⚠️ Important information
All models already include the Lightning LoRAs, except the SVI models, where Lightning is not included.
Do not use additional Lightning LoRAs on models that already have Lightning integrated, or the quality will be degraded.
For SVI models, you can use Lightning LoRAs if you want faster video generation.
2or3 ksampler (for svi) https://civarchive.com/models/2079192?modelVersionId=2668801
v2.1 (for nsfw v2) https://civarchive.com/models/2079192?modelVersionId=2562360
v2.1 with MMAUDIO (for nsfw v2) by @huchukato https://civarchive.com/models/2320999?modelVersionId=2613591
Another wf triple KSampler KSampler https://civarchive.com/models/1866565/wan22-continuous-generation-svi2-pro-or-gguf-or-32-phase-or-upscaleinterpolate-w-subgraphs-and-bus?modelVersionId=2559451
triple KSampler wf setup allows for more motion and helps prevent slow-motion issues. In exchange, your videos will take longer to generate.
For those having issues with my SVI workflow , you can try Kijai's wf here: https://github.com/user-attachments/files/24364598/Wan.-.2.2.SVI.Pro.-.Loop.native.json. Alternatively, you can try fmlf wf https://github.com/wallen0322/ComfyUI-Wan22FMLF/tree/main/example_workflows they will be simpler. There are others on Civitai that work very well too.
Qwen-VL workflow as an alternative to Grok for creating your dynamic NSFW prompts. Thanks @huchukato for his work: https://civarchive.com/models/2320999?modelVersionId=2611094
🟣 SVI Update – NSFW
⚡ Model Presentation (SVI-compatible version)
This update was made because the NSFW V2 models were not fully compatible with SVI LoRAs.
This version was created to work smoothly with SVI, while still functioning without them (though the workflow must be adapted).
Be careful, SVI LoRAs will only work with a workflow specifically designed for them, otherwise, it won't work.
There are two models available for SVI:
✔ Fast Move (FM) – Sexual scenes may differ from the Consistent Face model and will generally be faster.
✔ Consistent Face (CF) – Slightly better image quality, which may be preferable for anime-style videos; sexual scenes differ from Fast Move, but the difference is minimal.
You can also mix models between High and Low LoRAs:
FM (Fast Move) as High + CF (Consistent Face) as Low
CF (Consistent Face) as High + FM (Fast Move) as Low
Both combinations work and give slightly different results, offering more flexibility for your videos.
For this version, the main improvements include:
✔ Fully adapted for SVI LoRAs
✔ Greater flexibility: Lightning and SVI LoRAs must be loaded manually for custom workflows
🟣 SVI LoRAs – Strengths & Weaknesses
⚡ Overview
✅ Strengths
Best solution for making long videos
Excellent transitions between video segments
Reduced degradation compared to other solutions
Strong character coherence: the model retains information from the previous video, helping maintain consistency
⚠️ Weaknesses
Weaker prompt understanding
Weaker camera understanding
Videos are less dynamic
Sometimes slow-motion effect
(can be improved with proper Lightning LoRAs, dynamic prompts or triple ksampler)
🟣 SVI LoRAs – Download Links
⚡ Download
Note: Both LoRAs must be loaded manually in your workflow.
🟣 Suggested Lightning LoRA Combos (Optional)
⚡ Overview
You don’t have to use these Lightning LoRA combos. They are optional and allow you to fine-tune motion and degradation.
You can also use other Lightning LoRAs or assign different combos per video for more control.
🔥 Combo 1 – More Motion (Rapid Video Degradation)
High LoRA →
https://huggingface.co/Kijai/WanVideo_comfy/blob/709844db75d2e15582cf204e9a0b5e12b23a35dd/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
Weight: 4
or
💜 Combo 2 – Less Image Degradation
or
or
🟢 Combo 3 – Balanced Motion / Moderate Degradation
High LoRA →
https://huggingface.co/Kijai/WanVideo_comfy/blob/709844db75d2e15582cf204e9a0b5e12b23a35dd/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
Weight: 3
Low LoRA →
https://huggingface.co/Kijai/WanVideo_comfy/blob/709844db75d2e15582cf204e9a0b5e12b23a35dd/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
Weight: 1.5
🧠 Advanced Usage Tip
You can disable global Lightning LoRAs (in my wf) and assign different combos per video:
Combo 1 for Video 1, Combo 3 for Video 2, Combo 2 for Video 3.
Each combo produces different motion and degradation behavior.
If you want to create several-minute-long videos while maintaining high quality, it is possible, but it will take a very long time. You just need to avoid using Lightning LoRAs and use the full model instead.
🧠 Dynamic Prompts – Better Control & More Motion
You can use dynamic prompts to have better control and this helps to make the video more dynamic.
You just need to give this example prompt to an LLM like ChatGPT. It will be enough for it to describe the image you have and what you want as a video while keeping the same structure of the prompt in the following example.
⚠️ This will be a NSFW prompt; ChatGPT will not accept it.
You can use GROK https://grok.com, which can make NSFW prompt modifications.
For more examples of prompts with different poses, check:
Enhanced FP8 Model.
Give these prompts to GLM with the dynamic prompt structure if you want.
Example 1
(At 0 seconds: Wide shot showing a slightly overweight man casually walking down a city street, camera fixed in front, urban environment with buildings and cars.)
(At 1 second: Suddenly, a massive shark bursts from the pavement ahead, looking terrifying at first, pavement cracking, dust and debris flying, camera from side angle.)
(At 2 seconds: Medium shot from the side, the man stumbles backward in shock, while the shark dramatically slows down and strikes a comically exaggerated sexy pose, revealing large, exaggerated shark breasts, covered by a colorful bikini.)
(At 3 seconds: Close-up on the man’s face, eyes wide in disbelief, as he turns to look at the shark, small cartoon-style hearts floating above his head to emphasize his amazement, camera slightly low-angle.)
(At 4 seconds: Dynamic travelling shot showing the man frozen in the street, the shark maintaining its sexy pose, water splashes and debris still moving realistically, urban chaos around.)
(At 5 seconds: Wide cinematic shot pulling back, showing the man standing in the street, staring at the bikini-wearing shark with hearts above his head, epic perspective highlighting absurdity and humor.)
Example 2 – Anime NSFW
(At 0 seconds: The couple in a cozy bedroom, anime style, soft lighting highlighting their intimate embrace, her back arched slightly as he positions himself.)
(At 1 second: The man’s hips moving rhythmically, the head of his penis sliding effortlessly into her vagina, her body responding with a gentle, fluid motion, anime-style motion lines emphasizing the smooth penetration.)
(At 2 seconds: Her back arching deeply against him to intensify the pleasure, hips swaying with each thrust, breasts bouncing subtly, small hearts floating around them to capture the erotic energy.)
(At 3 seconds: Her face, eyes closed in bliss, a soft moan escaping, hands resting behind her head, anime-style blush on her cheeks, the air filled with a seductive aura.)
(At 4 seconds: The man penetrating her deeply, her body moving in sync with his, the bed sheets slightly rumpled, the room’s warm lighting enhancing the intimate, lustful atmosphere.)
(At 5 seconds: The couple locked in a passionate embrace, the scene exuding vibrant, seductive energy, anime style with smooth lines and soft shadows.)
🟣 Lightning Edition – NSFW I2V V2
⚡ Model Presentation (2 new versions available)
I originally planned to release only one V2, but some people preferred the NSFW V1 over the Fast Move V1 version, so depending on what you’re looking for, one version may suit you better than the other.
