MiniMax H3 came out yesterday. It mimics FiM style and knows a few ponies.

This video is a purely text-to-video workflow. I’m going to be doing a lot of further testing today to push the limits of this model and find out exactly what it knows and what it can do, and I’ll be posting my findings here.

Cutie Mark Accuracy is generally quite suspect across the board. Character accuracy varies by shot but is mostly accurate for most ponies most of the time–if it knows the character.

Characters it knows natively:

Twilight Sparkle
Rarity
Pinkie Pie
Fluttershy
Applejack

Characters it sort-of knows:

Starlight Glimmer
Rainbow Dash

Characters it doesn’t know natively by name:

Sweetie Belle
Princess Luna

Notably, they produce a moderately consistent horse, just not the right one. Princess Luna, in particular, looks kinda neat.

It can do Image-to-Video NSFW out of the box without any jailbreaking or LoRAs. Next tests are gonna be focused on what kinds of pony-specific poses it can handle for sexual positions, if it understands various anatomically correct sexual structures with or without reference images, and how it handles prompts with object manipulation.

Stuff I’ve seen looks really great!

I’m suspecting even with a gguf and turbo LoRA us 8Gb folks will be out of luck :p

My testing has been sidelined by a new project of mine over the last few days. I’m back in the saddle for a moment to post a few further reports about Text-To-Video. I had written up a veritable essay and then my GPU overheated and crashed my PC, so … Summarize time.

General Recommendation: don’t use Text-to-Video for NSFW. At all. Vulvas are misplaced, penises are wrong, anuses won’t draw no matter what you try, and it’s generally awful in every respect except the bizarre accuracy of nipple placement: I couldn’t get it to draw chest boobs without directly prompting for human breasts.

For SFW stuff, its okay as long as you are using characters it knows.

8/8 is Vore Day. I attempted to do vore with it. It doesn’t do it.

This concludes today’s updates.

@Teaspoon
To be fair, I didn’t try very hard, and it might still be capable, but a baseline test suggested it probably doesn’t have a lot of training on vore.