🗡️💀 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 (FP8 base) precision
🌟 CFGZeroStar patch (better results and prompt adherence)
🍰 Baked with latest VAE (2.1) BF16
🍀Optimised additional 22+ Parameters
🫥 lowered hallucinations compared to "CheekyDream"
🎯 Slightly more motion
🩻 Known issues
🫦 Addicting
FAQ
Comments (103)
the problems I found: 1) if semen is described in the prompt, it will definitely pour out of the character's mouth in the video 2) it is impossible to make the characters move (up and down, back and forth) in the sex scenes 3) the character constantly opens his mouth as if talking (in the posing scenes).
most likely, I have an error either in the settings or in the prompt. I would love to study the instructions or material for NSFW models based on wan 2.2, but they don't exist...
That's seems off to me, I posted so much examples and new posts, none of the describes things happens to me. In rare cases there is dripping semen from places, like mouth where it should not be, that's most likely a hallucination coming from the merged NSFW content. This is fixable by ether using better/other descriptions or multiple samples. Basic WAN 2.2 also have many of this things happening, even more on too high resolutions or too low.
The movements are all well in my samples. I think there is something off with your configuration. The talking thing was always a thing in all WAN checkpoints even more on WAN 2.1, but I found it far less in this one.
@darksidewalker your posts are really impressive. can you tell me how you generate prompts to create a video from an image? and which values of the init image creativity setting do you consider the most effective?
@seawolf338 At some point I may write a guide, but usually I just write 1 sentence what is the situation+appearance of the main char and than what should happen in 2-5 sentences. Init image creativity must be set to 0 (zero), that's a basic rule of i2v. If you set this to any other value it will blur and deform the start-image. On most images you have that "(i)" icon at the right bottom corner, there you can see the settings someone used, if that person did include them.
@darksidewalker I think the guide would definitely be useful, considering that this wan topic has just been born
@darksidewalker I don’t remember the creativity settings of the Init image in wan2.2
"if semen is described in the prompt, it will definitely pour out of the character's mouth in the video" -try adding a blowjob lora and setting it to -o.xx This helped me some.
@throwaway148 That's a smart idea to use a lora! 🤔
@saymannsk330 They are always 0. As for all i2v models.
im quiet new to this i need help ... how to use this model ? is this LORA ?
which workflow to use this in comfyui or workflow in swarmui?
This is a full checkpoint (mind the size), and "workflows" normally refer to comfyui, while "presets" refer to swarmui. Thats said, swarmui uses comfy as backend and can also use workflows.
@darksidewalker thanks finally i can use it .. but i found this model is really hard to mix with a lora ? why ? i already using the High Lora and Low Lora . and still the result is not as i expected .
any advice for me ? thanks in advance
How is the new model faster? I went from 27 minutes to: Prompt executed in 00:10:59. Fantastic! 15/5 stars!
maybe you were running out of memory before
@Light2020462 I don't think so, but maybe. I have a 4090 24GB and 64GB of system RAM. I haven't changed rez or frames, so I wouldn't think that would be the case.
27m are standard on WAN 2.2base without speed up, so it should be the speed up that's boosting the times.
@darksidewalker I was using your cheeky dream model before, but this new one seems to be twice as fast. Is that feature or am I special?
@DarthRidiculous That's should not be a special case, all models have the speed up, maybe it is something related to the VAE optimisation? Maybe you changed some settings? But 10m for 81 frames with 720p seems normal with speed up and 27m without any speed up for basic WAN 2.2
@darksidewalker Nope. I didn't change any settings and I am doing 144 frames per phase with 4 phases.
Dear Darksidewalker
I have been using your sweet spot pair since you released it. and I have no problems. none.
but since then, I have not checked the wan checkpoint list in my filter since I was working on my own lora concoctions. upon filtering the model list to wan, I see you added 3 more Pairs.
what are the differences among the pairs?
ps you have influenced me to download your illustrious model 😂
Hello mate!
I really love to hear, that you are having fun with my model. Thats the whole purpose I created it for, be a engaging jet easy way to enjoy WAN 2.2 with less restrictions.
The differences are listed inside the version details on the right side of the page. In summery, I just refined and polished it with every iteration.
I also hope you can enjoy my Illustrious model and if you mind, I would love to see some posts!
The Illustrious models are not only particular iterations, they achieve different styles as well.