For these V2 versions, I tried a new approach:
✔ I made sure that most sexual poses work, while the model is also good for SFW content
✔ More flexible for general use
🔥 NSFW Fast Move V2
Improvements included in this version:
Better prompt understanding
Better camera understanding
Reduced unnecessary back-and-forth movements outside sexual poses
(cannot be completely removed, but strongly reduced)Improved bounce effect on the buttocks
If a man appears, he will no longer automatically attempt to penetrate the woman when she is nude
This version is designed for those who want more dynamic scenes with more movement.
💜 NSFW V2
The difference between this version and NSFW Fast Move V2:
Less camera control
Less camera understanding
But body movements are less pronounced (breasts and buttocks)
Some preferred V1 NSFW to V1 NSFW Fast Move, and this version keeps that spirit
For varied sexual poses, check the previews — there are many.
You can use the shown prompts and adapt them to your images, but of course, other prompts will work as well.
Don’t hesitate to use other LoRAs for creating specific concepts.
🔧 Recommended Settings
Steps: 2+2
(Jellai recommends 2+3 for even better results, and I agree)Sampler: Euler simple
CFG: 1
This model already includes the Lightning LoRAs, don’t use them, or the quality will be degraded
You need to download both models: H for High and L for Low.
Here is an example made with the base workflow + the First/Last Frame workflow + upscaler (only on the starting images) + inpainting (cum).
https://civarchive.com/images/114733916
🔞 LoRA Examples (with links and recommended weights)
LoRA: Penis Insert WAN 2.2
LoRA weight: 1
Face → Doggy
https://civarchive.com/images/112864483Face → Missionary
https://civarchive.com/images/112864718
https://civarchive.com/images/112864835Face → Doggy (leg aside)
https://civarchive.com/images/112864885Another example
https://civarchive.com/images/112864972
Other examples with different LoRAs
Face → Reverse Cowgirl — LoRA weight: 0.7
https://civarchive.com/images/112865189Face → Cowgirl — LoRA weight: 0.5
Face → Missionary — LoRA weight: 0.3
https://civarchive.com/images/112865381Face → Missionary — LoRA weight: 0.5
Face → Doggy leg aside — LoRA weight: 0.6
https://civarchive.com/images/112865551Face → Doggy — LoRA weight: 0.7
https://civarchive.com/images/112865696Face → Spoon — LoRA weight: 0.7
https://civarchive.com/images/112868124
Recommended LoRA to obtain a very realistic vagina and anus:
https://civarchive.com/models/2217653?modelVersionId=2496754
💡 Tip for Anime Style:
If you’re working in an anime style, feel free to follow the advice of @g1263495582 thanks to him for this.
Try these LoRAs together or separately; they help maintain face consistency:
🔹 Note:
For these LoRAs, use only Low Noise.
For the examples mentioned above, use High and Low Noise as indicated.
LoRA weight: 0.3
Many things can be done with the model, but don’t hesitate to use other LoRAs for specific purposes. And don’t hesitate to lower the LoRA strength to preserve the face as much as possible.
📷 Dynamic prompt example with different camera angles:
(at 0 seconds: wide shot showing the woman standing in the snowy plain, a massive giant dragon emerging behind her, snow cracking and dust rising).
(at 1 second: the woman jumps backward onto the dragon’s back as it bursts fully from the sky, camera tracking the motion from a side angle, debris and snow flying).
(at 2 seconds: medium shot from the side, the woman balances heroically on the dragon’s back as it begins to run forward across the snowy plain, slow-motion on her posture).
(at 3 seconds: close-up on the woman’s determined face, camera slightly low-angle to emphasize her heroic stance, snow and debris flying around).
(at 4 seconds: dynamic travelling shot alongside the dragon, showing the snowy plain, scattered debris and ice fragments flying everywhere as it gains speed).
(at 5 seconds: wide cinematic shot pulling back, showing the dragon taking off with the woman riding on its back, soaring above the snowy plain, epic perspective with snow, wind, and scale emphasizing the drama).
For available camera angles, check further down in the “cam V2” model description.
🔞 Normal Clip vs NSFW Clip:
You can also use the NSFW version for your clip.
It can bring positive effects for sexual scenes, but it can also cause issues, as in this example:
NSFW Clip:
https://civarchive.com/images/112864204Normal Clip (same seed):
https://civarchive.com/images/112864295
Here are the links to NSFW clips (thanks to zoot_allure855 for correcting the BF16 version):
BF16 fixed version:
https://huggingface.co/zootkitty/nsfw_wan_umt5-xxl_bf16_fixed/tree/mainFP8 version:
https://huggingface.co/NSFW-API/NSFW-Wan-UMT5-XXL/tree/main?not-for-all-audiences=true
❓ If you have any questions, don’t hesitate to ask!
Many people ask me for help via private message. You can do that, no problem, but I would appreciate it if you could do it in the comment section, it could help other people. Thank you.
🟣 Lightning Edition – NSFW I2V camera prompt adherence
⚡ Model Presentation (2 versions available)
I decided to release 2 versions. Both can produce different results, as you can see in the previews (the seed was the same).
Fast Move Version: provides more movement, and the movements will be faster, with better prompt understanding and camera handling
(you can see it in the 4th preview "fast move high" where the man slaps the woman)Natural Motion Version: offers more natural breast movements depending on the situation and produces slower scenes.
👉 Check both and choose the one that works best for you.
This NSFW edition is, of course, focused on sexual poses.
You should achieve very good results.
To create specific concepts, feel free to use specific LoRAs — they work very well with this model.
🔧 Recommended Settings
Steps: 2+2
(Jellai recommends 2+3 for even better results, and I agree)Sampler: Euler simple
CFG: 1
This model already includes the Lightning LoRAs, don’t use them, or the quality will be degraded
You can also use this T5, which may improve understanding:
https://huggingface.co/NSFW-API/NSFW-Wan-UMT5-XXL/tree/main?not-for-all-audiences=true
📝 Usage Tips
Start with a clean, high-resolution image for better results.
The model will not change the face; if it does, increase the resolution.
Also adjust your prompt if you are not getting what you want.
Wan understands some terms well, but not all.
For prompts, check the video previews: adapt them to your image.
Other prompts will work too.
🎥 Training
I trained 2 LoRAs:
The first one using several videos and images of different sexual positions.
The second one to bring more dynamic motion.
Respect to everyone who creates this type of LoRA — it requires a lot of work.
🙏 Credits
This model wouldn’t exist without the incredible work of these creators:
alcaitiff : https://civarchive.com/models/1295758/nsfw-fluxorwan-22orqwen-mystic-xxx?modelVersionId=2300332
Sweet_Pixeline : https://civarchive.com/models/1844313/penis-play-wan-22
anonimoose : https://civarchive.com/models/2008663/slop-twerk-wan-22-i2v
dtwr434 https://civarchive.com/models/1331682?modelVersionId=2098405
A special thanks to Alcaitiff and CubeyAI, two very kind and humble people.
🔔 Important
Please don’t support my work with buzzes here, I don’t need it.
If you want to support someone, support the creators listed above — they truly deserve it.
💬 Feedback
Feel free to give me feedback, positive or negative, to help improve future updates.
Update WAN 2.2 V2 CAM I2V – NEW feature Camera & Prompt Improvements
This custom version of WAN 2.2 I2V has been updated to deliver better prompt comprehension and improved handling of camera angles and cinematic movements. It provides more accurate scene interpretation, smoother transitions, and enhanced control over dynamic.
Key Features:
Excellent understanding of prompts and scene composition.
Supports various camera angles and movements, including zoom, dolly, pan, tilt, orbit, tracking, and handheld shots.
