~~Work in progress~~
This is all still work in progress. Far from perfect atm.
The core is built as a base, its what Ill be building all additions on top of, It's Rank 64 only because my merger can only merge into rank 64 at the moment for HYV1.5.
Load order for general:
Hunyuan video 720p T2V FP16 > NSFW Core @ 0.25 > Oral @ 0.25 > Male Enhance @ 0.25 > Refiner @ 0.25
Load order for oral:
Hunyuan video 720p T2V FP16 > NSFW Core @ 0.25 > Oral @ 0.4 > Male Enhance @ 0.25 > Refiner @ 0.25
Due to training limitations, I can only train motion at rank 8, so visual enhancement is trained by itself as images at rank 32 and merged into the core.
This is my first motion lora so be nice ;), prompting may seem.. excessive to a lot of people, but, hunyuan likes scripted prompts, so i trained on that same style. ill post examples further down, its also in all sample clips.
This DOES work with lightx2v, which is why its split atm, you can run higher strengths without lightx2v, but it varies per part.
im also still training all this, the core is a merge of 3 separate runs do to OOM.
The oral is trained on-top of the core, but can run alone.
The refiner is built to break character overfit and to add more varied details.
The core is built NOT to be perfect, its built to have an understanding, and to be trained on top of, its a lot how I treated hyv1, Which is loading 32+ loras at 0.1 strength, then i can prompt anything.
Training set info:
Due to vram limits, I am forced to train motion at 256x256 @ 49 frames @ rank 8. The frames are 16fps 3 seconds clamped into 24fps 2 seconds, this allows this to produce outputs at 16fps (depending on lightx2v strength). - Currently epoch 80.
Each clip has 3 high res images, frame 0, frame 24, frame 49, trained @ 768x768 @ rank 32, this is to help clear up visuals and enhance whats trained during motion since motion training is so low. - Currently epoch 40.
The refiner is trained on several thousand images @ 512x512, all a combination of close-ups of anatomy, different body styles, different angles, etc. Also contains high res images from clips (just trying to avoid the refiner from breaking the core) Those are tagged in a mix of scripted prompts, and sd style prompts. - Currently at epoch 10.
Oral is trained pretty much the same, just, focused. - Epoch 30.
~~Prompt examples~~
Missionary:
([Style: Cinematic, Hollywood Movie, XXX, Pornographic, Porn, slow, passionate, romantic, missionary sex scene.]:1)
([Scene: Bedroom, Sideview, woman on her back in bed with man laid ontop of her.]:1)
([Camera: side-view of woman.]:1)
([Woman: 18 year old woman in her bedroom, 18 year old woman is petite, busty, large breasts.]:1)
([Man: 20 year old man, laid ontop of woman.]:1)
([Action: woman is laid on her back in bed with ler legs spread with man laid ontop of her moving his penis in and out of her vagina, having missionary sex.]:1)
POV Missionary:
([Style: pov, first person view, XXX, Pornographic, Porn, missionary sex scene.]:1)
([Scene: pov of woman on her back in bed as you thrusts your dick in and out of her vagina.]:1)
([Camera: pov, first person view.]:1)
([Woman: 30 year old woman, 30 year old woman is petite.]:1)
([Action: you thrust your penis in and out of woman's vagina.]:1)
Front view Prone:
([Style: cinematic, hollywood movie.]:1)
([Scene: woman laying on her face down on bed, man laid ontop of her with his chest against her back.]:1)
([Camera: front-view of woman.]:1)
([Woman: 20 year old woman, 20 year old woman is petite.]:1)
([Action: front-view of woman as man is kneeled behind her, fucking her in the ass.]:2)
Standing/seated:
([Style: cinematic, hollywood movie, horror, grungy, dark, creepy, sex, sittingsex.]:1)
([Scene: public restroom, nasty restroom, gas station restroom, dirty restroom, grungy restroom. young woman seated on restroom counter leaned back with her head against mirror with her legs spread around man, man stands between her legs, man's hands squeeze her waist.]:1)
([Camera: side-view of woman.]:1)
([Woman: 20 year old woman, young woman, petite, skinny, tiny breasts in tight black dress, tattoos, fishnet stockings, dark make-up.]:1)
([Action: side-view of young woman as man thrusts his penis in and out of her vagina. woman has intense orgasm and her vagina starts squirting.]:2)
Description
Refiner Version 0.5
FAQ
Comments (4)
Also, if anyone else out there is training for hyv1.5 using musubi and would like to train on top of the base, I can provide the musubi lora. the low level motion core is rank 8. (24gb vram limit)
Dont know which version do use for training, but for i2v generation (did not test t2v) the official distilled version from tencent is much worse (blury, less motion, robot like motion etc) as the hunyuan_video_1.5_i2v_480_quanto_bf16_int8 Version (try it out with the Hunyuan_video_1_5_LigthX2V_dpm_rank32 Lora, the difference is huge) from Wan2GP. The Hunyuan_video_1_5_LigthX2V_dpm_rank32 Lora produces better results (even with the official distilled version from Hunyuan 1.5) as the Hunyuan_video_1_5_i2v_step_distilled_dpm_rank32 Lora. In my tests, the 480p Model (same filesize), produces better results (picture detail is slightly worse, but better prompt following, motion, face consisteny and less drifting) as the 720p model (got bubble artefacts around hard shapes with fast movement, like with old hunyuan 1.0) and it's no wonder, both use the same amount of vram and both got the same size. If the 720p model got the same filesize and use the same amount of vram, then stuff had to be cut out. Hunyuan 1.5 is much better (compared to Wan 2.x) in guessing how a face should look when the face looks straight (and the face was looking sideways in the i2v image). Tencent got a ton of expierence with 3D stuff, and it helps to guess faces better.
@bluecrooked985 yeah, i did a lot of testing early on, on initial release.
i found varying results from each model and huge variances dependent on how they are used.
most of my testing/work is done on t2v model, i don't do enough i2v to spend much time with it.
what i personally found, is the 720p model, gives better motion at low res, its what I've used for all the car scenes i did because I couldn't get the 480p model to do the same motion, but the 480p model seems better at longer clips.
my personal setup, and its only just my preference and it's the fastest output in my system, is low res 720p generation (sub 480p generation) at 12-24 steps (even without lightx), then upscaled and 4 step 480p i2v distilled with lightx just to clean visuals (a lot like wan's moe setup, just with upscale in between)
i can't do that with these yet, as i've only trained t2v atm. wanted to figure out all the quarks to t2v before I add in the other layer of I2V requirements.
i can still do it using 480p t2v on second step, i just find the i2v model is better at upscale cleaning then the t2v model, that and the 480p i2v distilled model is small, so it can handle higher resolutions in 24gb vram
on t2v models, in my experience, it feels like the 720p model cut out a lot in temporal to keep the same files size, and of course 480p dropped visual details so its a trade off. but its weird the 720p model at low res (check out 320x240) is so insane with motion, so if you want a chaotic scene, fast moving scene, 720p low res is the best start.
currently building/training pov specifics.
pov lora added, trained on pov blowjob and pov cowgirl, but, can be prompted for other things.
style:, pov, first person, cowgirl-sex or blowjob-sex.
also started training a position tag
position: cowgirl-sex or blowjob sex.
ill work on adding all this data into refiner soon.