🗡️💀 DaSiWa-WAN 2.2 I2V 14B Lightspeed | FP8 Safetensors💀🗡️
My new flagship model for WAN 2.2 I2V generation - This is the best of the best!
This is a WAN 2.2 Model: You will need one pair of High + Low.
Version overview: https://civarchive.com/articles/23495/dasiwa-model-versions-and-timeline
🔮 Key Features:
🔥 LoRA-Free Generations
Generate high-quality videos without stacking Wan 2.2 LoRAs (unless you want adding spacial styles/concepts).☄️Fast: 4 step generation
Extreme versatile (more build in concepts)
Quality motions (less slowdowns)
🔞 NSFW + SFW:
Enhanced anatomy + poses + framing
Better understanding of sexual concepts
🪄 Better Prompt Responsiveness
🥺👉👈Better understanding of anime/manga style composition
🪡 FP8/FP8+ precision
⚠️ Read "About this version" details for the version you are using for more information!
🚫 Do not use any extra speed-up (low step) LoRAs, this is baked in already
🍒Workflow
Make sure to checkout my easy to use Workflows!
🍄LoRA's
Try first without additional LoRAs!
But: This checkpoint is not meant to replace all LoRAs, it is meant to:
Perform better overall at his own
As easy as possible to use
With LoRAs to be absolutely awesome
⚠️ Read the corresponding announcements.
📢 Make sure to check it out for in-depth information and a complex comparison!
🛠️ Recommended Settings
Steps: 4
CFG: 1
Sampler/Scheduler: Euler/Simple or Euler/linear_quadratic
Resolution up to 720p (native quality).
My go to settings:
0.52 - 0.83 MP
CFG 1
Euler/linear_quadratic
4 steps
16 fps
Sigma Shift: 5
Add other LoRAs with 0.3-1
16 fps, 81 frames ~ 5s
Dependencies
🩻 Known issues
Tell me 🫵🫢
🩺 Fixes & Feedback
If you use LoRAs, try to respect the LoRA training triggers and try some versatile descriptions, most LoRAs will work with 0.3-1.2 (start with 0.3)
Do not mass add LoRAs, just add 1 or 2 (x2 High+Low)
Negative prompting do not work with cfg 1, thats a limitation of speed-ups with cfg 1
Low resolution (e.g. 480p) are only for fast samples and will blur fine details, do a higher resolution if you want clear details
Before posting any questions I suggest reading my guide.
Update your ComfyUI ❗
🪧❗ Test your comfyui-backend with this absolute basic test-workflow before asking about errors.
🖤 Why I Made This
I was tired of using all these massive list of LoRAs, just to get a remotely good result after 10 generations, consuming hours of time.
So I can just make my videos with 1 or 2 concept LoRAs without pushing 6 till 10 LoRAs (Low/High) into a generation.
This checkpoint is also my personal playground.
Closing words
🤩 I want to thank all the fantastic other creators who made super nice LoRAs and concepts to play with! Support that awesome creators by using their LoRAs and post to their gallery and share the meta-data!
⚠️ I made all this with permissions or open-source resources (the time it is incorporated).
I share as much insights as I can without compromising my work. I'm doing this for fun as my hobby and just do not want my hobby to be destroyed.
More details can be obtained in the corresponding announcements!
If you would like to contribute in my awesome (😉) checkpoint or willing to share resources I'll gladly give credit! Just contact me!
✅ All credits / resources are mentioned inside the announcements! - Since different versions may have different resources.
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.
Custom License Addendum: Distribution Restriction
Notice: Notwithstanding the base license selected for this model, the following restrictive terms apply:
No Redistribution: You are not permitted to host, mirror, or redistribute this model (checkpoint, LoRA, or Safetensors files) on any other platform, website, or service (including but not limited to Hugging Face, Tensor.art, or SeaArt) without explicit written permission from the creator.
Attribution & Source: This model is officially maintained only on Civitai or other platforms where I explicitly own the repository. To ensure users receive the correct version, updates, and safety metadata, please point users to the original URL.
Usage: All other rights regarding the use of the model for image generation remain as per the terms and the restrictions provided per model.
Description
✅ Optimisations
🪡 FP8 (FP8 base) precision
💦 Better cum shots
🌊No more liquid waterfalls
🩻 Known issues
Sometimes it hallucinates things there are you did not prompt for 🫣 Must be a side effect of the NSFW adjustments
FAQ
Comments (96)
Since I started using wan, I downloaded a lot of light, each of which is more than 10g, taking up a lot of hard drive. I can hardly figure out their advantages and disadvantages. I am crying....