Ideal for cinematic storytelling, animated sequences, and creative video-to-image projects.
Flexible multi-step prompts, standard 4 steps (2+2), can be increased for higher fidelity.
Recommended sampler: Euler simple.
You can, of course, use your usual prompts; this is just one example among many.
Example Prompt with Different Camera Angles
(at 0 seconds: wide frontal shot of a man standing in front of an open fridge, cinematic lighting, subtle ambient kitchen reflections, the fridge contents visible, camera static).
(at 1 second: medium shot from the front as he opens the fridge fully, reaches for a can, slight zoom-in to emphasize the action, cinematic framing).
(at 2 seconds: camera shifts to a side medium shot, tracking him as he lifts the can to his mouth, fluid movement, maintaining lighting and reflections).
(at 3 seconds: camera starts a smooth 360-degree orbit around the man, following him as he drinks from the can, motion fluid, background slightly blurred for cinematic effect).
(at 4 seconds: close-up on his face and upper body while drinking, orbit continues subtly, fridge reflections accentuating realism, cinematic polish).
(at 5 seconds: final wide shot as he lowers the can, camera completes orbit to original angle, showcasing the kitchen space, lighting, and dynamic movement).
Available Camera Movements
Zoom / Dolly
zoom in
zoom out
camera zooms in on subject
camera zooms out gradually
dolly in
dolly out
camera dollies in slowly
camera dollies out steadily
crash zoom
Pan
pan left
pan right
camera pans across the scene
gentle pan left
sweeping pan right
Tilt
tilt up
tilt down
camera tilts up to reveal…
camera tilts down from…
Orbital / Tracking / Arc / Rotation
orbit around subject
360° orbit
camera circles around
tracking shot
camera tracks alongside subject
arc shot
curved camera movement
Other Movements & Styles
static camera / static shot
handheld shot
camera roll
Note:
LoRAs work perfectly with this model, offering full compatibility and consistent results across styles and concepts.
# SUPPLEMENTARY ADVICE
You can use negative prompts, but be careful: this will double the generation time.
Only use them if you really want to prevent something from appearing in your video.
In that case, enable the corresponding node; otherwise, keep it disabled.
⚠️ Important
The model must be used with CFG set to 1, so negative prompts do not work by default.
However, there is a simple way to enable them.
How to enable negative prompts:
Open the Manager
Search for kjnode and install it
In your workflow, add the WAN Nag node
To use it correctly:
Connect this node after the LoRA Loader
Feed it with the negative prompt
Then connect it to the first kSampler (High)
👉 Only use this option when necessary, to avoid unnecessarily increasing generation time.
And here is the negative prompt for unwanted movements:motion artifacts, animation artifacts, movement blur, motion distortion, dynamic distortion, shifting shapes, unstable render, instability, wobbling effect, jittering effect, vibrating render, inaccurate details, visual noise, distorted surfaces, rendering errors, warped shapes, exaggerated butt movement, jiggle, overanimated hips, unnatural butt motion, hyper bounce, extreme curves, distorted hips, unnatural pose, unrealistic anatomy, deformed body, disproportionate body, floating limbs, blurry textures, clipping, stretching, low detail, messy background, artifacts, butt bounce, moving hips, swinging hips, shaking butt, wiggling butt, moving lower body, moving pelvis, jiggling buttocks, bouncing butt, unstable stance, unnatural hip motion, exaggerated hip movement, hip sway, hip rotation, bottom motion, pelvis motion, wobbling hips, fidgeting lower body, dancing hips, pelvic movement, motion blur, unnatural movement
And here is the negative prompt from the official ComfyUI workflow:色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走
Note: If you use the default workflow without the node, this negative prompt will not work.
UPTDATE: Lightning Edition – T2V
I’m not really experienced with T2V myself, but a colleague who works with it a lot tested this version and confirmed that it performs very well. From the few tests I managed to do on my side, I got the same impression, although I haven’t compared it directly with the base version yet.
The settings are the same as in the image-to-video version: 2+2 steps or more if you prefer, CFG at 1, and Euler simple (though other samplers also work great).
I don’t plan to make an NSFW version for the T2V release,
but for the I2V version, I’m already quite happy with the first results.
Update WAN 2.2 V1.1 I2V
Updated version of the original Lightning merge — same settings 2+2 steps or more, featuring more movement and smoother flow (depending on the prompt).
The model already works very well in NSFW. Just use the right LoRAs, and the movement will improve.
WAN 2.2 V1 I2V
This checkpoint is based on the original WAN 2.2, with the Lightning WAN 2.2 and Lightning WAN 2.1 LoRAs already integrated. This improves image quality, makes motion smoother and more dynamic, and removes the slow-motion effect that can occur with the Lightning models.
A common setup is to use 2 steps on the high model and 2 steps on the low model, though other settings may work as well. Do not apply the Lightning LoRAs manually — they’re already included in this checkpoint.
My workflow
https://civarchive.com/models/2079192/wan-22-i2v-native-enhanced-lightning-edition
Description
FAQ
Comments (190)
Are these all GGUF?
If people request it, yes, of course! You’re already the first 😉
@taek75799 GGUF please!
@Stawy Second that. Q6 quant please.
Please GGUF! Thxs in advance
Q8 pretty please!!
Yes sir, "the people" demands it XD (please ? Damn... I really need to buy more RAM)
@BretChampagne Haha, thanks for your funny comment! I’ll publish them gladly. In the meantime, don’t forget to buy some RAM 😉
Is the new NSFW I2V only I assume? Based on your samples, I would say yes. Just curious
Yes, exactly, this model is very adult-oriented.
@taek75799 wait, this is I2V? dang.... I was hoping for T2V. This shows up on the wan2.2 text2video option in search. Arg.
@BigSad11 fun fact. If you feed your workflow a blank page it will act like a t2v. Sure...You might lose some frames to initiate, but sometimes it is worth it. Additionally you can merger the white image with 5% of an image like a cartoon scene, it can easily transition into a realistic scene. So versitile....
@Adaptalab0r Right on! Cool stuff. I'm going to try it now. Thanks
The new NSFW models are great, amazing motion! However, both the normal and fast versions are heavily biased towards male on female sex scenes. Anything besides that creates weird issues like Futa on futa for example. The male genitalia will just disappear and create female genitalia every time. I did not have this issue at all with the first NSFW enhanced version that has now been deleted.
Hello, and thank you for your feedback.
I had to make a few concessions for this version; honestly, it’s really difficult to include everything I wanted without introducing some other issues. But I definitely appreciate your feedback, and I’ll check into it when I have the time.
The only advice I can give you, if you don’t mind, is to use LoRAs, maybe lower the strength a bit, to help achieve the result you’re looking for. I hope you’ll be able to get what you want
@taek75799 Thank you for the reply! Unfortunately I tried with specific loras but these models seems very stubborn with male to female only. They are incredible at that but I will have to rely on the previous deleted version as it was responding almost perfectly to anything I wanted to make.
@Cosmicv Glad that you were still able to achieve the results you wanted. I kept the previous version just in case, so if you want it again, don’t hesitate to ask.
Will you be releasing a gguf for the latest nsfw move fast?
That depends on whether you liked the model or not! Haha, just kidding. Yes, I’ll do it so everyone can enjoy it ;)
Appreciate it! TBH my computer can handle the safetensors or the ggur, it's just that when I run comparisons the gguf can finish generating the same video faster than the safetensors.
cant you please split the versions and post them separately? one version that is only focused on camera shots and character consistency without adding any nsfw loras and one version with nsfw loras and release them separately? It starts to get a confusing with all the different models on one page.
