software, same CUDA version. It was like 5% slower.

Ada321 I'll keep trying, then. Maybe it's another setting somewhere.

try add Anime Style to your positive prompt

Ada321 If yes, why not?

It's for 480p wan

Download the video example, it contains the workflow.

Pusa lora is so large 4GB just for a lora 😥😥

fronyax Hmm, didn't realize it was that large, dang. You will have to work with what you can. I always go over the vram limit and work backwards from there until the dang thing works, you don't know if you don't try. With the blockswap I think about 16 - 17gb checkpoint model is the limit for 4070 Super 12gb cards.

The only parts that takes long is the fp8 model offload and the tile decode, the generation is "normal" speed. Sage makes it faster for sure.

I have to avoid the Pusa scheduler as it takes takes up too much resources and doesn't look as good. I use flowmatch_causvid scheduler with good results 6-10 steps if I like what I see with 4 steps.

I don't know how much 40 blocks is equal to in terms of memory, but that is the limit, more system ram won't do any good for the model part. More ram is great for offloading the other LoRA, encoders etc, and large upscaling.

--The Nunchaku team may be nearing completion in the coming months with their Nunchaku WAN 2.1 scheduled for their 0.4 release roadmap. Could be interesting, we'll see how accurate it is with the Nunchaku optimization. I do enjoy Nunchaku Flux and Nunchaku Kontext+Turbo LoRA so am excited about that. Even with Wan 2.2 out having a blazing fast low VRAM 2.1 is very good. Maybe the work done for 2.1 can easily apply for Wan 2.2--

hi, i don't find any information on how to add TAEW2_! and preview the generation, can you help me?

Pusa + this lightning lora + Fusionx is great

vAnN47 https://huggingface.co/Kijai/WanVideo_comfy/blob/main/taew2_1.safetensors

For the live playback of your generation. Put this safetensor file in your "ComfyUI/Models/vae_approx" folder

In your comfyui settings, go to VHS, at the bottom you toggle on the "Display animated previews when sampling"

Hope this helps.

no

It's up to lightx2v, not Ada321

I use Euler Beta and lightx strength 1

I am very new to this, but I have a similar GPU with 12GB VRAM. Q4 is too high, you should step it down to the Q3_K_M version, it works well for me.

tdougherty350505 It works with the native flow though. I ended up doing some research on this and I found a Github issue where Kijai talks about this. By his own admittance, the native nodes are much better at memory management, and they do it automatically. I'd rather not give up quality, so I'll continue using the native flow. I encountered another problem with this flow; increasing the block swap did get it to work, but the entire video is covered in what I can only describe as an orange filter. I decided it was not worth the effort to tinker with for now.

tdougherty350505 i have 3060 12GB and Q4_K_S 4steps total = 120 sec for 480x320

gambikules858 That's cool, but I want to generate in 6 and even 8 steps. It makes a huge difference in the quality of the motion

It's a moot point now anyway. I've decided that if I'm going to spend a lot of time on this to buy a better GPU. I can see the potential

Same