❇️💀 DaSiWa-WAN 2.2 I2V 14B Lightspeed | NVFP4 | Safetensors 💀❇️
NVFP4 Version of the model.
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
🪡 NVFP4 precision (FP16 base)
⚠️ 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:
Aspect ratio 3:4 (560:720) or 9:16 (608:1072)
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 🫵🫢
Approximate expected quality from quantization
Refer to this comparison article.
This are my tests compared to a full fp16 safetensor checkpoint taking prompt- and visual satisfaction into account on my DaSiWa checkpoints.
⚠️ Do not compare this with the unofficial quants of my checkpoint made by others, they are based on FP8 and not FP16 like my quants!
🩺 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
See high
FAQ
Comments (44)
anybody good results? i only get mosaik effect.... gguf and safetensor fp8 works great.(rtx5090)
Make sure you use high and low in the right place
@darksidewalker its a blurry mess if i change high and low.i did it correctly.but all resolution the outcome is bad.its like mp speed quality compared to fp8 or gguf q8.its little bit faster but quality is like 30-50% lower....
@hartweizen sure nvfp4 is like Q4...it is not some magical high compression same quality as fp8. It is just faster on rtx 50xx
@darksidewalker i think q4 is better in quality maybe.but i will stop using it until the quality improves a lot.its like using sage attention 3.0 on wan 2.2.... fp8 and gguf is great ;)
@hartweizen The quality of nvfp4 will not improve, it is just high compression for lower vram usage. It is between q5 and q4, but it is just a high compression working faster on 50xx cards.
i'm having good results fairly fast (for me), but just tested simple 2D
the strat from user aurelius is the way to go.High nvfp4 and low fp8.faster render times without quality loss.
@hartweizen with what workflow?
@mugibot72 i used svi pro version with that combo. At lower resolution its way faster.
@hartweizen wait how did you get a gguf and safetensor file working together? in the workflow I downloaded it's only or, like you can use safetensor or gguf, not together? Right now I just use gguf q8 on both high and low and that been working for me. but i wanted to try this combo since i have a 50xx card
@mugibot72 no gguf and safetensor does not work. Download dasiwa v10 nvfp4 High checkpoint and Saftensor fp8+ low checkpoint. And select them in the workflow. It works very good and is faster then before.
And since u are using rtx5xxx install sage attention 2.2 .it also helps a lot to render faster.DO NOT INSTALL SAGE ATTENTION 3.0.
@hartweizen hm... no idea why but whenever i try and use nvfp4 High checkpoint and Saftensor fp8+ low checkpoint i instantly go oom. it tries to load both and fill 25gb or vram at once. I'm on a 5070ti and only have 16gb vram, i got 64gb ram but it doesnt seem to offload. when I use gguf it works with no issues.
@mugibot72 Im running rtx5090 and have 96gb ram.~28gb vram and ~55gb ram is used with the svi pro(1-4stage)
Since u have rtx 5070ti.change the video encoder(video output) to av1nvenc.i only render with 0.53 to 1.2 factor resolution.
And use svi pro workflow if that works for u.
I never tried but DaSiWa has a VaE chunking?! feature.i dont know how it works but that might help.
Fp8 scaled text encoder.
@hartweizen hmmm okay I'ma see if that helps! either way right now I been getting really good results with the gguf version of fp8 the speed and details are a lot better than what i used to use.
@mugibot72 Yea since v10 at low resolution the quality is way better then previous versions. He already announced v11 xD
I use 5090, but v10 directly causes a memory overflow and cannot be used.
did u update comfiui? mine runs but the quality is very bad compared to fp8 or gguf q8
The model can not produce memory overflow, only your app can. You should check your installation of whatever you use.
You really should use a higher quality model than this. You could easily run the highest quality withyour 5090 assuming the rest of your system is as good. If you're running a 5090 and a system with only 16GB of ram and a shitty CPU then you prioritized the wrong thing.
i'd suggest you to try wan2gp
@darksidewalker hmmm.The resolution I generate is 1216*832. I've been using v8 and v9, both of which work properly. Only v10 causes out-of-memory errors in video memory.
@ACancd v8 and v9 have no nvfp4 version
@darksidewalker Thank you for your reply. I found that I have been using the 14b version all along and didn't notice nvp4, though I don't know what the differences are between these versions either.
The low noise NVFP4 is too much quality loss for me, but using the high noise from this one and the low from FP8 seems like a good solution.
interesting strat
Yeah most things can be mixed, but beware, the understanding of complex motion is low on nvfp4 high, you may missing motions this way. If you are capable of running better quants, I would do it.
@darksidewalker I tested some stuff and I didn't notice much difference, if any. The 360 spin worked regardless, for example. It only got messed up when I used the NVFP4 low noise.
yeah thats the way thanks :D.its way faster now 20-30% and with fp8 low safetensor everything does look the same.overall motion is ok for me.
it's wonderful...
This is really really good so far. The nvfp4 and it being lighter than regular WAN is amazing. Is there any chance for a NVFP4 T2V variant? i have no idea how possible it is, but i can't find any such version.
I really don't like T2V ... so the chance is really low 🫣
For all RTX 5070 12 GB owners, this is literally a lifesaver. The rendering speed nearly doubles—yes, the quality drops a bit, BUT not as drastically as, for example, in the GGUF Q4 and Q3 models I used before. I’m no expert; I don’t know how this magic works under the hood, BUT you can really see the magic in the rendering results. Basically, if your graphics card has Blackwell architecture and you’re short on VRAM—this is THE choice for you.
Oh my spark now is very happy thanks
Awesome! 13,4s/it instead of 31,4s/it with Q4_K_S+lighting loras, tested on RTX 5060ti 16gb, 480x720. Tysm!
i cant run this with ur recent workflow, its just canceling
Missing nodes or incorrect settings, my guess. Not the checkpoint.
my bad what i mean is the workflow, the recent one,
good, generated time 150% faster compare to q6 , quality seem acceptable
Is there ltx 2.3 nvfp4 yet
@darksidewalker tks, i see boundbite have been removed some altering loras like the previous synthsedution i used before
So i have to use separate lora for ns
fw content. If you don't mind pls share me the some src of the altering loras you used in previous version or some self recommendation
is there nvfp4 of synthsedution?
@anhcaycaycay657 You'll have to go through the articles, since there are much changes from iteration to iteration.
@darksidewalker is there any lora can improve eye quality, if the input character face is near, eye is big and detail then output seem fine
but if the character far away,small eye,then nah
@anhcaycaycay657 Not using high compression quants like nvfp4 can significantly stabilize details. Raising output resolution also helps.