One reason for a "clean" version is that if some one want to use a specific nsfw lora it might interact with a nsfw lora you already integrated into your model in a unpredictable way. another reason is, that people just want to use your camera model and dont want any nsfw stuff.
OK, again I'm not sure which is which. Fast Move - yes, obvious but is the "Natural Movement" version the one you call NSFW? And is the latter the same as the NSFW model you released recently (a week ago)? Sorry for my confusion, I just don't want to waste disk space with repeated or mistaken versions. One last request, could you make the download filenames more concise please? BTW - I'm using your week old NSFW model most of the time these days - thanks again.
Hello, and thank you very much for your question, it may help other people as well. Two new NSFW versions were released yesterday; they are brand-new. I removed the previous one because it had an issue.
For the difference between these two versions, please check the description, I hope it gives you a clear answer.
Regarding the name of the downloaded file, I'm really sorry. It definitely doesn’t have that name when I upload it. Civitai automatically renames the files for some unknown reason, and it's impossible to change those names afterward.
@taek75799 OK - so Civitai knows best, they think. As for me, I rename them myself as I download them but that can make it difficult to find the versions on Civitai again should I need to check the best settings. Maybe the downloads should come with a parameters guide along with the model. I use Lora Manager in comfyui_portable but, in my installation, it fails to process checkpoints.
Very good model, can you release a model version without acceleration, because Lightx's new acceleration model is updated very quickly, and the current version has a strong tendency to loop over 81 frames, and the package acceleration model makes any modification impossible. Thank you!
Hello, and thank you for your reply. To be honest, it might be quite complicated to publish it here. The purpose of the model is precisely to integrate the Lightning versions, but not only that — it also includes improved motion and better camera understanding.
However, I can provide it to you through a direct download link if you want?
By the way, regarding the loop, you can avoid it if you want… at least I think so. I’m not sure which workflow you’re using — a long-video workflow with a single prompt?
If that’s the case, after the first generation try modifying the prompt, or switch to a different LoRA.
First time I tried running it I went OOM. 32gb ram, 16gb vram. Do I need gguf or quantized versions?
Was my first time trying wan 14B so I wasn't sure what to expect. Previews look good.
Hello, and thank you for your feedback. You should be able to run it without any problem. Personally, I also have 16 GB of VRAM and 64 GB of RAM, but I know people who only have 12 GB of VRAM and 32 GB of RAM.
An important question: are you using Kijai's WAN Wrapper to run the model?
Another question: what resolution are you using?
Don’t worry, I plan to release the GGUF models, but the FP8 version here should work. We’ll sort out the issue.
@taek75799 Thank you! Very kind of you. I'm not familiar with Kijai's wrapper, is it recommended? I loaded your workflow but it looked very complicated and the manager failed to install a couple of the custom nodes and I'm new to Wan workflows, so I tried to use a simpler one from Pixaroma to start. It could certainly be a workflow issue.
@imafan927 Pixaroma makes very good workflows. I’m not sure if he used Kijai’s WAN Wrapper. Personally, I also tend to encounter OOM errors with the WAN Wrapper, even at fairly low resolutions, but that probably depends on the system.
I suggest using the official native workflow to see if it works. Here’s the link: https://blog.comfy.org/p/wan22-day-0-support-in-comfyui. Right-click on the WAN 2.2 14B i2V workflow and select ‘Save link as’.
Note: when using my model, do not use the Lightning LoRAs — they are already included in the model and activated in the workflow. To bypass them, use Ctrl+B on the LoRA nodes. Try starting with a square image; you’ll see that the default resolution in the workflow is 640×640. Let me know if everything works, if you want."
@taek75799 Thanks! I just tried with your workflow, and it got further but then went OOM again. The image I tried was 544x796, I should try something smaller next, but it also seems likely there's another issue I haven't figured out.
@imafan927 Thank you for your feedback. My workflow is just the native workflow with a few modifications to create longer videos. Actually, the base image won’t have the same resolution as the final video: you can use an 8K image and generate a 512×512 video.
Check the Simple Math Int node: it usually shows 720, so for a square image, you’ll get a 720×720 video. Another node is used to adjust the correct video ratio relative to the image.
I’m still surprised that you got an OOM error. In the meantime, I suggest trying my CAM Q4KM model, which is less resource-intensive. For this, you’ll need to download the GGUF node. To make it easier: download a video from the Q4KM model previews and drag it into the ComfyUI interface. You’ll get the GGUF node, and then you just need to go to the manager to download the custom GGUF node.
@imafan927 If it works for you, don’t worry. I also plan to release the NSFW Q4KM version later, as well as the Q6K version. The latter is more powerful than the FP8 version available here.
@taek75799 Thanks! There's definitely something else going on. I just tried the smoothmix wan2.2 Q5KM ggufs in the civitai smoothmix T2V workflow at 368x512. Memory jumps at the start of the second ksampler, probably expected. But I don't think it went OOM until the VAE Decode step.
@imafan927 It seems there might be memory leaks in a version of ComfyUI. If I remember correctly, this could be causing the issue. Have you updated ComfyUI?
If not, go to the ComfyUI directory and run update/update_comfyui_and_python_dependencies.bat.
@taek75799 It's a good thought. I might actually have the opposite problem--where it's too new:
https://github.com/comfyanonymous/ComfyUI/issues/10723
I reinstalled recently and just cloned the repo and have been pulling from master. I might need to reinstall it at a version prior to the memory changes.
I'm on linux. For a minute I thought it was the easy-use clean VRAM nodes, as disabling those got me a working video. But it wasn't repeatable: I tweaked the prompt and tried again and it OOM'd.
As of now I can use the 5B GGUFs, but not any of the 14B ones, so it's not about your model. I'll keep tinkering.
Thanks for your help!
Weird. I have the same setup and no OOM, even at 720x1280. I run the standard wan 2.2 workflow from comfyui templates. Maybe you're using too many frames, because even at your low res that's incredibly strange. To also add, I can run Phroot's AIO model which is 24GB, and the smoothmix model on civitai which is about the same, just fine too.
@rivdemon1221554 Yeah, it is indeed weird. I even tried rm -rf comfyui folder and reinstalling at a release version (v0.3.67), and just using the default Wan 2.2 t2v template with their recommended models and no custom nodes, and it still went OOM. Something is up where I can't handle any Wan 2.2 14B model, and I'd like to know how to fix it. According to everything I see, my hardware should be able to handle it (ryzen 5 9600X, 32gb crucial vengeance 6000 2x16, gigabyte rtx 5060 Ti Windforce OC 16gb, nvme 990 evo plus), so I'm wondering what's wrong.
Ok, I think I figured this out. 32GB is not enough memory to run Wan 2.2 14B: all models are loaded into memory at once, and you need a little for everything else. Of course you can still run it on 32GB RAM, as long as you have a big enough swap space. When I set up my system I didn't envision doing anything that took this much memory, so I only had 8GB zram swap, so my system would freeze or OOM-killer would kill the process when I hit the hard limit.
So now I added a 48GB swap file and I can run the Q3 14B smooth mix GGUFs. I haven't gotten back to trying your model or another larger model yet, I can try it, and it seems likely I'll want to try the GGUFs when they release.
Thanks again.
even at high step counts (14+ total) and not using any other lora's, most of my videos are very plasticky looking. anyone else? really good prompt adherence though
Hello and thank you for your reply.
Do not use 14 steps in total — this model only works with 2+2 or 2+3.
