🗡️💀 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+ (FP16 base) precision
🌟 CFGZeroStar patch (better results and prompt adherence)
🍰 Baked Latest Distillation (r64-1022)
🍀Additional concept optimisation (compared to MidnightFlirt)
🎬 Reward Attention (more realistic movements)
✨ No or extreme low transformation of details with anime/realistic images (lips, eyes, ears, breasts, genitals, ...)
🌠 Refined motions (compared to MidnightFlirt)
💌 It can guesstimate better details out of frame (compared to MidnightFlirt)
💫 As close to RL motion-speeds as I could get with speed-up tech
🍌Even less tries for good results (compared to MidnightFlirt)
🧩Even better compatibility with LoRAs (compared to MidnightFlirt)
🖼️ Usable with your preferred/custom CLIP (if compatible)
😵💫Additional reduced hallucinations (compared to MidnightFlirt)
💠 Zero prompt results capable
🚫Excluded CLIP
🩻 Known issues
🫦 The most delicious sin!
Tell me 🫵🫢
⚠️ Requirements:
Comfyui ≥ 0.4.0
FAQ
Comments (78)
Any recommendations on getting it to stop adding fingernail extensions to hands? Nearly every generation I do it gives the women (and sometimes men) fake fingernails or obvious nail extensions, even if I prompt otherwise.
When compared to "DaSiWa WAN 2.2 I2V 14B v8-v1 | Lightspeed | Safetensors" and "DaSiWa-WAN 2.2 I2V 14B TastySin v8 | Lightspeed | GGUF"
Besides the difference between gguf and safetensor, what is the difference?
It is all documented on the front-page, the announcement article and the "about version". 👍
I wont write down this all in a comment. 😅
I have only black screen as output, but if I use other GGUF ver, works fine... Any solution?
Update comfyui to 0.4.0 or later
The version requirements are in about this version mentioned
Yes, sorry, my bad, was sure its latest version - fixed, now its fine
is it something different with the mode? I still get black screens despite updating, I'm on wanvideowrapper
me too, I tried using --use-pytorch-cross-attention but now it says me: unet unexpected: ['blocks.0.self_attn.q.comfy_quant', 'blocks.0.self_attn.q.weight_scale', 'blocks.0.self_attn.k.comfy_quant', 'blocks.0.self_attn.k.weight_scale', 'blocks.0.self_attn.v.comfy_quant', 'blocks.0.self_attn.v.weight_scale', 'blocks.0.self_attn.o.comfy_quant', WARNING: No VAE weights detected, VAE not initalized. no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
@saehara151 This error is something else, you have to set clip and vae.
I can't get the model to work with WanVideoWrapper, is there a reason it would work with the workflow you provided but I'm not really sure as to why It wouldn't work otherwise, with the only difference being euler simple vs dpm++_sde with beta
I do not use the WanVideoWrapper, but I assume the problem is this node. You could go to his github and post a bug report why the node produces blackscreen with fp8 mixed quant ops system in comfyui.
BTW - I tested with WanVideoWrapper/Sampler and TastySin-v8.1 - It is compatible with it. It produces just fine videos. The set-up is convoluted, but completely possible.
Very excellent model! Five star rating! When creating videos, the original image can be accurately used as the first frame, and the consistency in the subsequent frames is also excellent. I have used many smooth and remix model workflows shared by experts, but I have been struggling for a long time without achieving the same effect as Dasiwa. This will be the only I2V model I will be using for a long time! Thank you very much, author!
Thank you for the kind words!
What workflow are you currently using? Would you mind recommending it to me?
Great model. Will you convert it to GGUF quants? Thanks in advance.
Serious question? 🤔
Did you not tried even reading the 3rd line after the headline of the front-page or tried searching over civitai search?
@darksidewalker Nope. I just look at the models at the top. Too many colors and descriptions. I gloss over it. Are we getting GGUF or do I have to wait for Bedovyy to convert it?