"Light" in this context means that you do not need the full 20 steps to produce a frame, only 4, so the "speed" to get a video is much faster.
Noob here, so do I use umt5_xxl_fp16 clip for this? Or can I still use fp8? This goes in the diffusion_models folder of comfyui yes?
FP8 is good, it is compiled with fp8 clip.👌
On SwarmUI the checkpoint (safetensors) goes to the "Stable-Diffusion" folder. Works also on ComfyUI for me.
@darksidewalker Alas, it only generates noise for me in comfyui. Don't know what I am doing wrong :(
@mygenaiessentials138 I did use a workflow testing samples, but SwarmUI for regular creation. I'll try to look into that.
@darksidewalker Thanks
The checkpoint works, did you maybe tried to use it as T2V? This is a I2V checkpoint.
@darksidewalker No, I did try it in my I2V workflow. That's weird!
@mygenaiessentials138 Did you overlap the steps, not 50/50%?
@darksidewalker Nah man, everything was good. No error on my part, it just generated noise for some reason. Probably cause I loaded it with some loras may be.
@darksidewalker Never mind bro. I got it working on a new workflow. Tried it again, and it works fine. Thanks. It's pretty fast.
Did you fix cumshots in hotspring?
18 h to download high model, 200kb/sec, LOL
same here
same, eventually I used VPN with Motrix to finish downloading in 40min each
As far as my test samples tell, it should be fixed.👍
Eagerly waiting for a Wan 2.2 T2V version :)
I'm not into T2V, so this may not happen with this checkpoint.
@darksidewalker I won't be giving up :)
Having no luck at all with this, it only produces animated static for me, using what I believe is a pretty typical workflow. Could you share the Comfy workflow you use with this?
On my testing the HIGH version was not downloadable correctly from civitai. So, I re-uploaded and than downloaded the HIGH version again. Generated 10 samples in 480p and all worked. So there is no problem with the checkpoint now for sure. I even added a new post with that test on higher resolution. If a problem occurs, check your swarmui/comfyui version. I did all on SwarmUI 0.9.7.0 and comfyui 0.3.60, sample https://civitai.com/posts/22808289
I'll make a comfyui workflow if there is time, but I normally use SwarmUI.
@darksidewalker I tried SwarmUI since you recommended it. It appears to be a wrapper around ComfyUI, so I assume functionally it's no better, and a little confusing. I'll give it a proper look when I have time. As for the model, I'll download it again and give it a go, thanks!
Solved, should redownload the High noise file, seems great!
I tried Hot-spring but it didn’t work for me. The workflow remained the same as my Sweet-Spot one, and the videos generated became a completely dizzy scene after 1 second. Any tip for the new edition?
Maybe you did use any speedup LoRAs with this, this will mess up things.
@darksidewalker I have the same issue; a few frames in, the video turns into noise.
My workflow is the simplest basic workflow, without any acceleration or nodes.
The same workflow works fine on 'SweetSpot,' but on 'HotSpring,' it turns into noise.
@s_07 I'll have to look into that, maybe my uploads got corrupted if this is the case. I'll download and try myself.
having the same issues as OP using Hotspring, just noise in the generation. No loras were used and on uni_pc simple sampler.
On my testing the HIGH version was not downloadable correctly from civitai. So, I re-uploaded and than downloaded the HIGH version again. Generated 10 samples in 480p and all worked. So there is no problem with the checkpoint now for sure. I even added a new post with that test on higher resolution. If a problem occurs, check your swarmui/comfyui version. I did all on SwarmUI 0.9.7.0 and comfyui 0.3.60, sample https://civitai.com/posts/22808289
@darksidewalker Yes, after re-downloading the file, the issue was resolved. Thanks for your work!
@darksidewalker Thanks, it works! I can see the liquid physics much better than with the old one!
Update: He reuploaded the file, it's working properly now!
Love what you're doing but it didn't work here. I tried it on my regular I2V Wan 2.2 workflow on comfy with the settings you recommended and no LoRAs, but after the first frame the video turns into colorful static.The static is quite pretty though.
I'm not sure what's going on, maybe my uploads got corrupted. My videos today went flawless with the checkpoint. I'll look into that.
