Generates WAN 2.1 videos in a fraction of time.
720p and 480p Version
Recommended settings:
Sampler/Scheduler: Euler/Simple
Steps: 4
CFG: 1
Sigma-Shift 5
Original Model from Lightx2v converted to FP8 quantisation.
☠️ Do not use any extra speed-up tricks or LoRAs or it may mess up your generations ... 🤬
⚠️ Hint: Most of the time the model is taking you by word. If you write "white" it is white. "Translucent" is translucent... like for the fluids. 💦 Now you know! 🫵 translucent whitish 🤫
⬇️⬇️⬇️⬇️⬇️⬇️⬇️⬇️⬇️⬇️⬇️
Recommended specs:
8 GB VRAM, 32 GB RAM
Sample times: <2 minutes for 81 frames, 4 steps on RTX 4070 Ti Super.
Compatible with 14B LoRAs.
I normally use 0-2 LoRAs, strength at 0.4-1 depending on how much the effect should be. 0.7-0.9 works best most of the time, not overwriting the style of an image.
At multiple LoRAs is seems best to tune the strength a bit down to 0.3-0.6.
Basic workflow example:
Here: https://civarchive.com/models/1811161?modelVersionId=2049602
My favourite UI:
SwarmUI https://github.com/mcmonkeyprojects/SwarmUI
Testing (my specs):
I can go wild on setting with this full checkpoint, even with added LoRAs:
121 frames possible: ~ 3 minutes
121 frames on 24 fps possible (more motion): ~ 3 minutes
128 frames on 24 fps possible (more motion and extended): ~ 3.5 minutes
Dependencies:
YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
Disclaimer
This models are shared without warranties and with the condition that it is used in a lawful and responsible way. I do not support or take responsibility for illegal, harmful, or harassing uses. By downloading or using it, you accept that you are solely responsible for how it is used.
Description
FAQ
Comments (95)
Can i use atleast two loras. its not speed up things but its to fix movements and textures. Will it affect performance?
You can use any LoRA, just not other speed LoRAs. I used up to 3 for poses/actions and so on...
Alright that's good news brother. What about your Wan 2.2 Lightning LoRAs, since you said "Do not use any extra speed-up loras.." Will it conflict with this Checkpoint i2v Wan 2.1?
This is amazing, it's so quick.
May I ask, do you have any tips on camera movement? It seems to ignore a lot of the movement commands other wan2 checkpoints are ok with. Do you have any camera movement tips?
Be more descriptive, I assume. All the speed-up LoRAs and tricks have a slight impact on prompt adherence.
darksidewalker Yep. That was the problem! Thanks :D
Wan 2.1 based checkpoints will be very hit or miss for camera movement without help from custom nodes and more elaborate setups. Even with perfect descriptions you may just have to try a different seed and roll the dice. They didn't train it as thorough as the 2.2 14B, and even that is really more 2.1.5
Cant get this to work at all. I have 12gb vram .. but always fails on model load using the provided workflow in comfyui. I've tried a lot of things. Help would be appreciated! Tried with the model in diffusion models using the right dtype. Which model is best for 12gb vram?
With comfyui I may not help much. Sorry. Maybe some other?
Its ok i spent a lot of time time messing around with chatgpt and comfyui. Eventually i have managed to get the small model to work. Thanks dude.
reed00112770 yo can you share that model? im in the same boat with a 12gb 4070
reed00112770 ComfyUI just straight up crashes for me as well when loading this model. Mind sharing how you fixed it?
Do you have first last frame to video?i used first last frame to video node and video last frame turn yellow
Unfortunately this one doesn't work for me. Probably because i'm the rare guy on a Mac M1 silicon using DT. I set it exactly to every setting you listed but it crashes while it's doing the initial "processing" stage. I WISH i could speed up my clips
i figured it out. "Draw Things" just did NOT like the LCM sampler. I switched to DDIM Trailing and now i get way better quality at quadruple the speed. Saves so much time. This is awesome. Thank you
这个会有动作损失吗？我的意思是这个模型是否比原来的模型动作幅度更小
works like charm, hope you make lightspeed for wan 2.2 either
This thing is like the invention of the Microwave! ⭐️⭐️⭐️⭐️⭐️
This is really strange. A few days ago, I was able to run it successfully and it worked quite well. But now, just a few days later, I can't even load the model—once I try, it causes my ComfyUI command line to freeze or pause!
