🗡️💀 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
FAQ
Comments (42)
Managed to get the model. This seems to work really well so far.
Looking forward to this coming out of EA. Can't buy buzz on civit anymore as it requires crypto.
This is basically one of the Quantized versions with merged Lora such as the Lightning Lora and NSFW Lora? If so... its way better to actually stack Lora a bunch of lora than use a merged model, Stacking gives you way more control on the Lora strength.. Using 2 concept Lora at the same time usually works better with if adjust the strength from each lora, merging all of them would be a mess.
Way easier also if the Creators update their Lora.. CubeyAI NSFW lora for example which is even experimental.
Nobody is forcing you to use it :) For me fine tuned merges did a better job in some cases. You are fully free to use a basic checkpoint with stacked loras.
I get what you're saying but I just built a workflow with this as the base and even on a very simplistic workflow with 1-2 additional loras I am getting some very NSFW stuff that is doing quite well adhering to a lot of different things. Usually my workflows end up with 8+ loras active on both high/low and it's kind of frustrating.
I know this isn't going to work for some, but it does for me. I think prompting is super important here but I am getting away with getting some real nasty stuff with specific prompting and like I said, before I'd be 8 loras deep, so 16 loras (h/l) which again, trade off of using 1-2 for me is WAYYYYYYY better. I can't stand complex workflows and micromanaging loras, which is kinda what this does. If you want a different type of scene/style/whatever you don't have to spend 5 minutes micromanaging all the loras.
I want to add, that I did not just lazy merged some LoRAs in, it was a process of 22+ optimizations and steps to get the result.
@darksidewalker I am about 25 minutes in of generated content so far and this is definitely by goto and base going forward. I can tell this wasn't just a "mashed together" but a thought out plan.
@MadCat2k Thx! I'm happy if it is useful not just to me. If you find some things that work good or bad, let me know! :)
Any chance of a T2V version?
I'm definitely not against the idea, but first I'll try to refine this one a bit more. Atm it has some issues I would like to fix.
Great job I am getting some good results. A lot of people are going to be upset they can't micromanage the strength etc and how each lora is handled, but for someone who hates micromanaging workflows this is great.
This is really good suff! If there will be a next version PLEASE exclude "bobbysmithy55555426 Male finishing LoRA" it ruined cumshot generations for me. Even using alternative loras at high strength the cum is waterlike fluid. Could you mabe replace it with another cum lora? But regarding the rest it really keeps a high quality image. I love it!
Thank you! You got me a nice hint there, I'll look into it, the process was more than 22+ steps and optimization to achieve this checkpoint. And I noticed the problem, like I wrote in the description. I definitely will try some things to get it work.
@darksidewalker Your determination to create another version fills me with hope and excitement. Thank you !
@Adaptalab0r hey, I updated the checkpoint, the problem should be solved on my testing.
@darksidewalker Thanks for the update, but to be honest I prefer your first version. Here is my feedback:
+ The motion speed is improved so it can generate decent movement even with 16 fps
+ NSFW motion looks more realistic/does what it's supposed to do
- Now it's the other way round: cum everywhere. Sweat is interpreted as cum. Every liquid that is dripping down is milky. cum squirting out of mouths (not using promts beside 1 word: Sex)
"cum"/"cumshot"/"ejaculation" in the negatives does not fix the problem.
- It appears that eyes are rendered more blurry. Overall quality seems to be not as good, but I cannot really put my finger on it (<5%)
As for right now, I use your sweetspot version which is awesome, and when it cumes to facials, I switch to Standard I2V + Lora.
And again: Thanks for your effort!!!
@Adaptalab0r Thx for the feedback, I'm still optimising parameters. I have a test-concept, but 20 samples are not enough to look into every detail, so this helps me improving things. I must admit that "hotspring" is a bit over tuned for cumshots ...lol... A next version may need some more testing. It is hard to get the right things on all edges.
This works great! Much better than stacking multiple loras approach that I am using. Thank you very much for sharing these models!!!
For the end to end workflow, do you mind giving me some advice? I like 4:3 ratio, so I generate t2i images using 784x1136 size as you suggested with 2x hires then create i2v videos using 784x1136 size + 81 frames + 16 fps. I wonder if there's a better approach you would use. Also I would really appreciate some tips for the Wan 2.2 prompting. Your example prompts looks simple but works surprisingly well(much better than other's long long prompts) and I wonder how I can learn more about it.
Hi and thank you! I can tell you what I'm doing, no problem.