Try it and let me know :)
@taek75799 thanks, I appreciate the reply. interesting, I tried 2/4 and you're right it's better. FWIW I use wan moe's ksampler (which uses one node for both high and low models, and uses those charts that were floating around a while back to determine how to split the step count across the high/low models for each pass). so without switching to a typical two-node ksampler and putting step counts in manually for high/low, if I use 6 steps, wan moe's ksampler splits it into 2 for high, 4 for low. (if I use 5 steps, it splits it 1/4). that gives pretty normal (not plasticky) result. impressive work, thanks again
@rjox Thank you for your feedback and encouragement.
I didn’t know about the KSampler MoE, I only knew the Lightning MoE — very interesting!
Here, there’s no need to add Lightning LoRAs; they are already built into the model.
@taek75799 oh something else I just figured out about the wan moe's ksampler when using your model.
default setting is sigma shift 5. when total steps are set to 5, it'll split the steps 1 for high, 4 for low.
if I set the sigma shift to 6 and total steps are set to 5, it'll split the steps 2 for high, 3 for low.
so I plan to use sigma shift 6 and 5 total steps with your model to get the appropriate number of steps you recommended.
Thanks for uploading your models. I'm Following you now. Regarding integrated LoRAs you made two yourself yeah? I get that. The one's listed in the Credits though, they inspired your own efforts or they're integrated into this release? If the latter, I'll not enable them in my flow.
hello, yes, that’s exactly it.
They are integrated at a fairly low strength.
@taek75799 Hey great to hear from you. Thanks for the reply - and so quick! Do you remember what strength? Also, VAE 2.1 or 2.2? I've been using 2.1 with SmoothMix i2v with good results and haven't bothered trying 2.2.
@SmutSurfer1960 No worries, you’re welcome. As for the strength, I don’t remember, sorry. For the VAE, use only VAE 2.1 for the WAN 14B model. VAE 2.2 is only meant for the WAN 5B model.
Is the old i2v NSFW version still available?
Hello, sorry, it’s no longer available.
Another clarification please. Reading this page it seemed you were saying you were getting... basically full camera control with img2video. You didn't say that it was img2video but said elsewhere that you don't really do much in the way of txt2video.
Did I mis-interpret? because I used the line:
"Handheld camera circles around subject, camera orbit around subject."
All I get though is a close-up - way closer than I normally get btw - and no sideways arc movement at all.
If it matters I'm using the SmoothMix workflow I've been using for a while. Maybe it needs something more for strong camera movement adherence?
Try using the Cam model with this: 'camera starts a smooth 360-degree orbit around the'.
Here, you can get an idea of what it might look like: https://civitai.com/images/108548197.
I don’t think it will be difficult for the workflow.
@taek75799 Once again thanks for the (quick) response. That seems like a txt2video prompt. (I saw it above). It has seemed to me that, for img2video of a sex scene, the model and / or the LoRA(s) lock the camera in place with maybe a little sideways side-to-side movement to simulate shaking? It'll be great if I can get camera movement working for img2video.
@SmutSurfer1960 The Cam model is an i2V model. It was created using a starting image. Try it — it will work, maybe not every time, but it really works.
@taek75799 Man you're the man! I tried your exact prompt after adapting it to a, uh, spicy bj image. I did get a fair amount of movement in it. Trying the exact prompt with a screenshot of the gif you linked worked a treat. LOTS of camera orbiting. Pretty much what your own example shows. SWEET!!! No more being locked into the seed images pov. Thanks man! Yours might be the only img2video NSFW model I use from now on.
@SmutSurfer1960 Thank you for your feedback. It was specially made for camera movement, which I find important for creating a good, dynamic video.
Tested i2v. The first thing I did was test a configuration similar to the "non-NSFW" version; the results were poor. I disabled all the LoRas and it worked very well; you just need to add a few more LoRas with a low weight to achieve the desired movement/details.
I love this model. It works great. (I'm using the CamI2VQ4KM variant). I'm getting this message though and would love to fix the problem. The AI chat bots aren't a lot of help. The message is a lot longer than what I have pasted here but you get the idea:
clip missing: ['encoder.block.0.layer.0.SelfAttention.q.scale_weight', 'encoder.block.0.layer.0.SelfAttention.k.scale_weight', 'encoder.block.0.layer.0.SelfAttention.v.scale_weight', 'encoder.block.0.layer.0.SelfAttention.o.scale_weight', 'encoder.block.0.layer.1.DenseReluDense.wi_0.scale_weight', 'encoder.block.0.layer.1.DenseReluDense.wi_1.scale_weight', 'encoder.block.0
...
Don’t pay attention to that, there’s a LoRA incorporated in the model that doesn’t use all the blocks. These aren’t errors, just informational messages.
@taek75799 Great! Thanks. This learning to create AI images and videos is probably the most complicated, nuanced thing, that I have ever learned. Or in this case, begun learning. Sooooo many variables.
honestly, absolutely cracked, we live in the future
Do you have the Full Precision Model available? I need to quantize it more than FP8 to run on my system
Hello Alexi,
I haven’t made the model in BF16 yet, but I will soon and post different GGUF files. Which one do you need?
@taek75799 Q3_K_S.gguf models work for me
@alexissilver124816 Oh really, you’re the first to ask me for this one. What GPU do you have? The Q4 doesn’t work for you?
@taek75799 NVIDIA GeForce RTX 3060 Laptop GPU. Main limitation is only 6 GB VRAM.
My setup cant run this unfortunately but I saw that you are planning on releasing the GGUF. When can we expect the GGUF version to be released?
HI, a lot of people have been asking me for the GGUF files. I’ll make them, but unfortunately I’m no longer on vacation and can only work on them during the weekend. I’m not sure which one to start with, but I think it might be best to start with the Q4 versions so that people who can’t run the model on their machine can use it.
Which version do you need?
@taek75799 it’s no worries. I use Q4 K_S or K_M for my setup.
@Adam_Lol I will make the Q4_KM version; it's a bit better than the 'S' one.
@Adam_Lol Thank you for your comment, Adam.
8gb RTX 3070 with 32GB DDR5 and was able to run the the 13gb NSFW fast move FP8 high/low. Windows Task manager only shows under 7gb video use for i2v, 5 sec, 512x512 video 4 step. My workflow is very simple tho
have you tested with this by chance?
https://github.com/GalaxyTimeMachine/ComfyUI-WanMoeKSampler
been using it with other wan 2.2 and it seems to work really well.
(edit: linked wrong thing)
also, im still learning but, i have a pretty good system and i see everyone does fp8 - for the kj checkpoint loader, what is the weight/compute type? i like to set things manually because literally 0 reason and test maximizing things to highest point to see what happens.
I use wan moe's ksampler too. setting the steps to 6 will split them to 2 high/4 low which renders a good result with these models. EDIT: using a sigma shift of 6 with a total step count of 5 will split them into 2H/3L which is better.
Hello, and thank you for your link.
You’re already the second person to mention these KSamplers to me, and I now remember having used them before, but I didn’t notice any difference. To be honest, my test was quick I should explore it further, thanks.
For the checkpoint loader, don’t worry leave it at the default. The model here is in FP8, so it will stay in FP8. For example, if you have a model in FP16 and it’s a bit too heavy for your system, you can set FP8 on the checkpoint loader, which is quite convenient.
I’m not sure if most people really use FP8; I think they do. But personally, I prefer using the GGUF versions in Q8, which are very close to FP16 and therefore more precise for WAN, Flux, Qwen, etc.
@rjox what other settings you make in WaNMoEKSampler???
@rjox Thank you for your feedback.
It seems that these KSamplers are well appreciated, so I’ll take a look at them.
If I may ask, have you tried the Painter node?
It allows the video to be less slowed down and gives much better camera control.
And thanks again for your comments.