I tried reinstalling ComfyUI and all the dependencies… but it still doesn’t work. I’m not sure if ComfyUI got updated or something… this is such a mess.
Loeee I may not help with ComfyUI, I use SwarmUI, but SwarmUI uses comfyui as backend and all is working fine. The only thing is SwarmUI just automates so much like choosing the right nodes and downloads missing models vae/clip...) and checking compatibility of ComfyUI. My ComfyUI backend is on 0.3.50 and it works fine.
darksidewalker I increased the virtual memory allocated by Windows, and then it started working again. My RTX 4070 with 32GB RAM kept running smoothly! I didn’t even know why I did it—but somehow, it worked!
Loeee Some other Windows user stated, that he need a significant amount of pagefile, maybe there is a bug inside windows that consumes huge amounts of RAM/Pagefile. Thats not a thing under my OS. So I may not help here.
Finally got it to run on ComfyUI, and while this is probably the best Wan2.1 model out there, quality wise, it's far from optimized. Also while the list might imply that you shouldn't do any "speed-up" tricks, you absolutely have to install at least Triton, it'll double your speed. (Tried it on a "clean" ComfyUI and got 40sec/it, with a fully decked out install - triton, sage, etc. - I got around 22sec/it. Same Workflow)
The main crux of this model, and why it's a hard sell, is the ridiculous need for up to 80GB RAM when loading. This makes the time to first inference around 20+ minutes and leagues behind loading both Wan 2.2 high and low noise models, while still being at roughly Wan 2.1 quality.
So for those who get a crash when loading this model in ComfyUI: Make sure that your system RAM + Pagefile have at least ~90GB total. (So 32GB RAM + 58GB pagefile, 48GB RAM + 48GB Pagefile etc.)
Just for comparison on my setup (SwarmUI/ComfyUI backend 0.3.50) @720p: Consuming 24 GB RAM, 4 GB (Swap/Pagefile), 13 GB VRAM. Thats not even close to 90 GB like yours. That's heavy!
darksidewalker Yea once fully loaded it'll drop down to occupying 40GB ram, during loading it'll peak above 80 and if the page file is smaller than that Comfy will just crash. (Also updated my backend to 0.3.50 and tried multiple loader nodes, all with the same results) Doesn't happen with any other model (Tried Wan2.1 fp16, fp8, gguf Q5 and Wan2.2 fp8 scaled and gguf Q5, has no trouble model swapping both Wan 2.2 models) Tried with both Torch 2.8+CU128 and CU129. I presume the two other people commenting that it crashes also run into a similar page file issue. This model seems very sensitive to some particular software/hardware configurations.
Good job in general, love the output quality, but to the edge cases like myself Wan2.2 is 66% faster.
InvictusAI Thats weird. And Wan 2.2 runs 2-3x slower for me as this.
it doesnt seem to work when it gets to the defusion model it stops and says reconecting at the top right then when i press run again it says TypeError: Failed to fetch
I'm using a 3050 rtx 8gb of vram
Worked immediately for me and super fast too. I noticed a lot of comments saying it isn't working but If you're having issues then either you have your files in the wrong folders or your hardware can't handle it.
I cant keep the characters mouth closed. Everytime they are always starting to talk
that's what marriage feels like i suppose
3060 12 GB. Now I can experience what it feels like using Wan 2.1. I always get useless results with Teacache and gguf, and I had accepted this as my life until now.
This thing is triple as fast and gave me results that I've never seen. Thank you.
160 frame vid, 5060 8GB RAM, renders from a half hour down to a few minutes. LCM magic done right. With some nsfw renders, it behaves like the model was trained with Japanese filter requirements. Legit renders, but weird fun wacky censoring over female crotches materialize. Dunno about that part otherwise yea, minutes on eight gigglebytes? Wow.
For nsfw content you need trained LoRAs for this. WAN is not trained for that like most basic models from companies.