I also generate an image using a model that fits my taste. I usually generate the image in the target video resolution or 2x of that. But I noticed that same resolutions work better and use a bit less resources and yield more stable results. I usually inpaint some details if the image is not as intended or has minor flaws. Thank I use this as initial image, sometimes I use that image as reference to create a second last frame image if I really want a specific end frame.
For WAN 2.2 prompting I really try to just write the important stuff. I find using 1 initial sentence for describing the main char and setting good to set the plot and than describing only the major things that should happen. With more and more description I found, that WAN has problems to put it all in the 5s, because it tries to do them all, but can't fit them in just 81 frames. With more frames one can describe more. Also I think it tries to do the things you write in sequence, so try to write what should happen after another.
Thank you very much for the detailed explanation!
Nice! Just what I needed.
Wan 2.2 itself made it a lot better for animation. I sometime mix 2.1 as High and mix 2.2 for Low.
I'm still trying to figure out how to reduce the amount of issues from making Pony images though. I love my checkpoint style and has a lot of details compare to some illustrious models. But damn the hands and toes keeps having more or less fingers/toes from here and there. Gotta constantly inpaint.
When 2nd pass + UltimateUpscale I need to reduce the noise to 15 as more will create random pussies between the pockets of inner knees/legs, but higher noise add more details. Such a pain.
Downloading both High and Low models now but I'm a bit concerned as I have a 4080 with 16GB VRAM and the High and Low are 14GB each! I hope it will run. Anything to avoid multiple Loras because, at least in my experience, the quality degrades with each extra lora.
[Edit] well, it did run on my 4080 16GB GPU but the quality was awful compared to my previous workflow just using WAN2.2 GGUF and Loras. Perhaps a link to a working workflow would help the inexperienced such as myself.
You only use 1 checkpoint at a time given, so 16gb VRAM should be fine, I have 16gb VRAM too, but 64gb RAM. Works for me.
can you share you're workflow please?
Using SwarmUI, I do not have a workflow atm.
Here are my updated basic workflows: https://civitai.com/models/1823089/wan-22-a14b-high-low-preset-and-workflow
getting 90 sec 480p on rtx 3090, seems good, thanks
170 sec on 920x640 OMG great quality and speed !
90 second length? 🥺
When I exported the video, color blocks appeared. This is clearly not a problem with the model, but with the VAE. I can't understand why. The VAEs for 2.1 and 2.2 should be the same, but some of the 2.2 models output chaotic color blocks, while others don't. I don't know what's going on. I tried every VAE I could, but I still failed. I hope someone can help me. Thank you very much!
I use the Wan 2.1 VAE and it works fine for me.
@darksidewalker Yes, I am using this VAE, but some models will only output a bunch of color blocks, while some models can output normally. I have no idea what is going on, and I wonder if anyone has encountered the same problem as me.
God damn another thing to download and try🤣
Pray to your Cat-Gods XD
But wait...where is wan animate version?? pepehmm
I'm already looking into Wan-Animate, but not ready to do some deeper research.
I think a version without baked in lightning would be cool too. Having a good time with this though
No way, sorry 🫠I do not have the spare time to test all the samples on native 20-30min generations per video.
lol np@darksidewalker
RTX4060 8GB VRAM runs 480x768 at 81 frames in 440 seconds, awesome.
I have no idea what i'm doing wrong. The high noise model just eats the 32GB of VRAM of my 5090.
It takes like 8 minutes to generate a video with 4 steps, 81 frames.
I got better results with models that are not supposed to be "light".
I'm also using SwarmUI.
I din't know if i'm doing something wrong 🤔
On my end it eats ~42GB RAM and 15GB VRAM, no matter if I use 480p or 720p. Times are 70s on 480p and 9min on 720p. I get almost same results on basic gguf checkpoint+lightning LoRAs with 90s/11min. All with 4 steps and 81 frames. I'm using standard pytorch-attn, not sage-attn.
Maybe you can get more in detail with your settings, so I can elaborate?
Thanks for taking the time to answer.
It seems to use 34GB of RAM and 31GB of VRAM, making the computer slow down drastically.
Though, i've got much better results later, a little less than 10 minutes for a 720p video, using two couples of LoRa. I guess it's fine even though i expected better results.
The model picture used has a bigger resolution but it doesn't seem to change anything.
I don't really know what kind of details i could add.
I may have been expecting too much haha
@Eligoal7 I added some hints of resolution and speed on the model page, for reference.