🗡️💀 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
See HIGH description
FAQ
Comments (119)
Excuse me, I encountered a problem: when I try to stack two girls together in an image, often their genitals collide, and sometimes a patch of skin ends up connected to one girl's genital area, which makes me laugh awkwardly 🤣.
Do you know how to fix this?
What kind of prompt would I need to avoid this?
The above is machine translation. English is not my native language, so there may be translation errors.
An image of a girl lying on top of another girl
I can not provide prompting help, sorry. There are to many factors, like the init image, the settings, prompt structure, quality of resolution/image ... No way to elaborate this. I never saw this happen, but I also did not do any yuri.
@darksidewalker Oh, it's okay, it's just a little try, but thanks for replying 😋.
Is RadiantCrush just an update to the last one? What changed?
It is an upgrade. Changelog is in the details.
Can you share with Prompts or You made all this on ComfyUI?
This is actually made with my workflow in comfyui
What's with that file size??
19GB?
Your first Wan 2.2 checkpoint?^.^
@darksidewalker are the previous versions not wan 2.2? Cause they are 13-15GB.
@Griphen116 NSFW Clip and VAE baked in.
@darksidewalker that makes more sense.
Thank you, Dev. I've paid the full version for support.
ERROR: clip input is invalid: None If the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.
It gives this error
You did use the latest WF without the latest checkpoint, I would assume. If you want to use the older checkpoint, try using the WF v1.1
Or rewire and just add your own clip.
@darksidewalker I downloaded it from here, isn't this the latest version?
@edomaster atm "radiantcrush" is the latest, working with the latest WF 1.2
@darksidewalker now another problem, the video is always 3 seconds, although I have me frames at a higher value.
Can you please give me an example workflow on how to use this merged checkpoint? Kind of new to this.
Already in the description and suggested resources.
are this new version mixed with the newest lightingx2v lora?
Yes, partly and finetuned.
how long 4060ti 16gb with 32gb ram do ? the size its 19gb im afraid my 4060ti cant do it
Don't worry. Your equipment is better than mine. My situation: Load with dasiwaWAN22I2V14B_radiantcrush checkpoint. Clip nsfw_wan_umt5-xxl_fp8_scaled. VAE wan_2.1_vae. RTX4060 8gb + 32gb dram + 128gb virtual ram. Video resolution 480x704. 81frames. ComfyUI WAN2.2 official workflow + sage attention. Total 6steps(2 high + 6 low), cost 401.9 seconds (first time), 406.05 seconds (2nd), 410.67 seconds (3rd), 376.83 seconds (4th). ^_^
@PotatoChipO3O thx
@PotatoChipO3O idk it froze my pc everytime i load the checkpoint
@Seii1 What is your virtual memory setting? Loading the checkpoint and other components requires 90GB of memory. The parts that physical memory can't cover need to be handled by virtual memory.
@PotatoChipO3O oh i need to set that , i didnt do it thats why, btw where i need to set virtual memory ? on my comfy ui disk?
@Seii1 Virtual memory is a Windows setting. You can refer to this video. https://www.youtube.com/watch?v=E9RWwYRE7wg
@PotatoChipO3O thx its works, but idk why u se the author setting the img always stretced
Radiant Crush - actually not better than teasing, I see lower details in all body parts (realistic video) , ( tested with your latest workflow)
Depends on what you do. I tuned down some things for better motion and prompting.
But that's the point, more motion means always less details in some situations. Overall I tested with 40+ samples and the outcome without any loras was significant better across all parameters. Also I incorporate latest inventions to make it as easy as possible to get good results.
But no problem the other checkpoints are available, if you like them more.
The next version will also be planed for future, improving things or doing alternatives better.
Unfortunately there is no "one-fits-all" checkpoint possible. I try to have it as versatile as possible and maintaining high quality. Some of the merges do some things better than others, that's why I keep them available. 👍
@darksidewalker ah, this answers a question...are these basically v1, v2, v3 or are they unique and worth keeping/using multiple releases and it seems the answer to both is "yes".
@BRai4now You got it. But I consider the latest the "best", in terms of generalised usage and output. I'm not a fan of v1...vZ, because most ppl think this is just better and better, but in terms of MoE and WAN 2.2 it is just not possible to get everything at once.
If you get it perfect on one thing, it will mess up all other, so one may always get a balance.
@darksidewalker yeah, everyone wants a magic bullet. It's why I collect so many WFs too, to see what might be missing from one, what's a real improvement and what's just someone's fever dream. And soooo many models.
I believe multiple Lora entries were merged, but from the user's perspective, it's unclear which ones were merged. When manually adding Lora entries, duplicates occur if similar ones exist. What exactly was merged?
Great power comes with great magic skills. Darkness lies within this forgotten knowledge. Use this tool of creation, but beware the consequences of your doing. Ahkn solari sum ohr nah kehm!
