You can click on the link below and try it out directly. If the effect is good, you can deploy it locally.
Online Workflow: Wan2.2 Remix i2v 3.0 Image-to-Video Automatic Reverse
Experience link: https://www.runninghub.ai/post/2033536710499373058/?inviteCode=rh-v1325
Online Workflow: Wan2.2 Remix I2V v2.1 SVI2Pro Fully Automatic Reverse Long Video
Experience link: https://www.runninghub.ai/post/2008833586203463682/?inviteCode=rh-v1325
Online Workflow: Wan2.2-Remix-T2V-v2.0 text-to-video workflow
Experience here: https://www.runninghub.ai/post/1991533843182264322/?inviteCode=rh-v1325
Online Workflow: Wan2.2-Remix-I2V-Comfy-Qwen3
Experience here: https://www.runninghub.ai/post/1986632318448267265/?inviteCode=rh-v1325
Now! Register to get 1000 points, daily login 100 points, play 4090! Experience the super power of 48G!
[B-site video]: https://www.bilibili.com/video/BV1hF2cBNECZ
[Hugging Face: Wan2.2-Remix]:https://huggingface.co/FX-FeiHou/wan2.2-remix
[Telegram Group]: https://t.me/wan22remix
The workflow is very simple to use, and I have uploaded a tutorial on how to use it on YouTube. You can try it out and follow the tutorial and Civitai's workflow to easily reproduce the results. Thank you. If you have any other questions, please leave a comment in the comment section
Model Overview — Wan2.2-Remix
Wan2.2-Remix is a research and creative model built to generate short, cinematic video clips directly from text prompts. It focuses on human dynamics, realistic motion, and scene consistency, producing lifelike movement and smooth transitions without requiring LoRA setup.
This model already includes NSFW capabilities, enabling the generation of high-quality NSFW-themed videos without the need for additional LoRA.
High-noise model T2V: Lightx2v Wan2.2 Lightning dyno version (quantized to fp8)
High-noise model I2V: Comfy-Org Wan 2.2 ComfyUI Repackaged (fp16, quantized to fp8)
Low-noise model: Comfy-Org Wan 2.2 ComfyUI Repackaged (quantized to fp8)
Enhancements: Blended with open-source motion LoRA data and refined pose training for enhanced realism, anatomical precision, and gesture diversity.
Integration: Fully compatible with
lightx2v, no extra setup required.Recommended CLIP: NSFW-Wan-UMT5-XXL
Recommended Settings:
Steps: 4–8
CFG: 1
Shift: 5–10
Sampler: Euler
Scheduler: Simple
This release is currently beta. While functional, it is not yet perfect—feedback and testing are highly appreciated.
The GGUF version of the Remix model is now available.
Thanks to @BigDannyPt for creating it!
Download link:Wan2.2-Remix I2V GGUF v2.0
Changelog
I2V v3.0
Added VBVR LoRA to enhance spatial understanding and improve scene consistency.
Introduced UnifiedReward optimization to improve overall video quality and generation stability.
Adjusted several NSFW LoRAs to improve quality and visual consistency in NSFW scenes.
This version is considered the final release of the Remix I2V series.Unless there are major technological breakthroughs in the future, no further updates are planned.
I2V 2.1
Enhance facial consistency in long videos
Adjust LoRA ratio to make video dynamics more realistic
The v2.1 version has reduced the weight of the NSFW in the low-noise model to enhance character consistency. If you want to get a better NSFW, you can mix and match the low noise model of version 2.0.
T2V 2.0
Optimized character stability in generated videos, reducing jitter and unwanted motion artifacts
Enhanced video detail for more visually rich scenes
Added dynamic motion effects to improve realism and cinematic impact
Updated base model to original Wan2.2 version to minimize differences across random seeds
I2V 2.0
Significantly enhanced character motion and poses for more dynamic animations
Greatly improved camera movement and cinematic framing
Optimized NSFW content weighting to reduce jitter in generated videos when producing SFW scenes
Enhanced explosions and particle effects for more impactful visuals
SFW version will no longer receive updates
I2V 0.8
Updated lighting acceleration module to Wan2.1 version, fixing slow-motion character issues
Integrated over 20 open-source LoRAs and 10 custom-trained LoRAs to enhance motion dynamics
Removed several low-quality LoRAs to improve overall visual fidelity
T2V 1.0
Reduced dyno weighting to improve stability across different random seeds
Added more Asian facial datasets for greater diversity
Enhanced character motion dynamics for fluid animation
Improved motion amplitude, gesture diversity, and overall visual quality
T2V 0.9
High-noise T2V model: Lightx2v Wan2.2 Lightning dyno version (fp8)
High-noise I2V model: Comfy-Org Wan 2.2 ComfyUI Repackaged (fp16, fp8 quantized)
Low-noise model: Comfy-Org Wan 2.2 ComfyUI Repackaged
Refined character motion blending for smoother transitions
Enhanced anatomical articulation and natural body flow
This model incorporates the following Loras. Thank you to the Lora authors for their work. In order to facilitate management, I will change the file name uniformly after downloading Loras, so I cannot correspond to their real names one by one. Please understand. If there is any infringement, please contact me and the next version will remove it.