@OzzyOsman So... You were looking at his HF page? Right? 🤣
@darksidewalker I try to quantize the GGUF models myself, but Wan2.2 is a tricky mistress. I like your models, have them all, but prefer how the GGUF works for speed. Anywho, keep up the good work.
@OzzyOsman I was joking around, because there is already a gguf version, pinned under resources, mentioned on the HF page of the guy who made quants of my older model and mentioned on the overview of my models. I really have no clue how I could "advertise" even more XD
Any ideas why the video tends to loop? For example, "a girl leaves a room," but in practice, especially if the video is longer than 7-8 seconds, she returns, often backwards.
The loop setting is disabled, of course — this applies to many wan 2.2 builds, not just this one.
In Wan 2.2, generating videos longer than 5 seconds often results in looping
WAN is made for 81 frames, all over this tends to loop back. It is what it is.
Exception is WAN S2V and WAN-Animate, they can produce virtually endless videos.
Check you didn't enable "pingpong" in the node setting. This is what causes this.
As someone already mentioned is the pingpong parameter what causes this. Check your node parameters, probably in the Video Combine node, if I am not mistaken, there is this parameter named pingpong, you have to set it to False and that will do it
This model is absolutely amazing. I'm looking forward to the next update! Keep up the great work!
Thank you very much! I planned one, but know it the time to make some art myself and pushing the model to its boundaries ^.^
Using WanGP2.2 I get results that just look like white noise. Previous versions like Midnight Flirt worked fine, but this one doesn't
So you may ask the WANGP2.2 creator why this might happen with fp8 mixed precision ops. 👍
You teased me about this one and I'm not disappointed ! I throw away all my Loras ! It works like a charm ! you did an awesome work once again ! thank you for sharing !
Thank you! But please throw not all LoRAs away, some may be needed and the awesome creators also need credits! :)
Where did the t2v version go? Will there be a new one?
Where? - There never was one.
Will there be a new one? - I doubt that ^^
@darksidewalker Just recently there was something, it seemed to me that it was a t2v model. But it immediately disappeared, I didn’t even have time to download it.
@Limonobatono I never did a t2v model. 🤔
@darksidewalker Perhaps I got something wrong. It's a pity. It would be interesting to see one like this.
Black screen.
Blue screen. 🤣
About this version.
Also getting black screen results. Not seeing anything that suggests why in the About section
@markharper80266 Look closer. There are "requirements".
Blacksceen because you use WanVideo Sampler ...
I have the same problem before. So I updating Wanvideo Wrapper. It works for me.
Hey thank you for adding the version numbers. I think i remember a comment around here that most of these seem more like flavors of focus rather than actual incremental improvement but smooth brains like me like ascending order. There are already so many choices to make, this make one easier, even if it isn't necessarily accurate.
👋 I decided to do it because of adding also gguf quants to the list, to give more clear hints what version is what. I'm not against numbers, but I like my version naming, so I'll stick with that too😁
Damn no "WanVideo Sampler" Support ... please write it on your model info page
KSampler works fine but i hate KSampler :D
You could make a report at his github and ask to support mixed precision, like comfiui native, than it would work, I assume.
I can not test any custom node, if it works native (and ksampler is) is is ok.
I won't write it on the model-page, because only v8.1 is mixed precision, so that would not be right for other versions or future versions.
BTW - I tested with WanVideoWrapper/Sampler and TastySin-v8.1 - It is compatible with it. It produces just fine videos. The set-up is convoluted, but completely possible.
LureNoir and TastySin are both really good, have excellent motion and prompt understanding. It understand concept such as scene changes, camera movement etc very well.
It would be really nice if there was also a text to video variant of this model, the prompt understanding on this model is incredibly good, most of the t2v models I've tried aren't nearly this good.
Note: You need to update the comfyui to the latest version, or you will get error when loding the v8-v1 checkpoint
Yes. Min version is stated in "about this version"
this one seems to have build in lighting lora, but when i use lighting lora with it i sometimes get much much more interesting result
No GGUF?
