🔊 DaSiWa WAN 2.2 14B S2V - Lightspeed 🔊
🔮 Key Features:
🔥 LoRA-Free Generations
☄️Fast: 4/6 step generation
🪡 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
💡Hint for best results: Make sure your audio length is as close as possible to the video length (frames/fps=seconds). Longer Audio is OK or cut the video at the end.
⚠️ The first frame burn/over-saturation can not be patched. It is from the VAE. Only a workflow can counter this.
🍒Workflow
Make sure to checkout my easy to use C-S2V Workflow with anti-burn, upscaling and ♾️ true infinite length!
⚠️ Read the corresponding announcement.
📢 Make sure to check it out for in-depth information!
🛠️ Recommended Settings
Steps: 4-6
CFG: 1
Sigma Shift: 10
Sampler/Scheduler: Euler/Simple
Resolution up to 720p (native quality).
🖇️ Dependencies:
🤝Honourable Mentions 🖤
For "LittleDemon"
Awesome initial AI voice and music was provided from: 🎦 YT channel Omega-VIII
The inspiration of the sample/front-character was made with Lyara 🫶
Spacial thanks to S1LV3RC01N for help with quanting this model and with the repo! ✌️
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
🔥 LoRA-Free Generations
🌟 CFGZeroStar patch (better results and prompt adherence)
🍰 Baked Latest Distillation
🚫Excluded CLIP
⚠️ The first frame burn/over-saturation can not be patched. It is from the VAE. Only a workflow can counter this.
FAQ
Comments (25)
Hey dark, i like your WAN I2V models, would you explain if this S2V works with, you know......
Spicy and naughty stuff or nsfw ?
It is not like the I2V atm. It's there is no optimized NSFW implementation.
It can be done with the WAN 2.2 or 2.1 Loras at some extent. But it is not a model that will heavily consider your prompt and not necessarily understand a NSFW situation.
At this time it is faster and slightly optimized, without the need of extra speed-up.
@darksidewalker Oh boi... I'll post examples that make you reconsider training including nsfw loras :-). You did REALLY GOOD with this one
@Adaptalab0r I'm definitely looking than forward to this🥰
This is wonderful ! thanks
This is a game changer! With this model you helped me re-define my nsfw content. Thank you so much. First impression:
1) Somehow the model understands Loras far better than the original model. With the original I needed to push the strenght really far (1.5 - 1.8) to get even far less impressive results. Here I can stay between 0.6 and 0.9 AND get crystal clear results. JUST look at my first examples! which brings me to...
2) The image quality, like with your other models, is sharp and crystal clear. Yet, I did not bother trying high res examples. The first generations look good already even below 480p. Just wow.
3) Fidelity: Great success rate. 30% Perfect. 60% Good. only 10% rubbish.
So If you excuse me, I need to get back experimenting with his gem!♦
Ok I'm diving in dude you convinced me XD did you just use his workflow template and connected Wan lora + prompting ? Cause There is no High or Low in S2V model right ?
@hboxgames132 You only need The Low Loras and IT really helps combining them even If it‘s one of The wan2.1 loras. Well sadly IT semms bot possible to Link wan2.1 loras in The Video descriptions. In some cases, i was able to inject noise i.e. another Video into The latent space, which works wonderfully. Even better than controlnet in some cases. So, I built my own Workflow.
So basically in cases, in which the loras Work Well in it's own (Like your's for example), i generate The wideos right away. If The loras hardly do what they supposed to, i create a i2v or t2v before Hand and inject IT into The latent space of s2v Generations. Guaranteed success in The second case
Is there a place in the workflow we can change audio encoder to support other languages?
I don't know what im doing wrong but i got black screen result im using your workflow and tried a the regular workflow aswell. I switched LOAD checkpoint to load diffusion model nothing so far. Any idea ?
NVM this was sageattention apparently : (
It happens with me several times especially with DaSiWa's checkpoints, even when I disabled Sage Attention,, I will say what I tried with it so far:
when I used the argument in the run_nvidia_gpu.bat to force the animated preview in Samplers to "--preview-method taesd" (make sure if you have 'UI Manager' to set the same option in it) , the Black Previews & sometimes the final video's Black screen Became less than before.
Not a final solution & may not work for everyone, but I am just a very humble user with Zero technical skills.
BTW, DaSiWa Thanks for this optimized faster S2V checkpoint 😉
@donzo15 yeah that makes sense. Sage Attention only works because someone (ie the original model creators) ran the appropriate distill algorithm and uploaded the results to the sageattention repo. You could, for example, take this model and generate your own sage attention weights and use them. Just rarely good enough reason lol
All good! Yep thanks for the checkpoint Dasiwa after some test this is far better then the original S2V lora work great with it!
Would it be possible to get this to use no photo and do a t2i or t2v instead?
It is a WAN S2V, it needs an image and a sound file. It is made from WAN AI to be like that.
I tried, but unlike the i2v models the results are really bad. So, It's not made to do this and even if I tried a workaround without an image always resulted in body horror, when I tried it.
@darksidewalker By the way: Your latest i2v Model "Tasty Sin" is the best nsfw T2V model I've tried, Yes "Text 2 Video" You can just insert a white empty image and after skipping the first 3 frames the magic happens. It's freaking awesome!
This is much better and faster than Wan2.2 S2V.
You did a great job!
Could you grant permission for commercial use of this model's output?
Please modify the Civitai license in accordance with the provisions of the Apache 2.0 License.
"Apache 2.0 is a permissive license that allows modification, distribution, and relicensing. When using, modifying, or distributing a derivative work, commercial restrictions cannot be added. According to Section 4 of the license, the original license must be maintained when distributing derivative works, and any additional restrictions (such as prohibiting commercial use) violate the compliance with the original license terms."
Thank you, wonderful that you like it!
...but:
No, I wont change the permissions, I do not see any benefit for me to do so.
I'm free to add anything to the license if it is not the original model. Because I altered it, I'm free to license it like I want, Only the original must apply fully to Apache 2.0.
That said, it is totally my decision how my published models may be used and licensed. 🤌
@rivdemon1221554 Well that's not correct and there are ways to enforce it.
This is an awesome model, the workflow is great too. First time I've actually managed to make audio work.
using mmaudio to make sex sounds; it seems the rendered video the female's lips aren't synced with the audio. Are there settings for mmaudio I should be using? Or what ai audio creation should I be using?
S2V needs English speech to be lipsync. Sounds are no speech.
Can I use lora on it?
If the lora is compatible/made with WAN S2V, than yes. Otherwise, I doubt it will work very good.