Wan2.1 - lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16
Wan2.2 - I2V_A14b_high_noise_lora_rank64_lightx2v_4step_1022
Wan2.2 - I2V_A14b_low_noise_lora_rank64_lightx2v_4step_1022
Wan2.2 - Fun-A14B-InP-HIGH-MPS_resized_dynamic_avg_rank_21_bf16
Wan2.2 - Fun-A14B-nP-LOW-HPS2.1 resized dynamic avg rank 15 bf16
Wan2.1 - Skyreels-v2-T2V-720P LoRA rank64 fp16
Wan2.2 - PusaV1_lora_HIGH_resized_dynamic_avg_rank_98_bf16
Wan2.2 - PusaV1_lora_LOW_resized_dynamic_avg_rank_98_bf16
Wan2.1-T2V-BBFW - Better Buttholes For WANv1
Wan2.1-12V-General NSFW-Low
Wan2.2-12V-Hip Slammin Assertive Cowgirl
Wan2.2-12V-PENIS LORA e496
Wan2.2-12V-DR34ML4Y -All-In-One NSFW
Wan2.2-T2V-Body Shots v2 14B
Wan2.2-12V-Deepthroat, Blowjob
Wan2.2-T2V-Masturbating Rubbing 14B
Wan2.2-T2V-Doggy Style-14B -v3
Wan2.2-12V-COWGIRL+REVERSE COWGIRL-14b
Wan2.2-12V-Combo Handjob Blowiobsafetensors
Wan2.2-Bounce-Test
Wan2.2-T2V-GodPussy
Wan2.2-T2V-MysticXXX
My Train:
P001-Side-Lying-Sex-Wan2.2-i2v v10
P003-Sexy-Enhancer-Wan2.2-i2v
P004-Lying-Bouncing-Wan2.2-i2v v20
P005-Chinese-Dance-i2v-v20
P006-Squirting-Wan2.2-i2v
P008-Running-Wan2.2-i2v v10
P009-The-Walk-Wan2.2-i2v-Light
P010-whip-t2v-v10
F006-AsianGirls-Wan2.2-t2v v10Disclaimer
This model is released for research and creative purposes only. It must not be used for illegal, unethical, or harmful activities. The author assumes no responsibility for any misuse or consequences arising from its use.
Description
FAQ
Comments (53)
Can we please stop with the mix-and-match languages in descriptions? This shit makes it IMPOSSIBLE to translate the whole page
I love how the description is chock-full of words, but none of it really explains what the hell it is or how it was made
lol thinkin same thing
So far this is very good. I am using 1.0 low noise for T2I and refining with Flux Krea, it creates excellent quality images of people. Better anatomic details that standard Wan2.2. Look forward to future updates and quants.
very nice but it exceeds my memory capacity, I wonder why your models needs more memories to work, with the same workflow, than the basics ones
I just incorporated some Loras into the original model, and the model capacity has not changed. It should not occupy more video memory than other FP8 models
comfyui
gguf version ? only had 4060ti 16gb
It is already in FP8 quantization format and does not occupy much video memory. You can enable KiJai's block swap node,
Do you have a workflow for the I2V model on your huggingface?
There is a supporting workflow on huggingface, which is actually the most common text-to-video workflow
Any plans to release fp16 versions for higher quality?
My internet speed is too slow, and uploading fp8 is very laborious, so I don't plan to release fp16 for the time being.
I have no idea how you have managed it, but this is incredible! Not only does it almost entirely eliminate the need for 99% of NSFW LoRAs, but it actually does a better job than many of them! Excellent work!