@taek75799 boundary 0.9
add noise: enable
control after generate: randomize
cfg high/low: both 1.0
sigma shift: 6 (5 is default but using 6 and a total step count of 5 will split them 2H/3L)
start at step: 0
end at step: 10000
return with leftover noise: disable.
For these models, I've been using euler simple.
For the normal wan 2.2 fp16 models I always used dpmpp_2m_sde_gpu simple as I find dpmpp gives more realistic details for photorealism and the gpu version is slightly faster.
@taek75799 I did download the painter node but having tried it out yet. anything in particular you'd like me to test when using your models?
@rjox CFG fall ratio high/low left at 0.5??
Why am I losing color correction?
@rjox Alright, thank you for all these tips. I’ll try that very soon.
@rjox I don’t want to bother you; just enjoy the model, but thank you for asking :)
Can't seem to get the camera to actually rotate around the subject or get a person to walk and have the camera/environment to move with them? For example moving through a corridor or whatever. Otherwise this is an amazing model.
Agreed. I don’t know how to make the camera rotate around the main object of action.
Hello, and thank you for your feedback.
Which model are you using?
The NSFW Fast Move version benefits from advanced camera control, but it’s not as powerful as the Cam version of the model. I had to make a few compromises, unfortunately, but I’ve already managed to make it work on the NSFW Fast Move version, although it depends on the scene.
Try this: "camera starts a smooth 360-degree orbit around".
Here, you can see a preview: https://civitai.com/images/108548197
@taek75799 I'm using the NSFW F8 model. I tried the "camera starts a smooth 360-degree orbit around" hint, but the camera wobbles back and forth and nothing else happens. I'm also using the Painter node.
@mmikemiller823390 Maybe try the Cam version if your scene is NSFW. Try using a LoRA — it could work. In any case, you’ll have a better chance of success.
All the models are I2V, aren't they?
There is a T2V version — look above and scroll to the right — but to be honest, it won’t be as good as the i2V version. I haven’t worked much on the T2V version.
However, several people have shared a tip: if you want to try it, just use a white or black image along with your prompt on the i2V model.
Of course, at the start of the video, you’ll get a few frames with the black or white image, but according to some, it works really well.
@taek75799 i2v is ok. I just wanted to be sure :)
was this renamed. i have wan22EnhancedLightning_v11Q4KMLOW and i cant find the original creator to give credit....
Hello, yes, it’s the 1.1 model from this page. I don’t know why, but when I upload the files to Civitai, the name changes completely, which is annoying.
@taek75799 thanks for the fast reply! now it makes sense why nothing looked different. i used the same model, named changed (didnt know) .
@goodstrk Maybe also try V2, if you haven’t done so yet. It provides better camera understanding. You can see how to use it in the description, if you want.
will definitely try
@taek75799 wan22EnhancedNSFWCameraPrompt_camI2VQ4KMLOW is this v2? for the q4 model low
@goodstrk yes :)
NSFW Fasf and NSFW -- butt cheeks are too 'jelly' to me
Hello, thanks for your feedback. I’m currently working on another version. This one won’t have those hip movements anymore, but it will have a much better understanding of the prompt and the camera.
Waiting for that good ol' Q4KM Qs.
What is the maximum length of video you can create with these?
WAN 2.2 is designed for 5-second videos, but you can go further. Check my workflow at the very bottom of the description.
However, note that every 5 seconds, there will be some loss of quality. We don’t have a choice for now.
For V2 camera you can do 7 seconds natively 100% guarantee.8 seconds only 30% will work, most of the time it's will loop the animation
Some models have a tendency to render back to the source image if the length is greater than 5 sec. I stumbled upon that not long ago when I switched the frame rate from 24 to 16 but kept the length at 121.
On my experience videos beyond 15 sec demonstrate severe degradation of quality :( https://civitai.com/posts/24517055
@PopHorn1956 I completely agree. I also try not to exceed 15. However, some people are less bothered by this. There are still methods to reduce transition degradation, like a good upscaler (SeedVR2 or others), but of course, the generations will be much longer. Unfortunately, we still have to rely on these kinds of tricks.
@taek75799 15 sec length actually great. Does the workflow here OK with 15 sec? I never heard about SeedVR2, will check it out
(you can see it in the 4th preview "fast move high" where the man slaps the woman)
where is that ?
Hello, this one — ‘the man violently slaps the woman's buttocks’ — appears at the end of the video: https://civitai.com/images/110323732
If I want to avoid any change in the body shape, for example, if a person has a very large belly, the model should not automatically reduce the size of the belly and include the person with the same features in the video. Which model should I use? Because your latest model reduces the body volume a little.
Hello, maybe try a body slider LoRA. This one is made for T2V, but often these also work on I2V. I cannot guarantee that this is the case here: https://civitai.com/models/1980236/wan22-skinnyfat-slider
@taek75799 Thank you very much for your attention. What I meant was that when a character takes off his clothes in the photo to video conversion, the models I downloaded still reduce the body size despite the explanation in the prompt. Is there a model that depends on the prompt?
@celloplayermj435 Hmm, good question. I haven’t tried that, so even with words in the prompt, you can’t get it to work: Wan doesn’t understand everything. The only solution I see is the LoRA posted above: you can control the body’s size after the person has removed their clothes. To do this, simply choose the LoRA strength like this:
Negative (-2 to 0): Slim to bony (use the prompt to help; below -2 it starts to degrade the image, but I got some results even at -4)
Positive (0 to 3): From big to huge (use the prompt to help; above 3 it works on some images)
@taek75799 thanks alot
If we could get a q8 gguf of the nsfw checkpoint that'd be great. Q8 GGUF's are always much higher quality than an FP8.
For some reason, the NSFW F8 model doesn't understand the undressing process very well. Either clothes spontaneously disappear, revealing the body. Or, when taking off a blouse in the middle of the scene, the subject's pants also come off. This problem wasn't present in the model with an improved camera or in the first versions of this model. But the NSFW F8 reproduces the actions themselves well. Maybe I'm doing something wrong.
Try this, I don’t know if it could work
She lifts herself up, takes her top off completely, and becomes fully naked, showing her large breasts and nipples.
or
A woman removing her pants with both hands exposing her nude body once she has the pants off she tosses it to the side. She is removing her shirt over her head with both hands exposing her completly nude body.
or
年轻女子随后脱下衣服，露出她赤裸而性感的身体。
Yes, you need to describe her movements in more detail https://civitai.com/posts/24517989
@PopHorn1956 The problem is not that he doesn't understand the prompts, but rather that the NSFW version is more aggressive in terms of undressing, and he misses out on certain movements.
How are you getting to the fps of the example images? The flows say 16fps, but they are definitely higher than that. I'm getting artifacting around the eyes in movement, but I think that might be because of the rife vfi node that interpolates frames.
Hello, and thank you for your comment. The videos are at 60 fps using interpolation. Check my workflow in the top left: you’ll find a node to do this (RIFE Frame Interpolation). For eye defects, try increasing the resolution if you can; this should largely fix this kind of issue.
https://civitai.com/models/2079192/wan-22-i2v-native-enhanced-lightning-edition
Sorry, I didn’t notice that you were already using VFI. However, take a look at the one I use; it interpolates much faster.
@taek75799 Thank you! I'll take a look.
@taek75799 I've tried a bunch of stuff and can't get past it. Things are just blurry, specifically noticeable in the eyes. I can't install sage attention, could that be the issue? With previous models I've still used the NAG and Patch Torch. Tried those on and off. I've had better luck with some than others. I removed clip from low path. I removed clip vision. Installed some nodes that you used which were different than mine. Used the rife node from your flow. Upped the resolution... Same end result. Any other thoughts?
