PinkCherry MM H3 (Beta)
this is the fl2va model, first frame last frame. plugging it the ref2va reference workflow will yield poor results likely, use the comfyui ref2va model for that. it wasnt trained on ref2va data
Beta 0.6
im using this workflow https://huggingface.co/Plaguekind/Minimax-H3
** beta improves t2v and motion, but its not perfect. still needs further training
** I post new snapshots on my huggingface periodically for testing. for those that want latest checkpoint we are testing. on beta-0.6 at the moment
** Aug 09 2026 - Still deep in training this model, next update should be a good one with lots of major fixes. but will not likely be ready until later this week
** Aug 08 2026 - received some great feedback/comments on weak areas in the checkpoint, and ill work on targeting those weaknesses with further training. thanks for all the input and please keep it comming so I know what you are seeing/hearing.
** Aug 8 2026 - both unpruned (32GB) and pruned (20GB) versions of 0.5-alpha are uploaded. Wont be any new versions of this model for at least a week. as im working on the ref2va model now
Join our discord: https://discord.gg/GCrr4Cj3D
Loras: any lora works with this base. We have had good luck using pussy synth, riding and coach bates penis lora at about str 0.4 to address the faults in this alpha release base.
I2V/T2V: Your biggest win will be with image to video, not text to video at this stage of training this model.
Model Type: First Frame/Last Frame FL2VA , however it worked fine in the reference workflow for me. Ill train ref model later
Model Training: Full fine tune of fl2va base model
Training Dataset: 5000+ 4k videos of various concepts
Model Training Target: FFN/Attn blocks only
Training System: Custom Training Pipeline
no guardrails or censorship has been bypassed in this model
This is early still in training model I am working on. Treat it as alpha level still. its the first last frame model not the reference model.
Fair Warning: Expect to find dicks on foreheads, exploding breasts and various other issues typical of alpha models. If that's not your thing, wait for beta.
Training Captions used: cock, pussy, vagina, penis, doggystyle, missionary, slut, moaning, orgasm, wet, hairy pussy, spreading, asshole, tight, fucked, riding, cowgirl, anal, reverse cowgirl, pounding, thrusting, hanging tits, fondling, massage, squeezing, dildo, sex toy, penetrate, squats, flexing, bent over, sucking, licking, tongue, blowjob, arousal fluids, jiggling, swaying, sex position, prone, nude, clothing, saggy, breasts, big tits, legs spread, panties pulled to side, cum, semen, creampie, saliva, masturbating, fingers, fingering, eyes closed, perky tits, rubs, lustful, areolas, pink nipples, mouth open, topless, exposed, voyeur, public, bounce, labia, folds, slut, whore, POV, overhead, low angle, high angle, close up, dripping, precum, round ass, amateur, glistening, facing camera, standing sex position, vulva, slutty, tattooed, pounds, from behind, hanging, stroking, jerking, pants down, panties, cock head, cock tip, wet cunt, passionately , rough, gangbang, group fuck, lesbian, threesome, handjob, double blowjob, throbbing, rapidly, rhythmically, wet lips, cleavage, jiggling breasts, cupping, tittyfuck, teasing, erotic, romantic, making out, blindfolded, thick, crotch, deep, steady, aroused, erect, soft flesh, slamming, messy, slapping, pink anus, bathroom
Description
version 0.5-alpha
FAQ
Comments (54)
nvfp4 is a tinniest possible, and run pretty well, i mean, in my gpu [ 4060 ti 8 vram ] is the only one that runs faster in every model type possible.
Have you tested how well this checkpoint generates video in text2video mode? I don't think it has any effect; it looks similar to the basic checkpoint, at least for text2video mode. Is it planned for the next version to be able to create decent videos in this mode?
its still alpha, yes I have tried it. depends what prompt and words. I think at this stage i2v is my preferred go to, but t2v will improve as training progresses. same story as ltx 2.3 was and took 8 versions :)
I think has to happen is I have to do targeted tuning of the weak areas, to really determine where the model is broken T2V and attack it head on, thats the 'beta' part. its work for sure
@sexgod1979 i was testing Kijai's w4a8_pruned 12Gig checkpoint in Wan2GP. it was crazy how fast it is - 1080p,20steps, 4sec = 6min on 5090. and just 22gb VRAM usage. compared to regular rendering with 33b checkpoint/40 steps/1080p it was almost the same quality. with just 12min for 40steps@1080p it has preserved reallife Vladimir Putin in T2V and his voice as well! i think T2V w4a8_pruned is the future...
