Quite often the problem with training is not following particular setups. If you can follow trussell or misubituner docs, the most important thing you can do is garnering good data AND properly describe it.

dngstn32 +1 for the JoyCaption mention, I feel like it's the only photo-llm I've been able to actually use

dngstn32 how do you use it? Is this a workflow? My only exposure to LoRa training so far is with Kohya.

Bleedy use the trainer? With the command line/powershell. They are describing a literal "work" "flow". Like you know, doing things. Lmfao. You can setup JoyCaption with a workflow, there are custom nodes that support it. JoyCaption is a whole other 8gb model. You could use it in command line like every other model or whatever. Just read the trussell docs lol.

dngstn32 yes, good link. I felt like adding it would just be more confusing, but this is exactly what I was referencing. One would use this in Step 3, but could also edit it to help with step 2 and 4.

makiaevelio543 Gotcha - yeah I was interested in actually creating a LoRa for video and was trying to figure out where to begin. I'll read through the article a bit more. Thanks guys!

dngstn32 i am trying to find out if i can train using my 4090 using your settings, how much vram did your training need?

The worst is knowing how often the models need to move around on your card, but most are already depending on that reality. So yes, it does take some extra time unloading/loading, but depending on your card, subsequent runs of the same prompt should go slightly faster -- than even Wan 2.1. It's impressive honestly. The first time I tried it when it launched I would agree, at least a little. Once I found a decent workflow (playtime_ai has great ones, simple enough to add your flavor of upscaling, grain, etc.), this model has been my go-to.

dngstn32 not if you use gguf

@dngstn32  Insane how some people will never fucking be happy with what exists. Even when it's what's best.

Buy a few billion dollars of GPUs, code your own better algorithms, and do it yourself.

Or just fuckin' enjoy the (literally) best AI video mode ever created by humanity since the Big Bang. It's not that hard.

@Baka_Oppai says the guy who posted 0 video... The MoE design in Wan 2.2 employs two specialized expert models to handle different stages of the video denoising process. Gguf runs on consumer-grade hardware with low VRAM, but it NEEDS both high and low models, components of the MoE architecture. Don't state what you don't know mate...

dngstn32 Yes, it looks good but the quality of your loras is MUCH SUPERIOR in terms of movements, skin realism and camera angles.

All the images just got oddly deleted or something. If you pull the metadata you might get the sample images. I had the workflow, but perhaps its too similar to K3nk's, or idk. It seems normal to me.

This video depicts a couple engaged in a sexual act on a wooden bench in a university locker room. The girl is very skinny, with fair skin, dark brown long hair in a high ponytail, and small, natural jiggling breasts, is lying on her back with her legs raised and spread apart. She has a navel piercing and is looking up with open eyes, blowing a bubblegum bubble with a face of disinterest. She wears a red and white cheerleader's uniform with the skirt pulled up, the front of her uniform reads "U of A". Her skin glistens with a sheen, likely from oil or sweat. The man is massive, muscular, tall, and bald, and has very dark black African American skin and wears a red polo shirt with a coach's whistle dangling from his neck. He is positioned between her legs, holding her right leg up with one hand and penetrating her anally, fucking her ass deeply and slowly with his massive veiny black penis. His face is not visible in the video. The lighting is bright and even, highlighting the sheen on their bodies. The overall style is explicit and realistic.

The camera zooms out to reveal the rest of the cheerleading squad standing around, cheering and clapping with red pom-poms and megaphones. A banner in the background reads "UNIVERSITY OF ANAL"

@Kiefstorm You have to add "roast beef" to the negative prompt.

Lol that comment ha ha I seen some of the guys on other comments talking about the vagsacks that had me losing it. I'm like, thankfully I'm not the only one getting the vagsacks.

same, on any penetration lora, no mater what ive tried it dont work.

@goodstrk looks like a curse.