🤠 Cowgirl Concept LoRA (Video | WAN2.2 5B)
Hey everyone!
This is my first video LoRA, trained for cowgirl-style grinding/riding animations using [WAN2.2 5B]. I followed the official YouTube tutorial and used Ostris’ AI Toolkit — surprisingly easy and works on windows.
I used the same settings shown in the video.
🔧 Training Details:
Model Target: WAN2.2 5B
Dataset: AI-generated 3-second videos (72 frames each) made with WAN2.1
Includes hair physics and breast motion, though the dataset wasn’t perfect
Training Steps: 3000 total, but best results around step 2500
Training Time: ~7 hours on an RTX 5090
VRAM Usage: peaked around 16GB
Workflow: Used the default ComfyUI workflow, just added a LoRA loader — worked flawlessly
Trigger Word:
c0wg1rl
💡 Notes:
Could likely train faster with proper batch settings (I didn’t use one this time).
While WAN2.2 5B doesn’t match the quality of the 14B model, it’s much faster and more resource-efficient, and still delivers good results.
Definitely room for fine-tuning and dataset refinement in the future.
Let me know what you think, and feel free to share any tips!
Hope this helps others get started too 🎬🐮
Description
FAQ
Comments (36)
Isn't the real point of the 5B how it compares to Hunyuan? I would assume it is better in every way, but is that the case?
No sure I never tested Hunyuan but the reselt are nice for sure with the basic workflow.
hunyuan is more flexible in terms of size and length, 5B is pretty strict about the resolution
Hunyuan Video is better and totally uncensored (it can do genitals and sexual positions without loras 'out of the box')
sirdrak ok then why it didn't have success like wan?
the i2v and certain LoRAs were key
iPiKo It's like vfx_ai says... The problem with Hunyuan Video was that it didn't have an i2v version for a long time, and then, Wan 2.1 appeared with a very good i2v version, so all the attention was for Wan 2.1 since that time.
we need work flow for low 6gb vram
did you only include a trigger word or did you caption as well?
Does the 5B need less vram to train?
I had only 16GB
and see 14B need like 24GB++
You can do it with less resoulotion or less frames.
I dropped your video into Comfyui and the workflow has more than 100 nodes - image generation, LLM instructions, penis enhancement, nipples enhancement, etc. Do I really need all those nodes? I tried to use your Lora in much simpler workflow but it does not generate penis, vagina or good penetration. Can you please share your simple workflow?
Thank you makes time to train💓🤏 ！ 5b is super handy, keep practicing a few more, no hesitation needed~
Thank you very much for your work. I would like to ask if the AI Toolkit currently supports training Wan 2.2. I noticed that the sample batch scripts have not been updated with relevant content. Could you clarify whether you trained it using the GUI or by configuring a batch script? Additionally, the author previously mentioned on GitHub that "WebP currently has issues." Has this been resolved, and is it now possible to use WebP video format as training material?
I used the UI, I used mp4, for now it only support the 5B model and 2.1, yes webp doesn't work it get detected as pictures not video.
This lora doesn't work in t2v, only in i2v I guess? Also : you use ollama to generate the prompt. What model do you use that would be uncensored?
It's qwen 2.5 instruct
wan2.2_ti2v_5B_fp16
c0wg1rl in your prompt. Work well with img to video
yerul563 i have not tested the effect of the trigger word to be honest not sure if it is needed or not, I just use LLM and let it do the job.
I’ve not been able to use it for T2V.
Great Work !! Thank u.
Do you plane to make the same Lora with buttocks in front?
Anyone would mind to share his 2.2 5B workflow? Theres very few of them, so far they have been a mess, either because its extremely slow, the video is a blurry mess or they crash my gpu, and yet with 14B i can achieve 6 seconds video and between 480-720p for 140 seconds generation time on average with a 16gb 5070ti... i must be doing something wrong.
Look for vGen workflow currently the best 2.2 5b I think in my recent lora posts I included a link.
请问wan2.2的lora是不是都是区分了高噪声和低噪声两个lora文件的？你目前只有一个，是怎么使用的？请教！
No only one model.
thank you for the work you're a champion and more thank you for taking the initiative to do his 5b models for Wan 2.2
Not having any success getting this working with on-site generation with 2.2 5B. No penetration, the girl usually ends up doing a dance instead. Issue with the on-site workflow?
I think the start_image needs to be correct. You can download a sample picture and try again.
very great Lora,very powerful!,thank you!
Nice work!!
I really wish there was one specifically for (fingering and masturbation). as there aren't any of these for this model (Wan 2.2 5B).
The model is great and relatively lightweight, making it perfect for fast generations on consumer GPUs.
Details
Files
wan2.2_5b_c0wg1rl_72_000002500.safetensors
Mirrors
wan2.2_5b_c0wg1rl_72_000002500.safetensors
wan2.2_5b_c0wg1rl_72_000002500.safetensors
wan2.2_5b_c0wg1rl_72_000002500.safetensors
wan2.2_5b_c0wg1rl_72_000002500.safetensors
wan2.2_5b_c0wg1rl_72_000002500.safetensors
wan2.2_5b_c0wg1rl_72_000002500.safetensors
wan2.2_5b_c0wg1rl_72_000002500.safetensors
wan2.2_5b_c0wg1rl_72_000002500.safetensors