Have fun!
thank you! cheeky dream is queued in my JD2! @darksidewalker
If you refer to the WAN 2.2 process, yes it should.
@darksidewalker https://skrinshoter.ru/sXzAYRcPfcc
@Renessance As you can see in the examples. Everything.
I won't elaborate the link/workflow, sorry. I use SwarmUI and it works. Basically it is a WAN 2.2 checkpoint.
@darksidewalker Got it, thanks for the answers.
Sorry if this is a stupid set of questions, but it should be quick to answer. I am going with what you're suggesting to everyone, and though I'm quite used to, and overall more familiar with ComfyUI, I am using Swarm.
That said, I just simply want to know two things.
1. How do I load both the High & Low models at the same time, as I've read in other comments it's needed, which makes some obvious sense. ((I keep getting static, save for two times, once when it generated only a picture, the second did the video with my exact Image to video, but it was horribly low quality.))
2. Is more of just a simple yes or no, is "Init Image" required, or will it function simply with image to video? I believe my actual video result had "Init Image" on with the same picture I used for the image to video section.
Thank you for the model, and forgive me, I'm quite new to video generating locally.
Hi!
1) At SwarmUI on the left side there is a menu called "Image To Video", expand this and check the box "Display Advanced Options?", there will now be 2 separate "Video Model" and "Video Swap Model" boxes. On Video should be the HIGH and on Swap the LOW model.
2) An Init-Image is always required for all I2V models.
3) The models will be uses after each other, not the same time and 50% steps should be used for HIGH and 50% steps for LOW, this results in using the option "Video Swap Percent" with 0.5
4) In this "Image To Video" section are also the total step count and CFG to set for the video
-> Hint: The SwarmUI git has a very good "docs" section where everything is explained in detail.
@darksidewalker Thanks, friend! I think I've got mostly everything set up properly. I'll reply again later to let you know if I got it working properly. If I come into any issues, I'll try to figure it out myself before bothering you with more questions, as I can see you've been answering quite a bit of them non-stop.
@darksidewalker Update: Damn, dude! I got it working without a hitch. (Had to do some paging on my OS drive as well) but holy smokes, the quality of this model is impeccable, I mean, absolutely fantastic. It's like magic. Thank you VERY much for your upload, work and help.
@AddyRahl Glad to hear you managed to get it working! So have fun and if you mind, post some nice stuff :)
Does it matter if the file type is .gguf or not?
Sure, they are completely different checkpoint format's. This is not a gguf.
Dude you update model so fast that i I don't have time to test it at full 😂😂😂
But more serious question in new model you say that new Vae was baked. What Vae and this mean that i dont use wan2.1 vae as default?
The thing with safetensor is you have to use a VAE when saving the checkpoint as reference, but the VAE is not hard-included like in gguf, or at least you can alter it afterwards. I decided to compile it with the latest 16-bit VAE for reference, to optimise the checkpoint a bit.
You are free to use any other VAE with it (2.1). Most likely all VAE 2.2 will result in wrong dimensions and throw an error.
@darksidewalker What Vae you using when generate video?
@KeMiliUs WAN 2.1 VAE, as for all WAN 2.2 14B models
The post of this model seems awesome,, I can't wait to try it！！
Now this model is suppoted in Wan2GP!
What do you mean?
@darksidewalker
Well, FP8 support has been added to Wan2GP. But FP8s are also different. I wrote to the author of Wan2GP on git and sent him links to the high and low models. he updated the engine and now the model works there. https://github.com/deepbeepmeep/Wan2GP/issues/1020
@saymannsk330 I see :)
Was the TeasingKiss model updated today? I downloaded a few days ago should I redownload? what were the changes?
Released today. Its an upgrade. Changes are documented on the page~
@darksidewalker oh right lol, Forgot it was early access. Love the model. Thanks!
Hi im just wondering if this would work in comfyui or should I switch to swarm? Thank you
I am running the both high and low checkpoints in comfyui with no issue. I downloaded his comfyui workflow found in the "suggested resources".
@rondayne09535 cool thanks for the response. I should have checked the suggested resources before posting. I appreciate it.
Its working very well at ComfyUI
I downloaded cheeky dream and was going to use it while I waited for donations to be over for teasing kiss. but When I checked Teasing kiss was paid for already so I got that one.