@darksidewalker Yea that's no surprise. I'm noting the anomaly because I am using lora's, prompt magic, tweaking cfg/strength/steps/etc, yet the force is strong in this one. Overall I'm seeing a number of very strange renders reminiscent of gen1 SD1.0+. Tweaking cfg in increments of 0.1 from 0.5-2.0 does alter the composition (Lightspeed is a really sensitive model), gone up to 10 so far and it grinds its paintbrush in expected ways. Yet still, the artifacts are weird. I've had to negative 'blood' and 'gore' due to said weirdness as an example.
@catwalk_supermodel lightning/destilled models are meat for CFG 1, others do not work as expected. You could change the sampler+scheduler. But stay at CFG 1. Also negatives do not work if CFG is 1, negatives are disabled this way.
@darksidewalker I've done enough testing now to confirm the problem was two-fold. 1. Must use weight fp8_e4m3fn_fast, 2. keep the lora strengths ~0.1 +/- ~0.1. Lightspeed is amazing.
It works better than any quant and is indeed as fast
phenomenal... blazing speed, high quality, and none of the range of motion issues from typical lightning accelerators. Thank you!
I love this. Will there be a WAN 2.2 Model??
Maybe, but for now there are not quite the merges/loras to do one. Wan 2.2 is really new.
@darksidewalker yeah i havent even tried 2.2 yet. This Lightspeed has totally changed the game for me. I dont know how you did it but its wonderful
@Saratoga77 Thx, but it is not my work, I used lightx2v model and made a merged quant for easy usage. But glad it is useful for you!
NO WAY.... ITS SO FAST... UNREAL!!! and the quality is insane!!!
At cfg=1 the prompt adherence is really weak. Anyway its a good model because its so fast, so I'll stick with it for the moment. Thank you!
I don't what type of magic you did but this is awesome,thank you!
Thanks, but the most awesome stuff did lighx2v, I just made a quant/merge for easy usage:)
@darksidewalker still, nobody else did it so i really apreaciate it, sadly , seems like some loras from wan 2.2 seems to not work very well but overall a awesome job, speaking of wich, do you think you could do the same thing to 2.2 for a lightspeed version?, that would be a game changer
@alternative_Universe I really would like to do it, but atm it is tricky because of the split checkpoint wan 2.2 is. Plus the distilled versions are far from good atm, usable at best.
@darksidewalker damn, seems like I will stick around 2.1 for a while, thanks tho
sirve para 8gb? que alguien me ayude
yes it does. Im running it on 8gb and making videos in less than 3 minutes.
@EbonFantasy como lo hago amigo? cual es el workflow
@Mrfenderai https://civitai.com/images/99275606 you can download this video and drag it into your comfyUI. If you are using ComfyUI
is there a wan 2.2 version of this checkpoint?
Not exactly the same atm, but my first attempts to create something similar: https://civitai.com/models/1981116
@darksidewalker Cool thanks for the rapid response.
Sorry to piggyback off of your comment, but can someone briefly explain to me the difference between the 'WAN 2.2 14B' models and the 'Wan Video 14B i2v 720p' models? Feel free to link a resource if you have something, I can't find info on it. Thanks!
This is the successor of WAN 2.1 and WAN 2.2 FP8, the WAN 2.2 GGUF (FP16 base):
https://civitai.com/models/2190659
WAN 2.2 is always optimised up to 720p.
The old WAN 2.1 models had two versions one for 480p and one for 720p optimised.
@darksidewalker Okay, thank you. I didn't mean to post two comments earlier. This one wasn't showing up after I posted it, so I posted a new one. I'm assuming I should keep using the WAN 2.2 14B Models, then.
@dingledust definitely if you can. WAN 2.2 is absolutely superior to WAN 2.1 🤝
what is the difference between 720p and 480p models?
i was using this 480p model and was having a blast with it, today i wanted to try out the 720p one but everything, from the model size to output quality seem pretty much the same.
One is optimised for 480p the other for 720p, the 720p can do 480p too. If you use a low resolution both will do the same.
