Lora for Framepack F1, trained in Musubi Tuner.
How to prompt:
Experiment with strengths, I've found 1.0 is fine mostly, but if I see weirdness lowering to 0.7 helps. I've not extensively tested all ways of prompting, But I've found the easiest way to get RC is to just follow the pattern of the training prompts, look at the examples on the post (for settings as well).
Although the model was trained exclusively on reverse cowgirl POV, I've found it can do a lot more like cowgirl from the back, insertion, pov cowgirl maybe if you lower the strength. That's just how powerful framepack is, it generalizes so well. A lot also depends on seed, I recommend testing multiple seeds as the result can vary a lot. all examples were genned with sage attention. But good to use teacache and latent preview when experimenting, so you can quickly check for good seeds.
More prompt examples:
```A light-skinned Asian woman with dark hair and a noticeably large ass is straddling a man, facing away in the reverse cowgirl position. She wears only a baseball cap, her head turned slightly to her left. She grinds down slowly, her large ass jiggling with the deliberate movement as his cock slides deep into her pussy. The man's torso and legs are visible below her. The scene unfolds on a bed with white sheets, a grey headboard against a teal wall, and a bright window with white curtains. Two black bedside lamps flank the bed, with discarded red underwear nearby. This is shot from the man's POV.```
```A 3D animation depicts a Caucasian woman with long blonde hair and a medium build straddling a man. Positioned in reverse cowgirl, she looks back over her shoulder towards the camera, her face visible. The main focus is the sexual act; she's riding his erect penis, her pussy gripping his shaft. Her medium ass is bouncing and jiggling with the medium intensity upward and downward movement as she fucks him. Only the man's torso and penis are visible. They are outdoors on a wooden park bench situated on sandy ground, with a large rocky structure and green bushes in the background. The scene is viewed from the man's point-of-view (POV).```
```An ebony woman with dark skin, short blonde hair, and a notably large ass is straddling a man, looking back over her shoulder towards the viewer. She wears a light-colored one-piece swimsuit. The main focus is the fast reverse cowgirl movement; her large butt is bouncing and jiggling intensely as she rides the man's thick penis, which is visibly penetrating her vagina. The man's torso is visible beneath her, and his hand is placed firmly on her jiggling ass cheek. The scene is set outdoors during the day on a grassy patch next to a paved path, surrounded by dense green bushes and trees under a blue sky. Filmed from the man's point of view (POV).```
training details:
4572 steps, but at different latent windows:
3180 steps at latent window 1 (Batch size 8)
1392 steps at latent window 4 (batch size 6)
dataset 158 clips from 4-10 seconds.
Note: If one of your images was used for showcase and you'd like it removed, just let me know.
Description
This version is trained on latent window 1, then further trained on latent window 4.
FAQ
Comments (44)
thank you for this. Do you care to explain how you trained this lora in musubi tuner? what model did you use for training? what specs etc?
I trained it on the Framepack F1 model (you can find training info in Musubi trainer: https://github.com/kohya-ss/musubi-tuner/blob/main/docs/framepack.md). total time was about 20 hours on a H100 (large batch sizes seem to help framepack training).
I can write a small walkthrough of the training process if people would like.
@tunnah you think i can use framepack mode with the diffusion-pipe i already have for hunyuan? just changing the base model?
@K3NK The framepack training isn't merged into the main branch of diffusion-pipe. I trained a couple loras on it but it didn't support batching yet so I moved onto Musubi Tuner, which was really easy to use.
But yeah if you want to use diffusion-pipe anyway just swap the model and read this PR: https://github.com/tdrussell/diffusion-pipe/pull/242
are you using comfy? im starting to think that gradio webui has better support for loras than kijai's wrappers?..
Hey guys, I'm thinking of what lora to train next. I want to see if a smaller dataset with smaller batch size is as flexible. I think the content will be blowjob, but please share any other suggestions for things you need in framepack
I'm not sure, but I think that there are no POV doggystyle Loras for Hunyuan yet, or at least I haven't found one.
Inspired?
@Bbird good idea, I will do a quick bj lora tomorrow, then get onto a POV doggystyle next. I've also not seen many of them
I think you can train a mating press lora because no one has trained it for Hunyuan.