I ran 3 prompts that I always had problems with previously - 2 non-sexual and 1 sexual.
I ran all with 81 frames because by then it was 1am and I wanted to see what it does quick. and was it was quick?! couldnt even watch through 3 full tiktoks
but I encountered a problem with each prompt. and this is a ME problem, not your model problem.
I, being me, use KJ load diffusion model load with Sage attention int4 cuda fp16.
I, being me, use patch Sage attention with int4 cuda fp16. essentially doing it twice
Also me, used Teacache after sage attention, and for 16gb vram with offload, its not required.
what I found is that teacache kicks in on first run after starting comfy. instead of 4 steps, it does 3. but the video generated is perfect. - takes 120 seconds. without teacache it takes around 3 minutes.
but I think because of so many sage attention patching, aswell as teacache - the issue Im having is that when I load a new 1st frame, the latent from the previous video blends with the new latent. i think teacache more because cache isnt purging after generation is over? so Im getting a noisy video with the old first frame. restarting comfy fixes this.
I will run with teacache bypassed and let you know, and do more frames and more steps without rope and see how far I can push without oom.
4060 ti oc 16gb btw
there is a chance (or an huggingface repo) that this model will ever be quantized? meybe a q5 or q6?
This will likely not happen any time soon, because it is already quantized to fp8 and a direct conversion to gguf is not possible as far as I know. My resources are not enough to convert the full fp32 model with all the steps involved to a gguf. If I could find a way to do it, I would.
@darksidewalker understandable, it is a pity... there is any chance that the full fp32 model is shared somewhere? maybe in the near future (when my job let me have some time) i'll quantize it with some cloud resources (giving you credit and the quantize files to be published under your name)... or if not me i think that if it is shared someone with more time than me could do it sooner.
@Visnis The problem is I do not use a full fp32 model to create this, so there is none. That would be the checkpoint from WANAI itself.
@darksidewalker ok, sorry, i missed that part, reading again your first comment it was clear. it is a pity because yours is one of the most interesting checkpoints for wan, a gguf version would be great... hope that someday in the future it will happen ^^
Your hardware and settings are your limit. This could help: https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition
I did a 8 seconds video just fine: https://civitai.com/posts/23351286
I don't think it's a hardware limitation can generate 15-second videos. That's not the problem. The only problem is that the action repeats
@nilsb812140 Prompt to short for the given time :) It starts to guess.
turn off your "pingpong" option in "video combine"
set "loop" to "0"
you used some "first frame to last frame" lora maybe
@Sylvanlonay I use the comfiui workflow without lora. i cant find a "pingpong" option
wait you can go over the 5 sec limit with this checkpoint without sliding windows?
@SolidBold It depends on your specs, but sure, did it till 9 seconds on 16GB VRAM, just for testing.
i have 5060ti 16gb and 32gb d5 ram, but can't run the model through the low noise model loaded, is it the the gpu or ram issues?
@darksidewalker im running comfyui on a docker,
Adding virtual RAM (swap, pagefile, zRAM)
i think the swap part and zRAM part will help, may test and result later, thank for your ultimate guide.
are these i2v or t2v lora's?
You could read the headline and file details. This will answer your question! I believe in you!
@darksidewalker somehow i managed to miss such subtle clues, my bad. sorry for that
every time i try to run the model comfi gives me the same error "shape '[1, 16, 40, 2, 40, 2]' is invalid for input of size 2150400" there is a vorflow somwere ? thanks!!!!!
I suggest reading this carefully: https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition
Short answer, probably either image resolution or wrong vae.
I had the same issue and figured it had to do with the vae (I was using a WAN 2.2 vae instead of using a 2.1)
@NeyoAlt Exactly as written in the article and the description of the checkpoint.
@darksidewalker tank you !
Hey, how can i run this model?
I mean, i saw "On 16GB VRAM, 64GB RAM, 4 steps, cfg 1, 81 frames", but the high + low models size are much larger then VRAM that you have. Is it really worth it? I think when model is forced to use RAM, the generation time is starting to get really long. Can you please explain to me? I have 5060 ti 16gb VRAM, and 48gb RAM.
I'm fully readed the "https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition", but still quite cannot understand how this works. Are people really doing this? Is this fine? Using the not quantized model which size of are much larger then the actual VRAM gb that you have. I always download quantized models so they cannot fill up my VRAM use to full 100%.