Seems pretty good. Faster than the other merged checkpoint I've been using, smooth wan. Haven't been able to gauge/compare overall fidelity to the 'new topics' they introduce. I do notice though that it suffers a similar affliction to smooth wan, that of bouncing hips when no bouncing hips are requested (and indeed, when they're mentioned in the negative prompt). Must come from a lora that both have merged, but it is a rather unfortunate/annoying.
Negative prompt is disabled with cfg1 (speed up). The sometimes not requested motions come from distillation. This is a side effect of the speed up. You can alter your prompting to mitigate that problem.
@darksidewalker Yeah, I run around 1.3 usually.
@Numerius CFG over 1 effectively is as disabling distillation and reward systems. Just for info.
@darksidewalker I'm afraid my understanding of the subject isn't sufficient to fully grasp the significance of that, but that explanation does strike me as somewhat inconsistent. If the unrequested hip motion comes from distillation, but distillation is disabled when I run at CFG above 1, then why would I be getting unrequested hip motion?
@Numerius Okay, lets give a little break down. CFG 1 is made for distilled and reward systems, to speed up the process and get a result in few steps. With >1 you gradually disables that feature, the systems will start to ignore the weights and just do what it would do with regular checkpoint. But with low steps this leads to hallucinations. At some point, I would guess cfg 3.5, it will fully have disabled that systems. But, mind, that's not linear.
It will get complicated here, but in the end higher cfg leads to longer generation times, because of disabling the speed-up's. So if you do not increase the steps at some point you will degrade your output.
So the one is giving you hallucinations or the other ^.^
But as always it depends on many things, cfg >1 can give good results, because prompting matters more with higher cfg.
There is no perfect formula, but I would all speed-up's normally requests a cfg of 1, because they are compiled this way.
You could try to experiment with "sigma shift", it will allow 3-8 for various outcomes. 5 is just the standard.
Oh man! Your new model is amazing 🤤🤤.
With your workflow default settings: 560:720 speed is super fast, high 68 seconds, low 68 seconds 😮.
cfg 1.5 high 138 seconds, low 136 seconds. I’ve never experienced such fast generation speeds 😄.
The only issue is that the model takes a bit of time to load, maybe that’s on my side. But thank you for your model 👍👍👍.
My configuration is: RTX3080 10GB, 64GB RAM, 49.152GB virtual memory, running on a mechanical hard drive.
Thanks to your composite model and clean workflow, even with my low-end setup, I can run it super fast 👍👍👍.
Thank you!
very good keep it up
Will try my best :)
Marked improvement to prompt adherence and movement. Great Job!
Thank you!
Incredible, I never thought it would even be possible yet here I am running 40GBs worth of WAn 2.2 checkpoints with a 16GB 5070ti, yet through your checkpoints and workflow here I am. 169 seconds for a 81 frame 720P video.
Glad you are this happy :)
I just dont believe some of these were made without loras lol too detailed. So what all nsfw actions does it understand and what other contexts? Cum without having to describe the fluid in detail and how it acts? Why do none of these examples have prompts to understand how you prompted to get such results?
I see, you are new to this. No problem! 👌
You do not need to just belief me 🤌
You can just download the videos, drop them in comfyui and ... BOOM! ... it is all there, metadate, workflow, prompt, checkpoint used, seed ... everything! .. Thats why there is "workflow" as prompt, because the video contains the complete workflow used. 👍Sharing is caring!
One thing, not everyone is sharing every prompt/wf, sometimes I also want some things be my secret magic, but most of the time and most of the ppl share their prompts/wf this way.
Hello, I'm new to comfy ui. I just downloaded your workflow and models, and I've noticed that when I use them, the video generated by using a high-noise model is blurry, and the video dynamics are poor when using a low-noise model. Comfy ui crashes when one model is high-noise and the other is low-noise. What's going on? This has been bothering me all day.
Hi, than I suggest you read this carefully first: https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition
@darksidewalker I finally succeeded in making the video. The strange thing is that I had to generate the video with a relatively low resolution (320*480) and then change it to a high resolution (I am currently using the 608*1072 you recommended) before it can be used normally. I did not change anything except the resolution.
@again10719971101233 I can start every resolution, it just eats more time. So this is maybe a limitation of your specs or the model was not loaded fully because you loaded something first? Don't know.
But glad you managed to get it work.
First off, excellent work, this plus Gladas' workflow is the closest I've gotten to my ideal output after trying many models and modifying many workflows, so thanks.
I have a couple questions for someone who knows how things work more than I do.
1. Traditionally more steps = more quality. This is a "4 step model" so does that rule still apply here? If I did 8 steps instead would there be an increase in quality? Would it decrease the quality because of something to do with the model?