Same here. Unfortunately i can confirm that after using the same workflow that i used on "sweetspot"
@Adaptalab0r 😔 sorry, to hear that. I'll look into that as soon as I can.
On my testing the HIGH version was not downloadable correctly from civitai. So, I re-uploaded and than downloaded the HIGH version again. Generated 10 samples in 480p and all worked. So there is no problem with the checkpoint now for sure. I even added a new post with that test on higher resolution. If a problem occurs, check your swarmui/comfyui version. I did all on SwarmUI 0.9.7.0 and comfyui 0.3.60, sample https://civitai.com/posts/22808289
@darksidewalker Thanks, mate! I'm going to give it a try!
@darksidewalker Thank you, I'll post something when I succeed :-)
is there a gguf version?
Not any time soon. gguf merges are a horrible thing.
@darksidewalker Well, the first FusionX merge got some GGUF versions and they all ran pretty nicely... but I respect your opinion
@H_for_Hi What I mean is doing a gguf merge is horrible, not gguf is horrible :) If I'm satisfied with the checkpoint and consider if finished, I may do a gguf, but my focus is make it as good as possible before doing other quants.
@darksidewalker Got ya, mmm, well, so it is time to wait, then.
q8 or q4 in future?
Do not count on that any time soon.
Sorry for stupid question. Can i generate photorealistic video? Or only anime?
yes, you can
Thank You!
Yeah should be fine, it just understands anime better as the basic one.
@darksidewalker, my great thanks!
Your jump in estimated times based on your 4070 ti:
480 pixel side length (384x576): 70-90s
720 pixel side length (592x848): 120-140s
960 pixel side length (784x1136): 9-11 minutes
Its that last jump to 960, is that major time jump because you run out of VRAM and have to resort to swap? I plan to run this in a server environment with a 96gb card like the a6000 blackwell or one of the H series gpus. Would that avoid the suspected swap issue?
I can tell you that I do not use any swap, because my system has basically none. It uses ZRAM, which is basically using RAM as swap. So this may slowdown a bit but I don't think it is that significant compared to swap. But ... I really can not tell you how good it will run on a production grade GPU.
@Kriston The RTX Pro 6000 Blackwell is a seriously powerful card and has plenty of VRAM, if you do not pair it with 16gb of system RAM then you will not run into any problems. On a 4090 with 24Gb VRAM and 96Gb sys ram I can render 80 frames at 832*1216 with 6 (3+3) steps in ~160-170s using WAN2.2 GGUF Q8 and several LORAs. This is with SageAttention/Triton and Torch compile.
Any chance of a quant version sometime? 14GB is still pretty hefty for those of us with plebian gamer-tier GPUs, but a q8 version could probably cut that down quite a bit
I notice you mention the training images were made with an "illustrious" model, does that mean it wont follow prompts as well as wan 2.2? (I remember trying illusterious and could not for the life of me get my head around the prompt engineering for images, kept having to look up danboru tags....which i still cant understand lol)
The initial images where made with illustrious, because this is a I2V model, you need initial images to produce a video. The training was not made with illustrious.
@darksidewalker slightly off topic i guess but has illusterous got better at prompt coherence? as i say last time it was a list of comma seperated words to prompt it some made no sense lol
@AI_man2025 Most models use tag prompts. Illustrious, eg my checkpoint understands normal sentences better, but it is still more useful with tags. Flux, Gwen and WAN understands sentences well.
@darksidewalker cheers. been looking into the tag thing, cant fully get my head around converting a proper prompt to a tag list and i cant find tags for a lot of stuff... but I am giving it a try lol
@AI_man2025 for WAN you do not need to use tags, only for like illustrious models
@darksidewalker I am guessing though that starting with an image from illustrious helps though (specially with more adult videos, given wans lack of knowledge of those areas lol)
@AI_man2025 This is a I2V Model, you always need to start with an image or it will not work. A last frame image is optional.
honestly got this to work a single time, fiddled with the things, went to work, forgot what I did and didn't save meta data. Now all this thing does is make the character vibrate lol
I don't know, the settings I use are in the description. They work 100% of the time for me.
You could try my workflows if that helps: https://civitai.com/models/1823089/wan-22-a14b-high-low-preset-and-workflow
EDIT: So I'm continuing the experiment and the problem I'm having is no matter what I try with an nsfw prompt the output will just be ejaculating and nothing else. Can't do penetration or anything.
