Open-sourcing Banodoco Hivemind, a live dataset of 1M+ Discord messages featuring talented artists/engineers discussing open image and video models. Designed as an agent-driven skill to equip your agent with the tacit knowledge it needs to push open models to their limits.

I'm glad to announce that I've finally finished the Minimax H3 version of my crossview-warp model finetune. For those who don't know what it is, it's a model that lets you regenerate your existing video clip from a new camera angle or with a new camera movement. Link to model👇

Yue2 allows for custom voice training on a production-class music model for the first time. Sample by gueykhalamari trained w/ AI-Toolkit using Samuel L. Jackson’s voice, link below.

seitanism injecting frames into specific positions with h3 - see how it handles 1 frame every 1 second:

(gen by orabazes)

Wan Animate 2 allows for highly versatile motion transfer in a wide variety of styles. Examples by Kijai:

On out-of-distribution requests, H3 can produce highly coherent visuals and audio - example by @nebsh83:

How your 6 y/o potato GPU reacts when you try to run Minimax (gen'd by _kruss_)

Inpainting based on a reference w/ Minimax H3. Examples by AbleJones, feat. @PurzBeats:

76 five-second clips exploring different animation styles with MiniMax H3. All generated locally on a six-year-old GPU by the_shadow_nyc:

The next generation of vid2vid will go to some magnificent, disturbing places (generated locally by Dawnll using Minimax M3)

Historically, the lag between a frontier video model launching and a broadly comparable open-weight model running locally has averaged roughly nine months. If that trend holds, we should expect a Seedance 2-level local model sometime between August and December this year.

Now available as a Huggingface dataset! 1M+ Discord messages feat. years of nerds pushing open-source AI art models to their limits - available for your training/research. Link: huggingface.co/datasets/Bano…