2. Lightning Loras typically slow down motion somewhat, a common trick to combat that is to add a 3rd Ksampler that runs a single high noise step without the lightning lora, then carry that noise into the next high noise Ksampler that has the lightning lora in order to guide to generation towards faster motion. Since the "lightning lora" is baked into the model itself, there's no way to separate it from a Ksampler so how do "force" faster motion? Just better prompting? Stronger lora usage?
Hi!
1) Short answer: 4-8 is ok, rest will degrade/add no value.
2) 1 Sampler step with base 2.2 high model or better prompting/stronger lora usage
Moe KSampler, which is a single node handling both high and low has resulted in limited slow-motion for me, using no lightning on the high noise side. It also determines optimal shift based on the samper+scheduler combo, because it'll be different for each one and most people are just poking around in the dark on that otherwise.
@BRai4now I read of that, but there are many posts on git for problems with that one. I'm still testing myself.
@darksidewalker haven't had any issues. Been doing i2v and t2v with it and the streamlining is welcome. The triple sampler is one way to go but it's so messy and also guesstimating.
@BRai4now Oh, hey, how is T2V working for you with these? Like, it does output what's prompted (follows prompt much better than I'm used to), but I'm getting much lower quality than proper T2V checkpoints.
@voboyso oh, I wouldn't even try something labeled 'i2v' in a t2v context, though the other way around tends to work pretty well. It's sometimes frustrating in that I'm finding t2v more interesting than I initially thought. It's just i2v has some of the best look and style lora support.
Nice work! The models run pretty fast for each video generation.
However,I've noticed this model seems to be biased towards making people drool. For example, I'm using image-to-video to create the effect of precum or breast milk... But the AI mistakenly thinks it's the character's drool...
There is no explicit drooling trained in, maybe it is just bad luck/prompting. One thing is WAN 2.2 does not identify genitals and mouth 100% the time when used in conjunction with nsfw. So this could lead that effect. The only thing you can do is playing with Sigma-Shift or better prompting, maybe adding a extra lora for your purpose with 0.5 till 1.2.
@darksidewalker Thanks for your insight! It seems that the "licks her own lips" prompt will lead to this kind of issue.
@maxabcr2000709 Lips are on more than 1 place in nsfw tagging.^_^ You may be more specific like describing the way she licks her lips with other face details.
这个版本,动作上在4步的效果比原版好多了,但是风格化的动漫人物的细节会慢慢过渡到广泛的AI风格,比如眼睛渐变逐渐消失,一些棱角变平滑.特别是在眨眼后眼睛的细节一下会有很大程度的变化.我试过好几个改版的,都存在这个问题,在动漫人物上可能原版本的人物风格保留最好.
確實如此，使用官方的也是，應該是WAN22對動漫效果實在不好，只能在等等了
我也遇到了這個問題...這個模型什麼都好，就是表情變化上破綻很大，五官變化容易變模糊、風格跑掉。
我發現我用原生的WAN2.2 搭配 Start-end frame可以比較好的保留原圖的風格...
不過我使用原生模型也有自己的問題...就是產出的影片跑到最後會開始變暗...
it seems to have trouble creating pussy its just like a long clitoris fix please
This is not how it works, it will not create perfect genitals just out of the blue for any situation for you. It's still I2V. And if you want a spacial situation you should use a Lora for that.
@darksidewalker ah ok
This merge far surpassed my expectations! Great motion and great prompt adherence. Will definitely be using this all the time from now on. The only minor gripe I have with it is how it really likes to make the hips move in just about every generation I've done. Not always unwelcomed, but annoying when you specifically don't want it. Still a great model though! 👍
Could you share a prompt and image example of the "problem"? Could be over PM if you like.
Do you have a workflow for FLF2V(assuming it works too)? it works great but without FLF2V workflow, i cant make a looping video. workflow i was using for 2.2 FLF2V stop working properly for some reason..
The built-in WAN2.2 FLF2V workflow in ComfyUI is usable (you need to turn off the lightx2v LoRA), and I often use it for consistency fixing or creating loop animations.
I'm also working on a next version for my WF and a version for FLF, ... soon^^
@PotatoChipO3O i kinda managed to make it work with the old flf2v workflow i had
the problem is, it did work but the quality dropped and it got slower
as of now im constantly testing different setting and abling/disabling nodes
@darksidewalker nice, looking forward for it
@GensouKAIsaer I did noticed that too, but atm I'm not aware what the issue is.
@PotatoChipO3O Do you use the latest checkpoint "RadiantCrush" for FLF?
After some testing I get this with every WAN 2.2 checkpoint (native or not), with or without any LoRA/Speed-up. Steps don't matter, resolution, upscaling ...maybe a comfyui bug ...
I tried 4 different and my checkpoints with various settings. shrug
Changed VAE's, CLIP's ...