@darksidewalker sure thanks I'll try that. Yeah I used your settings. Somethings gone wrong on my end. Not using swarm even though I prefer it. Seems like I'm better off learning comfy in the long run. What I got the one time this checkpoint worked was some of the strongest prompt adherence I'd ever seen. It could be that my prompt is screwing everything up.
Not bad model, but when I tried to generate some SFW video, always can get some white 'liquid' CUM from unexpected places. Well for example, from mouths of characters. Really weird.
By the way, I can not control the camera, any suggestion on prompting?
This is an issue with prompts. It is sensitive to prompts that could be interpreted as referring to a sexual act. For example, the word "blowing" could be associated with the act of performing a blowjob.
The checkpoint is optimised to be more sensitive and biased towards uncensored content. Therefore, be more descriptive or specific, or avoid sexual innuendos.
For content with just SFW things I would not recommend this checkpoint, you could just use a basic Wan 2.2 one.
Is it just me, or did all the files get removed? Versions are still there, but the file sets themselves are empty
My fault, re-uploading all atm ... could need some time.
@darksidewalker I just bought a new video card and want to try your 2.2 version, and it's such a lucky moment).
@Sing_Love Thats awesome! I wish you so much fun with AI! But be kind, the checkpoint is not perfect atm 😅
It just doesn't work for me because of the error "Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 21, 142, 98] to have 36 channels, but got 32 channels instead".
I'm using the same workflow as for Wan 2.1 Lightspeed, and GPT says this error most often occurs because of upgrading from 2.1 to 2.2. Have you tried this in ComfyUI? If so, we'd appreciate it if you shared the workflow for the new version.
I did try this in ComfyUI and use SwarmUI and updated my basic workflows. The error may occour, because you using the Wan2.2 VAE instead of WAN 2.1 VAE?
Also WAN 2.1 does not work the same as WAN 2.2, you also need to update comfyui to the latest version.
@darksidewalker I use:
umt5_xxl_fp8_e4m3fn_scaled
wan_2.1_vae
In general, I use the same things as for wan 2.1 Lightspeed.
@darksidewalker What Clip Vision are you using? Maybe that's the problem?
@Sing_Love umt5_xxl_fp8_e4m3fn_scaled
I got this error in the past if I used the wrong VAE and or had a old backend comfyui version.
@darksidewalker umt5_xxl_fp8_e4m3fn_scaled — it's just CLIP.
There's also Clip Vision, which is used to convert an image to CLIP_VISION_OUTPUT. I'm using clip_vision_h.
@Sing_Love You can look up my settings here, I do not use clip_vision:
https://civitai.com/models/1823089/wan-22-a14b-high-low-preset-and-workflow
@darksidewalker Thank you so much! I'll check it out a little later!
@darksidewalker Same error...
@Sing_Love And your comfyui is on version 0.3.60, right?
@darksidewalker Yes.
@darksidewalker I'll try updating all the nodes now. Maybe that's the problem. There just can't be any other reason for this problem anymore.
@Sing_Love I really do not know whats going on on your install. I can say that my comfyui is almost clean. I did only install "rgthree", video-helper and gguf nodes. Besides that I use SwarmUI and not comfy to render my videos. Maybe you try SwarmUI, makes a lot of things smoother.
@darksidewalker Updating the nodes didn't help either. I don't understand...
@darksidewalker Are you using this VAE?
https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors
@Sing_Love I can't tell the exact source of my VAE. This was auto downloaded from SwarmUI. But my VAE seems a bit smaller on bytes - 242,1 MiB (253.860.343 bytes)
Can you please just do a clean install? That will fix it
@mich3lang3lo Yes. I just reinstall ComfyUI, and it begin working (lazy to use translator, so it's my great english).
had same issue, had to add --disable-all-custom-nodes to the bat file and reload and it worked, now my issue is grainy images
@masonkriss155 this can be 1) you did not use a initial image or 2) you did not use cfg 1 or 3) mixed up High and Low on steps
@darksidewalker had only a high noise lora in, didnt know i needed another sampler with a feed in from a low one as well. Now i need to make it from 10 min gens to 3
@masonkriss155 All Wan 2.2 models are MoE models, you need both. 50% steps on high and 50% steps on low. Also this is not a LoRA is is a full checkpoint. But you can mix in LoRAs.
@darksidewalker lol yea it took me a second to stop and read the directions, but I figured it out