@darksidewalker thanks for the quick reply, and a great merge as well. i usually do square inputs with 640x640 pixels, which one should i use? the generation speed is identical but maybe i couldn't see the small details.
@pogid97357345 640x640 is low res, the 480p is great. If you go like 960x960 the 720p would be better.
@darksidewalker awesome, man, thank you very much!
Does this work with Wan2GP? I have it in the ckpts folder but it is not being recognized. have restarted, etc.. no dice..
Don't know. I don't use this backend.
Same problem here.
I have tested 3 steps, it give exactly the same result than 4 steps
remember that the more steps the better it will process whats prompted, including loras and textual inversions, sooo, if you stack loras you need more steps for it to process, for me i get really high quality at 10-20 steps !!!!
Wow. Mr RTX 4070 took 1200 seconds for a 5 second video at 480P. it improved litteral 10 fold as it takes 120 seconds now.
only thing that is noteworthy is the ai understanding of physics is more often prone to... garry's mod physics. Could be my setup. but i commend you none theless.
Hi! If you want the good physics, you have to go to WAN 2.2 :)
Can someone briefly explain the difference between the 'Wan Video 14B i2v 720p' models and the 'WAN 2.2 14B' models? (Not dasiwa specifically). Feel free to point me to a resource for me to learn about them, I'm unable to find anything on it. Thanks!
You confusing WAN 2.1 with 2.2, but if you want to know more about WAN 2.2 here is some guide: https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition
@darksidewalker So the models labeled 'Wan Video 14B i2v 720p' on CivitAI are WAN 2.1 models then? I think that's what was confusing me, that it doesn't have 2.1 in the label. So I'll keep using the 2.2 models then. Thank you.
@darksidewalker And thank you for this guide. It looks like this is what I need.
@dingledust The headline has the WAN 2.1 label, the model names can not be any longer, maybe I should change them
Thanks, I love it.
I'm using 480p (480/832) but I'd like to get a little better quality out of it.
So I'm setting it to 576/1024 (like in FusionX), but it doesn't significantly improve the quality.
I have a question.
Is it better to use 480p and set a higher resolution than the standard, or 720p and set a lower resolution than the standard (720/1280).
I have 8GB of VRAM, so 576/1024 is my sweet spot.
If you want better quality you may use WAN 2.2 instead of WAN 2.1
But if you want to use WAN 2.1, than 720p and set a lower resolution.
@darksidewalker Wan 2.2 with 8gb vram, is it possible?
@Rathma With a lower quant or enough RAM.
i guys, I have a problem using WAN 2.1 14B, This is my workflo: https://civitai.com/models/1802623?modelVersionId=2086227 , my problem is that its not generating any video, the final output is just a black square. I already tried debugging with gemini and chatgpt but was completely useless, could someone help me?
Update comfy.
In CMD screen I saw this "WARNING: No VAE weights detected, VAE not initalized." Is this a problem? I'm using your Workflow.
Well, you need to use a VAE.
Yes, i am using the vae. I think my gpu is weak (4060 8gb).
"D:\Stable Diffusion\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-videohelpersuite\videohelpersuite\nodes.py:130: RuntimeWarning: invalid value encountered in cast
return tensor_to_int(tensor, 8).astype(np.uint8)"
@darksidewalker It looks like I can use the 480p version of the model. After changing it to 480p it works as expected :)
I was also using sage attention and probably removing it helped too.
I would recommend you to use WAN 2.2 and maybe a q6 or q5 quant gguf, that's better than WAN 2.1.
WAN 2.2 can also use lower resolutions.
What does "Sage Attention, Flash Attention, Radial Attention, Q8-kernel, SGL-kernel, and VLLM are built in, so there's no need to enable them" mean?
Where can I find "dasiwa-wan2.1-i2v-lightspeed.gguf" ?
Also this works better than wan2.2 for what I was trying to do ^_^
Not available as far as I know and at least not from me.
@darksidewalker Ahh, it said to put it in the diffusion models in this workflow lol, went searching for it. Keep up the great work <3
@EngelsXIII Thank you! But there will unlikely any update to WAN 2.1 anymore, since the tech is not developed anymore.