Probably cause it will take 1 hour to generate my 10 seconds video, and every second matters.

@tickaestore799 Why don't you use google colab? Then you get a A100, it is not too expensive.

yeah it just did that to me too. :( Maybe with only 1 of the files for I2V would work better, I will have to test.

In reference to video I lowered the strengths to 0.7 and 0.2 and it works a little better, try playing around with that

@kaileykk I am mostly doing I2V but I will try. thanks :)

Missionary only as god intended 😂

@pocketpie okay with missionary, but they can poop then?

@forfreelsd368 You'll have to go on discord for those sort of loras but they do exist I've seen it!

@pocketpie oh, I got one for wan2.2, from some dump on huggingface. Funny, that lora worked extremely good, almost every time picks right hole. I wonder why I keep such things (I test them on some pics with other loras, than just put them in thematic folder and forgetting them, but sometimes really funny to just dig it and remember how much stuff there, and when almost everything from wan2.2 i2v was squeezed... now comes h3, which almost everything do from the box, and I'm like: damn, I have no Idea what to try next, lets try to dig in wan2.2 stuff, and yes, poop was tried also, and... wrong hole. not like I worry too much, I was just excited how simple was to manage that, but without proper anus lora, separated from vagina - many stuff will just not work, as anal sex for example - when dick not already in). Oh yes, h3 can do b3sti4lity (ref2v, I'm not into i2v or t2v right now) better than normal sex, hm... I repeat, I'm not some pervert, but possibilities are really fantastic for someone with a good imagination (and that's not me, sadly).
Discord - not my stuff, I badly write on English, not mention to speak it, but I created couple of "discord loras only" myself for wan2.2, not very good because lack of experience in creating (sometimes - collecting) a good dataset and prompting it correct. Also funny thing: I created wan2.2 i2v lora on single scene where Kratos from GoW3 torns off Helios's head. And you know what? H3 knows how to do that from box, if describe it in colors. Amazing. Me, who created whole (well, only high noise for test) lora for that: wtf, h3?

I2V / ref with 2 loras + 4step lora. This is like a horror movie. :(

@stormstrike52382 in mine wf even without 4step loras those two sometimes broke face consistency and every time broke prompt adherence. Still think it's some wf stuff (used one from this site).
Anyways, I'm still don't get about this model many things, for example how to use bf16 or bf16_pruned version for generations, if white noise all the time I replace int8_convrot...

lmfao like all of them have worked. H3 is more showing which users are literate and which arent

@makiaeveli so you really are crying in every single h3 lora comment section aren't you... do better

@MiraStudio lmfao I could leave a comment in each thread if you'd like. I've commented on most new model loras as they come out since before Hunyuan. "Do better" lmfao. And what, my liege, must I do instead? You could downvote, my liege. Or possibly, people could read the manuals and stop shitting on every other creator? Americans are something else lmfao. Seriously, who's the problem here? "Oh my eyes are so harmed by pixels you were allowed to place here, oh how dare thee!!!" I mean you aren't even attacking the content of my comments. This is the definition of ad hominem, so... do better? See, I was attacking the person's argument. That they were wrong to say all the other loras don't work. You are saying I should... shut up?

@HearmemanAI  Thank you so much!

Lemme know when you find one of those looool

Do you manually prompt your videos or do you have an LLM arrange it? I feed the official prompt docs into the system prompt of a local vision model (gemma4 uncensored in this case) with a small explainer at the top explaining its job, and when I tell it to be a static shot with no camera motion it seems to get the message on zoom. But I couldn't tell you what it's doing to make that happen lol.

@sansmia I used manual prompts. I may have to try the LLM approach.

the bed and the walls 😅😆🤣

Needs normalization in the training data

LOL

Base model is clearly only trained on shaved mutilated vags :p

Problem is, you'll never be able to train the anatomy of a vagina when its covered with pubes.
Now that the anatomy is pretty much perfect, one could do some additional fine-tuning steps to bring in some variation.

I think @HearmemanAI did the right thing by doing this in stages. First stage is recover the anatomy in Minimax h3, second stage is bring variance