Reaching perfection
im sure having fun with it thats for sure.
great quality but still get crunchy carrots for audio sometimes.
like during blowjob scenes, unnatural sucking sounds? model is confusing eating with sucking. prompt it the sound in detail until I can fix it
@sexgod1979 right now its during vaginal, thrusting as if trying to make a sound for it. prompting it out hardly works, sadly
@NUGGZ1616 im trying to picture what exactly that sound is. is there typically a sound in real life. remember this checkpoint was only trained on real sex. I dont see many clips in the dataset where there is typically much of a sound like that. Use a lora? its likely just that the model hasn't learned that sound yet, it will just takes longer as this is still alpha and its under trained.
@sexgod1979 its not consistent, so hard to know when it will happen, but I've noticed it with reverse cg positions. Literally sounds like someone chewing
@NUGGZ1616 ive seen it in most situations. might just have been my type of generations
Love the quality of v0.5 but motion is stiff and not good enough for JIGGLE and BOUNCE especially with natural breasts (small or large), it's just very fake-like material and the actions are SLOW I hope you can add more dataset on more NATURAL boobs jiggle / bounce and especially FASTER more INTENSE motion because the native MiniMax-H3 weakness is that SPEED MOTION is not amazing sadly, but everything else is great.
Keep up the good work, I hope v0.6 will have these suggestions! ❤️
what prompt are you using, I bet I can make them jiggle :)
look at this video, the breasts move naturally. Now this checkpoint wont give you surreal boob jiggle its only trained to be natural, use a lora for that. https://civitai.red/images/139148573
another example of boob jiggle (in zero gravity) https://civitai.red/images/139130647
I could probably always hit it with a targeted boob jiggle dataset though. are you using the pruned or unpruned version?
@sexgod1979 Look at the boobs, That's not jiggle, that's super stiff... try it with the NATIVE MiniMax-H3 and you don't even need to prompt JIGGLE because it's a native physics that trained on, what the MODEL trained on is probably a lot of fake-boobs so it's ruin the NATURAL motion and jiggle, don't confused with BOUNCING to Jiggle the fat/skin which goes together.
Same with ASS, it does not JIGGLE by impact or natural motion easily unless you need to prompt it very hard and it's very technical motion, try the native MiniMax-H3 you can get better natural jiggle fat/skin without even mention it, by a motion. something with the model was too harsh to make it stiff.
I appreciate the smaller pruned model. Most things look on point or that I can tell is headed towards a good direction.
My thoughts is that it needs to train on some penis data. H3 is very good with adding life like movements and I find that it can be very unforgiving in certain positions where the penis 'pops' out. Unfortunately whenever that happens the penis looks like a fleshy door knob. Likewise the same problem lends to some issues during some thrusting movements.
I know there is a penis LORA, however I find that using that LORA frequently makes the penis regenerate outside the vagina as if it is desperate to spawn the whole thing instead of supplementing missing parts.
In any case thank you for your hard work so far.
okay, thats fair, ill work to address the floppy dick
Same with female cunnilingus, when the tongue retracts a thick piece of flesh can be attached to the receiver and the giver usually licks it or sucks it and sometimes they just continue with it disappearing back into the tongue and reappearing with the motion. This also happens on baseline model and it might take a dedicated lora to fix this problem.
@SIMJEDI more pussy licking training. awesome, its a price im willing to pay :)
@sexgod1979 I did a few tests and adding a vagina LORA (I used the innie LORA at 0.5 strength) fixes a few things and especially helps a lot with the penis popping out. However I think the lack of penis training is preventing the model from trying to do proper thrusting. Just thought I'd share.
How many steps do you need for the model? it is merged with the Lightv2 loras or the model need the regular 15-20 steps?