@rando2048 I don’t think this issue is caused by SAGE ATN. It only speeds up the generation and degrades the video very slightly, almost unnoticeably. I believe you already tried everything. It might be due to your original image, or maybe the face is too far away. You could try upscaling the image with SeedVR2 — that usually helps, and it’s what I use to make my previews.
What is the maximum resolution you have tried? Personally, if I have a square image and I want to make a video, I go up to 1024×1024.
@taek75799 Images are mostly gens from sdxl. 832x1216 or 896x1152 for the most part, but sometimes I tinkered with those. Then I use upscale more like a refiner with a different prompt at 1.2 1.25 or 1.5x. One image I just tried was 1120x1496. I wasn't cropping, I was setting the width to 640 and height calculated to maintain aspect. It helped to change it to 832 or 720. Then I set it to crop and hard set the values to 832x1088 and 720x1280, which seemed to have helped slightly more. So it it does seem to be resolution based, but not as in "not high enough". Also, not head shots, but face is plenty large enough.
I'll keep tinkering. Side note, lace clothes have similar issues, it's like small intricate things don't transition in movement well, but you don't notice it on something say like an arm. That's kind of why I originally thought rife may be doing it. I see it in the individual images in "Save Image" too though. If I have time later, I'll upload an example.
@rando2048 Thank you for your feedback. Feel free to show your example and, if you want, even leave the metadata on the video so I can check the workflow — there might be a small thing to adjust, you never know. And if you want, you can even post the image; I’ll try it on my side.
Godd luck.
Anyone else having issues getting it to generate pussy? Like with no cock/insertion? The generated video looks really wrong. Original image has no genitalia showing
Hello, try using a LoRA if you want: https://civitai.com/models/2109996/wan-22-pussy-and-anus-lora
I just tried it. Of course, it’s not perfect. Try incorporating the LoRA mentioned above; you will get better results:
https://civitai.com/posts/24513774
I don't know what I doing wrong, but I have a lot problems with a camera. It not works as it should. But when I switched back to default Wan 2.2 - everything works well. Even wan22RemixI2V works better. I tried gguf 4KM, 8 and fp8 tensors.
Hello, and thank you for your feedback. I’m currently working on a new version where the camera will be improved. But when you say camera, which angle are you looking for?
Si vous voulez que j’adapte le ton (plus pro, plus friendly, plus simple), dites-moi !
@taek75799 I tried "fast dolly out 3m". I have a picture with a sitting girl, so I need to move the camera back so she will be fully in the camera frame when she stands. But this model tends to move the camera closer, even if I was able to move it away. I have a complex set from 5 continuous steps (the last frame is a start frame for another) and what surprises me that it works as intended on the default model from huggingface. But I still cannot figure out the logic.
@Dylfin I completely understand. I don’t know which model you downloaded here. There’s a model called Cam, which is specifically made to improve cameras. Try it if you haven’t already. However, an important thing to know: Wan can’t do everything regarding the camera. It can do a lot, but not everything. The prompt is also important. Check the description of the Cam model: you’ll also find an overview of the prompts, and don’t hesitate to look at the previews of the Cam model — there are examples of prompts with the words to manage the camera.
will there be quantized versions?
Hello, I'm still unsure about releasing the GGUF versions. I managed to improve the model, so I don’t know if I should publish this new version directly instead.
@taek75799 I'm asking simply because I have an 8 GB 3070, so I'm unlikely to be able to run a 14 GB model.
@zombieleaver I’m wondering which GGUF model could run the model on your PC. Does Q4_KM work, or would something lighter be needed?
@taek75799 q4-km and q3-km worked and started, but I didn't try the others, not counting on them.
@zombieleaver If the Q4_KM works for you, that’s perfect. They are better quality than the Q3. Don’t worry, I will post them for sure, just a little patience ;)
@taek75799 Thanks for that.
A Q6 would be appreciated, too! I'd like to try the model, I am able to run Q6 on my 8GB 3070 TI.
@woodenpickle I will release the Q4KM and the Q8 this weekend. I wasn’t planning to release the Q6K, but if you need it, I imagine you’re not the only one. I’ll probably release it on Wednesday. Sorry for the wait.
@taek75799 Q5_KM of NSFW and CAM would be also much appreciated ... but no rush. Great work btw. :)
@nicknack443540 Hello, and thank you for your comment. I was also hesitating about making the Q5KM for the Cam model. You’re the first one to request it, so I will do it, but I don’t know when yet. However, to be honest, I’m not sure if I’ll make it for the NSFW version. I’ll soon release V2, which is, in my opinion, much better than the one posted here. Thanks again.
Thanks for the models and workflows you provided, they r really constructively useful and efficient. Im so surprised that the generated videos r both real and fast on my pc !:) It's just one more thing: could there be any possibility to improve the integrity and accuracy of the body? I mean I do put in some prompts for perfection about body parts like breasts, for example, round and strong upper pole fullness. But they always seems to be the sagging ones, even though I put reverse prompts in the negative clips. It just does not have any effect. Could this problem be better solved? Im very satisfied about the work you've done right now but it'll be perfect if these details are concerned. :)
Hello, and thanks for your feedback. It’s pretty hard to please everyone: I had to make some choices. It’s not personal, I just wanted to keep a certain coherence. Because of that, I couldn’t add everything: we would have lost coherency in the face, for example. All I can advise you is to use LoRAs — most of the ones for Wan are excellent. I can show you a few if you want: I don’t know if these ones will suit you, but don’t hesitate to look, I’m sure you’ll find what you want:
https://civitai.com/models/1971502/wan22-breastchest-slider
https://civitai.com/models/1648982/wan-nsfw-posing-nude?modelVersionId=2391828
https://civitai.com/models/1980236/wan22-skinnyfat-slider
https://civitai.com/models/1410875/perfect-body-wan-t2v-14b-lora
Got it, thanks for the reply, I understand what you mean, it's probably for the fact that i've just get into these things recently, but now i think i know how to deal with it, im sure that i will check them out, thanks again! :)
@izark00326 I see. I hope you’ll manage to achieve what you want, and feel free to give me feedback.
Maybe this is a stupid question but is this safetensor too large to use on a web hosted server like wavespeed.ai? I'm trying to figure out if I can do it
Marvelous models! Only one "issue". Is there a way (prolly by a specific prompt) to reduce the bounce/sag/jiggle of the soft parts of the body on I2V? I'm on Fast Move version. Would like to keep the carefully sculpted attributes firm and tight. ^^
Hello, and thanks for the question. To be honest, maybe it would work with some LoRAs, but I’m not really sure which ones. Maybe this one, but I haven’t tested it myself: https://civitai.com/models/1648982/wan-nsfw-posing-nude.
Have you tried using negatives, for example? If not, check the description to see how to do it — it might help.
I will make sure to improve that in the next version.
@taek75799 Yup, tried various positive and negative combos, didn't really work. Thought maybe there are specific words the model responds to. Thanks for the reply! ^^
@Futopia You’re not the only one who noticed this issue, which isn’t really a problem for some people, but I prefer to remove these movements when they aren’t specified in the prompt. A new version should be coming soon, without this issue and with additional improvements. Thank you for this feedback, it really helps me.
I appreciate a lot the fast model. I agree that maybe you should give us a trigger to have a bit less bouncy on the soft part of the body. Still very good work
You are the third person to tell me this, and I imagine others must feel the same. I’ll take it into account to improve the next update. Thank you for that.
I really like the flexibility of prompting for this model, though the jiggle physics are a little out of control.