Needs elaboration...
@darksidewalker i notice a dumb reason why my gens was dropping quality and it was just because i was using lower reso (480x720<my default) than your default reso (528x768<ur workflow default)
i remade workflow with combination of your i2v and some other people FLF2V and it now generate faster and quality "fine"
cant say if its better or worse but its good enough for me
its kinda surprising its working fine even without dedicated clip model and vae....
your model is really impressive
@GensouKAIsaer Glad, my issues came from the MOE-KSampler. After some testing it does degrade things in some constellations.
@GensouKAIsaer dropped my FLF workflow, if you like try it out :)
@darksidewalker Yes, I use the latest checkpoint 'RadiantCrush' for FLF2V.
what was added on latest version making each file 5gb larger ?
never mind i found it, Baked in with latest VAE (2.1) BF16, NSFW Clip, Latest MoE Distillation (10/2025)
I'm really enjoying your model:
But where can I find the breakdown by model version?
In "About this version"
very good model, but prompting is a little bit different than original wan2.2 with some loras.
這個模型太優秀了，速度飛快效果還很好。比我用的Q8_0 GGUF好多太多。4060都能用
Unfortunately, this one doesn't load for me, unlike your previous versions. I run out of ram halfway through the generation, crashing during the loading of radiantcrush low. I was quite surprised how well I could run the previous ones on my hardware.
You tried with my workflow or did you use another that loads clip and VAE seperetly?
I was loading clip and vae separately. I just now ran it at 720 x 720 with only the vae ran separately, and it is working like a charm! Thank you <3
Worfkflow ran just fine for me but few hours ago I got same issue as the others guys commenting, when it reach checkpoint 2 low it crashes, weird
@Sexybeast10 Have you tried drag and dropping a previous generation way back to when it worked?
@MedliKnight Guy named caffeinsplash Found a fix for me that worked, add an unloadmodel node between the two samplers, feed the latent through it and attach the high noise model
@Sexybeast10 could you show me a screenshot how to use this node and how to do it?
@Sexybeast10 How do I do this? The node only has two options: value and model. Where do I feed what, and between which samplers should I insert it?
@MrDxrk On value u got 2 dots on left and right side that can go into Ksampler High, the left dot connecter and the right side dot goes into Ksample Low
@Sexybeast10 while I was waiting for answer, I looked for fixes myself and making big swap on ssd fixed problem
@MrDxrk Haha yeah I ended up buying a new ssd disk as I was using my old 2tb HDD disk, the ssd seemed to have fixed it all, which make me wonder how much faster it would go on my M.2 hdd as its even faster than a SSD hd
it doesnt load somehow, i have an 32gb physichal ram and up to 90gb virtual ram but seems lile it crashing when second checkpoint loads
add an unloadmodel node between the two samplers, feed the latent through it and attach the high noise model
@caffeinesplash That fixed it for me, was working fine for 2 days, restarted pc and then crash on checlpoint 2 Low, but adding the unloader made it work again, thanks
Recently comfy and some nodes seems to get broken because of incompatible code. I'm experiencing freezes since 2 days. Still investigating the cause.
@caffeinesplash comfyui should unload automatically between experts. Had no issue with larger models with 16GB VRAM and 32GB RAM. I had crashes for a while and found out torch compile was causing it. Not a single one since.
@rivdemon1221554 What did you do with torch to fix it?
@darksidewalker i made 140gb virtual ram and it works, good model though, thank you
@RLG I did solve the issues for me with the new workflow by eliminating mtb and easy-nodes.
Generating a video with my fav resolution (720:912) and my WF accumulates around max 50.6 GB memory (30-34GB RAM, 14.1 GB VRAM and 2.5 GB zRAM). That said, 140gb virtual ram seems massive ^.^
Hey buuuuuuddy, so your DaSiWa I2V checkpoints look really good, and without paying attention, I used them on T2V, which is obviously not optimal, but it followed the prompts really well!
I'm wondering if you're going to release a T2V version, as that/those would definitely be my main checkpoint(s), if so. I was super impressed by the outputs, other than the fact that the quality is quite degraded with T2V, so I just wanted to check and see if that is a plan/possibility.
Unless I missed something, I understand that I2V checkpoints do not work well for T2V, correct? Like, I'm not just doing something improperly to get good results with T2V, right? Hope you keep creating! Thanks for doing such great work and releasing these for everyone!
Hi mate!
Thank you for your kind words!
I'm not into T2V unfortunately. BUT ... there is @DigitalPastel making a awesome T2V checkpoint! It would be a major waste of time and resources to compete with such a good work. His i2v is not my taste, but from what I have seen the t2v (Smooth Mix) is really good! He is releasing a v2 atm and the samples look impressive 👌
@darksidewalker Thanks so much for the reply!