I use 20 steps. It is not merged with Lightv2
@sexgod1979 please don't ever merge a turbo lora with this by default. That would be like lobotomizing someone and everyone saying how they are more pleasant to be around. Yet ultimately it would be profoundly sad.
@whitespider9999 I havent merged anything into it, its just the fine tuned weights raw. I dont merge any loras into at all, if people want to use a lora (which is great) then thats for them to decide not me
@sexgod1979 Perfect!! Thank you!
Heading in the right direction. Truly appreciate your efforts sir.
Latest alpha seems to be doing insertions more successfully.
thanks. 0.6 will be a big jump, I just tested penis popping out of vagina haha and putting it back in!
@sexgod1979 That's a nice touch!
what about groping? touching? grabbing? molesting? why no one care about those ones anymore?........
use the term 'fondling' as per the training data below in model description. if youre not getting anything that word than wait for 0.6 to fix (im getting it in v0.6 I just tested it, the man fondles the woman's big tits). groping is used in the dataset but its tiny. massage and fondling was the words I used
lmfao because of VISA and Mastercard and Americans in general
@makiaeveli Yeah, before you speak, maybe you should do some research on where these payment processors are actually based and who's actually fucking with things in general.
@hushicho Visa and Mastercard are based in the United States. Visa Inc. is headquartered in San Francisco, California, and Mastercard Inc. is headquartered in Purchase, New York. [1, 2]
Just cause they get to wrap it in some Aussie or British organization of "saving the children", I don't really think your argument stands. All these people use American dollars. What, cause Trump believes in AI that means anything? Ask all the states that can't use porn without government ID. It's the same in there, too. All the other western governments are only doing what "the most free country in the world" does to their own citizenry.
Also its always incredible when these "do your own research" types give absolutely no reference to what they believe is the true answer. Probably because I'm speaking to another member of the flock -- er I mean, an American.
@sexgod1979 thank you for replying, i'v always use those terms to create hugging caressing scene images or naughty roleplay before the actual "action"
That would be great! Could you include some femdom-related content? When I generate facesitting scenes, the audio sometimes comes out strange.
It seems like most DOGGY STYLE on-four or standing or other cases, is mostly UP and DOWN instead of Forward and Backwards THRUSTING direction, even when suggesting it multiple times in the prompt it's random and insist to do UP and DOWN, probably it was not trained on enough Forward and Backwards thrusting compare to RIDING or bouncing up and down, I hope you can improve that in v0.6 🙏 that will be great!
If she is on all her fours."The man is kneeling behind the woman.thrusting his hips back and forth while holding her hips.he leans foward....
That should be enough.
@hartweizen yeah alpha-0.5-C on huggingface is better at this up and down issue. working on this exact issue though. thanks
@sexgod1979 You rocks! ❤️
Hi, have you seen the ``new GGUF``??
Its called w4a8, its basically the size of INT4 but with 99% INT8 quality and a boost to speed, often TWICE as fast.
https://github.com/comfy-org/comfy-kitchen/pull/90
Just an idea for the future.
/flies away
I couldn't seem to get that working.
in my case, it had seriously bad result every time,
bot comment
one of h3's biggest strengths is it's exceptional video to video abiltiy, it can faithfully reproduce and continue pretty much anything in the style of the original content. no other video model has come close in this area. Here's hoping when you are continuing to explore this model that you find your way to that capability,that would lead to even more wild creations.
no clue, looking at the videos in our discord and on civit looks totally different. youre getting orgasms, penetrations and wet blowjobs in the base model? did you try alpha-0.5-C? I guess keep in mind this is still an alpha version. not beta, and not a fully trained model that is finished. I think in the next week or so it will clearly differ from base model. if you have time give alpha-0.5-C on huggingface a try, would love the feedback.
There’s a lot of slowmo! She almost always generates it. No matter what I write in the prompt.
idea for your research and testing process:
make a workflow that is configured correctly and provide together with the checkpoint. that way you can rule out that your checkpoint is the problem when most of the time i think it is misconfiguration.
Yes, Train Ref2Vid and your soul is mine.