Quality-wise, the main downsides I found are that limbs like to vanish often, or torsos will contort so the upper and lower body are facing opposite ways. Also almost every video starts with a weird facial gesture for some reason.
But still I think it's one of the best here, great work and thanks!
Hello, and thank you for your feedback. I’m taking note of your comment. I’m preparing a new model that is almost ready: it should fix most of the issues and have a much better understanding of the prompt, along with other improvements. I’ll keep the rest as a surprise.
For the torso position, I’ve had this issue before, but it was due to my prompt: the image I had didn’t match the prompt, or I needed to specify the correct position in the prompt.
I dont see a Natural Motion Version edition. I only see the fast edition. Where is the natural motion edition?
Hello, and thank you for your comment. The difference between the NSFW version and the NSFW Fast Move version is that one offers slightly faster movements. That’s also why the preview videos are exactly the same, so you can compare them. I know it’s not very noticeable, but depending on the scenes, it can become more apparent.
The next model should have much more natural movements. ;)
@taek75799 Ah got it. The normal NSFW version is the natural motion version. I actually think in some cases the fast edition can be more natural - it really depends on what's happening in the scene
@ArcFXSpank Indeed, I think the same. It can depend; honestly, it's really hard to have something, with WAN 2.2, that is flawless in every situation. We have an excellent model, but not a perfect one yet.
Very nice. Motion issues are a huge part of why I usually stick to 2.1, or at least use 2.1 as a base with 2.2 lightning with tripleksampler. There's quite a toolset here, all of them well worth checking out. Well done.
did you train it to understand this format
(at 0 seconds: ).
(at 1 second: ).
or just recomendation
Try it on WAN Native, it will work, but not as well.
Hey, thank you very much for the checkpoints and the thorough usage explanation. I've just tried the Cam Version and did notice the face changes immediately. It changes completely in about 1-2 seconds. You wrote to increase the resolution, do you mean the resolution in the Wan Video node or the resolution of the base image?
Is this a quirk of all releases of this checkpoint or is it a Cam thing only?
Apart from that it's fun messing around with it. Nice work. Thanks again for your work.
Hello and thank you for your feedback. The starting image must be good, but the most important thing is the video resolution. If it’s too low, the face will change on any WAN model.
@taek75799 Thanks for your reply. I usually have the resolution set at 720 x 480. I didn't notice face changes on any other models I've used. The starting images usually have a resoltion of 1216 x 832. Again, with other models I didn't notice any face changes. I wonder what could be wrong here.
@fatberg_slim Thx for feedback.
To be honest, from my position, I couldn’t tell you. I haven’t noticed this problem on the CAM model; if I do, it’s due to too low a resolution. If you wish, you can send me your prompt and image so I can try with my wf it myself and show you the result, of course if it’s not too personal.
did you bake any loras into this?
Hello, yes, check the description: the information is there.
@taek75799 ok, so it seems that the loras added were your own, and not released separately, right?
I just don't want to double-stack loras, e.g. you've got a large pectoral lora in here and I add the same large pectoral lora into my workflow, skewing the weights too far.
@ForeverNecessary737716 Yes, you can use lower values like 0.7. But let’s say that when you do a merge, adding the same LoRAs that are already included in the merge won’t systematically degrade the video, unlike adding them twice in the base WAN 2.2 model. That’s one of the advantages of merging. The ideal is to test the model and see if you want something specific; in that case, adding a LoRA to help the model would be a good choice.
the models are awesome, even if i sometimes have to switch from the nsfw model any time a dudes junk is anywhere near a womans and she isnt supposed to be humping it : |
Hello and thank you for your feedback, haha. This will be fixed in the next update.
@taek75799 dude no worries honestly i dont even know how you fix that effectively. you spend ur whole time trying to make these ppl hump and im telling them not to?! that just aint fair
@Fosphur Haha yes, I see what you mean. It’s already fixed for the next version. I ran some tests: the man will be able to appear and perform other actions.
The best NSFW model I've encountered. Previously, I used a Remix model, but in Remix, even in half of the simple poses, the characters start squatting, bouncing, convulsing, etc.
The only thing I don't like about your Wan Enchanced NSFW model is that after 1-2 seconds of character movement, its skin becomes too clean, losing all the skin details that were on the image, which makes the characters look artificial and 3d-like.
Also, I have encountered the issues that others have mentioned in the comments, such as sudden switching of the front and back of the body or too active physics of soft body parts when the body makes minimal movements, but as I have seen, you will find a solution and compromise for these issues.
Thank you for you hard and great work!
Hello and thank you for your feedback. The lower body will no longer have this issue as much; it might still happen sometimes, but much less. To be honest, it is extremely difficult to make WAN understand this outside of sexual scenes on an NSFW version.
Regarding the body changes between the front and the back, I don’t think it’s a real problem, as I’ve had it myself before, but I managed to fix it with the prompt.
For the new version, the prompt understanding will be improved, and the same issue might still appear, but it will be easy to fix.
Thank you again for your comment, and I hope the next version will please you even more.
@taek75799 The loss of skin detail (texture) affects not only the lower body, but also the upper body and face. But I think we understand each other ;)
@missiya 168033849530312
Does anyone else have the issue that in "normal" scenes (no sexual activity) it is not possible to make the characters stop moving their hips? I tried both the fast move and slow/natural move version. For example, i want to create a video where a character is standing still, and just looking at the camera. The character always thrust their hips/butt forth and back or left to right. I don't use any Loras. I don't know if it is the checkpoint itself or maybe my setup with CFG=1?I am using NAG for negative prompting so i don't have to give a negative prompt and higher cfg. I have this setup: "Load Clip" node, which is feeding into the clip-input of "Super Power Lora Loader (rgthree)"(high an low noise lora loaders). The "Super Power Lora Loader (rgthree)"(high only) has the CLIP output connected to the positive and negative prompts. The "WanVideoNAG" nodes (high/low) take the negative prompt as conditioning-Input and output their model both to the high and low Ksamplers. The single "WanImageToVideo" node takes in the "positive" input from the positive prompt. This is outputted to the positive-intake of both ksamplers. the negative Input of "WanImageToVideo" is a "ConditioningZeroOut" node (the conditioning-Input of that node is the positive prompt, the CONDITIONING output is to the negactive input of the "WanImageToVideo" node.
Hello and thank you for your feedback. To be clear, yes, this is due to the model. You should know that this issue will be partially fixed in the next version. I emphasize partially: this effect will be much less noticeable. It is very difficult to remove it completely. I can do it, but the sexual scenes would be of lower quality. However, I think it will be much less present.
I have the same issue, using the regular Wan 2.2 model with various loras. It's like a monkey-see, monkey-do thing. (If one person is thrusting in and out then all of the people watching need to do it, too.) It's annoying, for sure.
@taek75799 thank you for the reply! I tried other nsfw checkpoints as well, the behaviour is the same. I have a workflow that does multiple videos in sequence. So basically what i would need to do is: Use the official wan 2.2 high/low diffusers for anything non-nsfw, change diffusers for anything nsfw. don't know if that can be done. Something like video1 (a character waves at the camera):wan 2.2 , video2 (a character pleasures itself): NSFW_fastmove, video3 (a character walks out of the room):WAN 2.2
@DigitalDegeneracy You can try this, it’s ideal. Unfortunately, on a long video workflow, if I do this, I get an OOM even with a VRAM purge. What I can advise is to use the V2 CAM model: it has no NSFW LoRAs, it is just optimized for prompt understanding, offers more dynamics, and better camera comprehension. If what you are looking for isn’t sexual poses, it will work well.