@Hell_Yeah I think you misunderstood something. Full quant of WAN 2.2 would be FP32 and never fit into any consumer GPU atm. FP8 is scaled and fits on 16GB VRAM, 64 GB RAM. Also, as I wrote, it is always only 1 model used at a time.
@darksidewalker Ah, you mean that when using high model, it will use 14~ gb of VRAM at the time, and after that when low model will go, it will also use 14~ gb of VRAM, and VRAM usage of previous high model will clear itself? I'm using cursed workflow with wan 2.2, and maybe i'm really kinda silly. I'll check the size of my models, and try yours after that. (English is my second language, sorry for misunderstanding <3)
@darksidewalker Checked my models, both Q4 wan2.2, are 8gb~, so when i start the process in the ComfyUI, models are loading at the same time. 8+8=16, exactly how much VRAM i have atm. I think with ComfyUI i'll just get OOM error if i try using your model. I mean, i just don't know how to make ComfyUI use only 1 model at the time, unloading the other one. It's bit sad to be honest... If you can help, i would love to hear your thoughts, otherwise, thank you for the feedback
@Hell_Yeah I am using this model with only 12 gb vram. Check workflow from this post. Download video and drag and drop it into ComfyUI for import. https://civitai.com/posts/23366027
@Discocat Your WF is not good if it loads all models at the same time, thats not the point of MoE. Also q4 is for cards with less than <6/8GB VRAM. You should use another WF for sure.
@darksidewalker OMG! I CAN REALLY USE Q6 MODELS AND ETC NORMALLY! I was so silly when i thought that i need to use 8+8=16 to fulfill my GPU requierments. Thank you both guys!!! New generations are looks much much better now! You're opened the whole new world to me, thank you!!! I'll go try teasingkiss models right now
@Hell_Yeah You could try my updated WF (https://civitai.com/models/1823089?modelVersionId=2310177) should work well with my model and any other WAN 2.2
How can I use this checkpoint to maintain consistency in the shape of the mouth or eyes?
How can you not?
So, if any of the down-voter's can provide a good answer of this question, go ahead. I do not have an idea why the face is inconsistent, all my generations are with consistent face details.
Why is everyone so serious ...
@darksidewalker Every time I do i2v with this checkpoint, even in states where I don't want the mouth to move, the mouth keeps moving as if it's talking, and I want to stop that. Is there a prompt or helpful LoRA for doing so?? No matter what I do, the mouth moves and I can't get the desired video. I need help
@kimrude0940912 I can not much help here. If you do face close-ups WAN 2.2 speed-ups tend to talk, if you do not write long prompts for other things and every aspect of the face. There is no lora that would change that known to me.
Turn off any other speed up lora... i had this issue
@FollowTheWhiteRabbit That's mentioned in the description. I assume he read that before.
@darksidewalker Even though I'm not using LoRA at all and only using the Dasiwa checkpoint, that problem still occurs. I'm not using the speed enhancement LoRA either. It's so sad... Boo hoo..
@kimrude0940912 Maybe read my article about wan 2.2 and figure out what could help you? Just a guess.
This is cool and all but how come the loras don't show up anymore, i'd like to use loras with this workflow but they no longer show up on wan2gp, maybe i should clarify, i understand this is suppose to substitute loras, but it completely eliminates the ability to use one or 2 loras in addition to it ?
The checkpoint does not disable Lora usage, that must be a problem with your app.
@darksidewalker figured it out, for some reason it only detects loras in the "Lora" folder and none from the "loras_i2v" folder. Weird.
Any tips on getting faster motion? Like for a dancing scene; a lot of times they seem to be moving slowly.
Reducing the resolution of high noise points can improve image responsiveness but will lower the final image quality. Additionally, increasing the Sigma Shift value may speed up image responsiveness. It is uncertain whether adjusting the cfg value will accelerate image responsiveness, but you can try increasing it to see.
Unfortunately that's a most common problem with all WAN 2.2 speed-up's. You can also use a dedicated dancing lora or use plain WAN 2.2 without any speed-up lora/checkpoint. This is most likely a problem from the distillation process of speed-up's.
As mentioned from @MXQHXQ , this could help too, but will degrade quality. Also increasing CFG is like disabling the speed-up's and will result in raised generation times, therefore you go the plain WAN 2.2 route there.
you can increase fps per second when saving after interpolation