I combined some NSFW loras that I carefully selected with the loras that I trained myself, and adjusted the weight ratio through a lot of testing.
Any chance we can get a version of this without the lightning lora built in?
According to my current fusion method, there is no way to not integrate lightx2v
@FX_FeiHou But On my t2v Generation I still have to load the lightx2v with a Lora Loader, right? And can I use this for i2v generation aswell?
@MartinPointner1978 No need to load lightx2v anymore, a sampling step of 4+4 can achieve good video quality. This model is not applicable to i2v.
Looks great! Awesome job with the checkpoint/model
Extremely versitile checkpoint. My Favoriten right now. Only one thing is bugging my gernerations: when i do Missionary sex scenes the man is always generated with a vagina, when he ia schien from behind. And hints how to avid that ? Or maybe the database lacks poses Mike "mating press"
Perhaps the data is not comprehensive enough, you can write more detailed prompts, such as seeing the male anus and testicles from the back.
Thank you. ill try and See if i could post the results
Great JOB!
用起来很不错，感谢大佬开源
is there any i2v models for this?
There is a preview version of i2V on Huggingface, but I think it's not good, so it wasn't released to Civitai.
any chance for a gguf version , Q5-K-M
Most wan2.2 models creators refuse to make any gguf🫤
@Mikeflower I'm not refusing to make GGUF, but I don't know how to do it
for those with 8gb of Vram (like me) i can use the model with Ksampler MOE ( or double stage ksampler) , you don't need to use block vram , On the workflow on the sample videos posted here i'm OOM , not with my workflow , i will post some exemple with metadata soon :),
@Meandre66 and just use the same on high and low?
@tethor11 nope moe ksampler use the High and low
@Meandre66 Do you have any uploads that were made with this method?
@mobdik17378 https://civitai.com/images/105507466
my exemple 832x480 and 81 frames, can be up to 1024x576 but in 49frames max, 1280x720in 33 frames
and i have 32 GB of DDR4 RAM, if you don't have vram u need ram
@Meandre66 The link turns up a 404
@mobdik17378 i see it right now on the model showcase, the link may be ok now
@Meandre66 Still 404
@mobdik17378 https://civitai.com/collections/13188233 ?
works fine on 3060 12gb (comfyui with default workflow t2v wan 2.2)
had better results with v0.9
How do you add loras to your workflow?
Great model! Res_m and res_s bong sampling seems to produce poorer results from testing though as opposed to Euler simple (lots of degradation). Additionally running at higher step count does not seem to produce higher-quality results but rather, seems to impact quality and degrade it again (not sure if anyone has better settings as I have been running 8 steps 4 high 4 low cfg 1) which produces fine results but not INSANE highly realistic ones. Not using lightning LoRA's at all either as quality is far better without. Tried 30 steps at eueler simple with 12 high rest low, but seemed to produce poor results as well.
For me, the quality is the same as in the original Wan2.2 with lightning lora 4 steps. 2 low 2 high.
You get also different outcome if youre using different umt5. There is also a difference between the the different umt5 NSFW/Uncensored text encoders. Its worth to try around with them. Also this rank64 distill lora makes a big difference how the person moves. I get the best quality only with dpm++_sde/beta.
@lechuck777 Yea def been noticing quality is solid all around, although primarily been testing with NSFW umt5 encoder. Find I don't use lightning LoRA as quality tradeoff is better without it imo (only takes ~3-4 min for me without it for a 5 sec gen at 8 steps 4 high 4 low). Some weird artifacting was happening with res samplers for me (which always seem to be best for highest, passable real almost at 30 steps 12 low rest high), but will test more with dpm samplers.
Can anyone make a gguf version? (better Q4_K_M.gguf)
I also want to do it, but unfortunately I can't do it……
https://colab.research.google.com/drive/1JFRdqy1EdsZNU6STYnf3ft2a4DUHvKXm?usp=sharing
Colab gguf converter based on city's code, uses fp models uploaded to hf and yields a downloadable gguf, works within free account limits without gpu (though may timeout at the last step, a rerun should then succeed). Tested on wan 2.1 finetune, so I'm guessing it should work with 2.2.
So, the answer is, anyone who can run a colab. Probably.
Details
Files
wan22RemixT2VI2V_t2vLownoiseV10.safetensors
Mirrors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Wan2.2_Remix_NSFW_t2v_14b_low_lighting_v1.0.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.