Could you please share your workflow
I am confident that you can create a great POV squatting cowgirl LORA. Because you've proven the quality and reliability of your data set selection with this one. Whatever you make, I'm sure I'll give it a shot!
i NEED ALL POSE !!!!
Hi, I am getting an Error loading LoRA reverse-cow-w4-000004: list index out of range when trying to load this Lora to Framepack Studio F1.
Is it maybe only working in Comfy?
I'm not sure if Musubi LoRA work in Framepack Studio. I Haven't used it myself. There's this UI that wraps the Musubi inference scripts that's pretty good: https://github.com/maybleMyers/H1111
But I think comfy will be the easiest for most people
Same error here...
@sirdrak Hey, sorry it's not working in Framepack studio, would be great if we can ask the maintainer to add support for Musubi (since it's the best training repo for Framepack atm).
Framepack is something that is installed outside of ComfyUi no ? I mean it's like a separate thing or am I missing something ?
To use in Comfy you can setup Kijai's wrapper: https://github.com/kijai/ComfyUI-FramePackWrapper
There's also some T2V nodes you can install if you want to do that instead of I2V.
It works great! Best motion among reverse cowgirl loras. If you make another lora of variable positions, it must be great too. :D
Thanks for posting your gens, They're really nice!
@tunnah If the quality is good, it's all because of your LORA. I'm excited to be able to breathe life into some of the images of hips I've generated in the past. Thank you so much.
I could never get the other cowgirl lora's to not produce artifacts when using I2V and Framepack. This one did it perfectly without changing any workflow settings. Amazing lora!
you can share a workflow?
Hey, There's a couple ways to inference Framepack right now:
1. Comfyui: https://github.com/kijai/ComfyUI-FramePackWrapper (not sure about F1 support, but the lora works with regular Framepack too I've heard) (any default workflow that loads loras should work)
2. Musubi inference Scripts: https://github.com/kohya-ss/musubi-tuner (no GUI)
3. GUI for Musubi Scripts: https://github.com/maybleMyers/H1111 (I've been using this lately for quick testing, works great)
There's also Framepack Studio. that I'm told doesn't support Musubi Lora format right now. But hopefully they will soon.
@tunnah Thank you for sharing. Looking forward to more Fremepack position LoRAs.
From the front view, it seems that the mouth moves randomly. The frequency of opening the mouth is too high. Is there any way to keep the mouth closed?
Hmm unfortunately Framepack seems to always have anime characters talking. Since I didn't train on front view in this lora, it won't change that behavior. You could try negative prompt: "talking", but not sure how well it will work.
@tunnah thanks,test it doesn't work.It's a great pity
@AshCanc that sucks, I will train a pov cowgirl lora soon. So that should fix the mouth movement hopefully.
@ACancd prompt for an specific expression, i havent try too much anime tho..
@K3NK i cant understand/(ㄒoㄒ)/~~
Bravo for these prompts. Your lora didnt work as its not FP studio compatible but i had another cowgirl lora and slightly modded the prompt for the picture i wanted to use and the very first section of the video produced the best results for myself so far.
so used to cancelling the job after 1 or 2 sections to change seeds si t was a nice change.
So thank you very much
May I ask if you have a lora that can be applied to Framepack Studio? Did you train it yourself? Thank you:)
If you switch to the 'develop' branch of FP-Studio and use the new lora loader it is now compatible.
need more NSfw framepack lora !!!!!!
I've had some luck with regular Hunyuan lora, but it's inconsistent which ones will work well.
Hi @huiyu9144 , today I released my BJ lora, please give it a try and see if you like it: https://civitai.com/models/1685979/blowjob-framepack-f1-i2v
Outstanding work. Please do more Loras I will support you if you have a platform.
Thanks! @Afenix today I released my BJ lora, please give it a shot and see how it is: https://civitai.com/models/1685979/blowjob-framepack-f1-i2v
plz do riding front legend
New to this ai Generation stuff. What software do I need to download to use this? I got Comfyui, Framepack, and stable diffusion. If it’s comyui can someone direct me to a Workflow that I can use to apply this Lora. I’m kinda confused how this Lora works and I wanna generate a character using this. Thanks y’all!