You made the same comment for Smooth Mix Wan 2.2 and deleted it after 1 minute
https://civitai.com/models/2190659
https://huggingface.co/Bedovyy/dasiwaWAN22I2V14B-GGUF
enough?
@qek Yeah after reading I saw the link for smooth mix GGUF. Thanks for these links for this. Merry Christmas :)
No "WanVideo Sampler" support means I won't be using this one :(
Well, or you could write the creator of wan video sampler to support mixed precision, like the normal ksampler does. I can not know if any custom node support less than just the standard sampler and I can not change the checkpoint at this moment.
It's seems Kijai never fixed the problem with loading fp8 models
I'll probably never test/support it officially.
Because this custom node is a mess with 1.1K open issues, more VRAM consumption and a bunch of unnecessary extra nodes needed.
BTW - I tested with WanVideoWrapper/Sampler and TastySin-v8.1 - It is compatible with it. It produces just fine videos. The set-up is convoluted, but completely possible.
I have tried the TastySin GGUF, MidnightFlirt and LureNoir and I can say the best camera movements came from LureNoir and most facial consistency from Tasty Sin GGUF, given the same seed and prompts. Given a chance, I'd want to keep LureNoir and TastySin and I'd probably skip Midnightfirt to save SSD space.
Looking forward to the Tasty Sin V8-1 (FP8) to see if its camera movements better LureNoir.
Excellent work DaSiWa, I do think NSFW Enhanced Q8 and your checkpoints are true rivals.
Thanks for the great model! Is it possible to get an SFW version?
You can just use it for SFW content no problem, nobody is forcing you to do NSFW content.
@darksidewalker In my opinion, different NSFW Loras that were merged into the model will harm SWF requests. Correct me if I'm wrong.
@bbx749 I don't think so.
@darksidewalker I would gladly agree if my sfw characters didn't have semen dripping from their mouths so often.
@bbx749 I do not right get your point. If it does not work for you why you are not just using basic WAN22 or any other checkpoint that fits your needs? I have no problem to get SFW with correct prompting.
@darksidewalker For some reason, your model works much better than anything I've tried before! That's why I'm your fan. Perhaps you could publish a list of the sfw loras you used in it?
@bbx749 I did publish a complete resource list inside the announcement.
Hello!
This is the second time I've commented on one of your posts with my beginner's questions.
I've been using your MidnightFlirt model and I love the results I've achieved. So I'm very grateful to you for that work.
I have a question. What are the differences in terms of the results generated between this model in .Safetensor format and the other model in .GGUF format? Or do they only differ in terms of configuration?
I read something in your Usage Guide - Definitive Edition, but it wasn't very clear to me.
Thanks!
You could read my announcement made for this model, too.https://civitai.com/articles/23271/release-of-dasiwa-wan-22-i2v-tastysin-lightspeed-or-gguf-or-safetensors
Safetensor is just the better format, faster, better memory management and in this case a bit smarter.
The results should not differ very much.
Would you consider making a T2V version?
Not planned atm.
Reckon you would boss it.
@smokeymcpott158455 There are other things to come, I have limited time^^
If you use wan2gp, any wan model you can use no input image and it defaults a black frame first frame and generates video same as if t2v.
Details
Files
DasiwaWAN22I2V14BLightspeed_tastysinHighV8.safetensors
Mirrors
DasiwaWAN22I2V14BLightspeed_tastysinHighV81.safetensors
DasiwaWAN22I2V14BV8V1_tastysinHighV81.safetensors
dasiwaWAN22I2V14B_tastysinHigh.safetensors
Dasiwan2.2_Hi_V81.safetensors
DasiwaWAN22I2V14BV8V1_tastysinHighV81.safetensors
DasiwaWAN22I2V14BLightspeed_tastysinHighV8.safetensors
DasiwaWAN22I2V14BV8V1_tastysinHighV81.safetensors
DasiwaWAN22I2V14BV8V1_tastysinHighV81.safetensors
DasiwaWAN22I2V14BLightspeed_tastysinHighV81.safetensors
DasiwaWAN22I2V14BLightspeed_tastysinHighV8.safetensors