Lora for Framepack, trained in Musubi Tuner.

VERSION 1.0_xtreme Some important things for inference:

1. Try different seeds, best way to get different motion

2. Use full descriptive prompt (use an LLM if you don't want to type more). Hunyuan needs long prompts

3. Try latent Window 4, just like seed this can sometimes give very different (and nice) results

4. Try high steps, I use 35 steps now and it seems a lot cleaner. even more steps and higher resolution (if you have VRAM) is preferred.

How to prompt:

The model is pretty strong, but as always seed matters a lot. Strength 1 can be good or sometimes go a bit crazy. Step count is also important for good results, if things start breaking down, increase the steps and maybe reduce model strength to 0.7-0.8, combine with other blowjob loras too if you like. All examples are on strength 1 and 25 steps to show model limitations.

Directing with prompts isn't perfect, but some things that might help you customise results are.
fixed camera,

handheld camera, (if you want it to move),

sucks tip,

sucks half penis,

sucks full penis,

engulfs penis,

strokes penis while sucking,

hollows cheeks,

holds deepthroat, (to create a pause at the bottom)

slow/steady/fast movement, (choose one to get desired results I like fast in most cases)

smooth movement, (can helps stop jerky moves)

As always follow the prompt examples, here are some more to show the structure:

A 3D scene of a light-skinned woman with long redhair, wearing around green earrings and white top, is kneeling and performing a blowjob. she strokes the penis with her hand and her head bobs up and down with rapid fast movement as she takes the half the penis into her mouth with each descent, her cheeks hollowing as she creates immense suction, her hand strokes in sync with the sucking. The scene is set in an indoor bedroom. The perspective is POV.

A woman with shoulder length, red hair , wearing a blue skirt with her huge breasts out is kneeling is performing a blowjob. her lips wrap around the shaft, her cheeks hollowing as she applies suction during rapid fast movements, her head bobbing up and down energetically as she takes the full penis into her mouth with each descent. The scene is set in a jungle with trees and green shrubs. The camera is fixed. The perspective is a side view close-up.

A dark-skinned woman with shoulder-length, straight auburn hair, prominent made-up eyes, and full lips, wearing dangly red earrings, is positioned in a bathtub, leaning forward performing a blowjob. She begins with short and steady movement of her mouth up and down, hollowing her cheeks, her lips wrapped tightly around the head of the penis. Her hand, adorned with long, colorful, pointed nails and a ring, firmly holds the base of his shaft. As the action intensifies, she transitions into deepthroat with fast, energetic up-and-down movements of her head along his full length, until pausing at the bottom and looking into the camera. The scene takes place in a white jacuzzi-style bathtub filled with murky, greenish water; jets are visible on the side of the tub, and a white towel rests on the tub's edge. The man's lower torso and groin are visible. The perspective is POV with an unsteady camera.

3D scene depicting a woman with fair skin and green hair adorned with a red mask featuring white accents and a decorative symbol, her breasts partially visible, is performing a blowjob. She sucks the full penis with slow movement, her lips wrapped completely around the shaft as her head moves up and down with smooth, deliberate motions, hollowing her cheeks intensely on each downward stroke to create powerful suction as she takes the entire length into her mouth repeatedly, the tight seal of her lips sliding along the full shaft in a steady rhythm. The scene takes place in what appears to be an indoor setting with wooden flooring visible in the background, creating a warm-toned environment against a darker backdrop. The camera remains fixed throughout the scene. The perspective is POV, showing the action from the receiving partner's viewpoint looking down at her as she performs.

A light-skinned brunette woman with her hair pulled back, visible tattoos on her chest and shoulder, wearing a red top and small stud earrings, is performing a blowjob on a man whose erect penis extends from the right side of the frame. She begins by taking half of his thick shaft into her mouth, her cheeks hollowing inward as she creates suction, her head moving up and down with slow, smooth movements while her lips grip tightly around his girth. As the action progresses, she maintains the same depth, continuing to suck half of his penis while her cheeks remain hollowed, but her pace increases to steady, smooth movements, her mouth working more rhythmically up and down his length as her lips slide along his shaft. The background shows a bedroom wall decorated with various anime posters and photographs, with blue neon lighting visible creating ambient illumination. The scene is captured with a fixed camera from a close-up side view perspective.

developments notes:
I took a lot longer with this than I had initially planned. I did a quick BJ lora 2 weeks ago and did not really like the result, it was ok, but not consistent. I also tried a bunch of experiments, with trigger words and further training on the same dataset, and none of it works. Turns out a good dataset with strong captions is the most important thing.

I've not extensively tested this model, but have played around enough to like it. It consistently gets some nice motion and can be pretty dynamic/creative. By no means is it perfect and sometimes can be a bit strong, especially on animated content (change seeds if you don't like the stretch lips effect, I'll train it down in the next iteration).

I've found Blowjob content is extremely varied and no lora is gonna have the perfect motion for every situation. I'm calling this version 1.0 but might further train the model (maybe on more data) to try clean up the wildness. Blowjobs have been much harder to get right than penetration lol, which I didn't expect.

Future plans:
I plan to train a regular FP blowjob model next (pretty much the same dataset, but just on normal FP) it has some strengths I've noticed that F1 is missing. This model works with regular Framepack, but not to a good enough standard.



training details:

5325 steps, latent window 1, dataset 76 clips (11 epochs)

3080 steps, latent window 4, dataset 163 clips (38 clips overlap) (7 epochs)

videos from 5-15 seconds at 480p

Note: If one of your images was used for showcase and you'd like it removed, just let me know.