Here are some of the current samples with the default prompts in AI Toolkit generated using the current state of the model. The AR model has clearly grasped how to create images that match the prompts.

you don't need to pay too much attention! full attn is overkill and should be used sparingly calibrated krea2 from full attn on 48 heads into 12 heads full attn while the rest is scanline attn (band diagonal attn). scanline attn is block sparse by design (see the gif)

Captioning music is difficult. The LLM has to be greedy because you need to be precise, but this makes it susceptible to fall into loops. Lyrics often repeat over and over, which makes it want to loop more. You can penalize this, but that breaks real loops.

Added initial support to fine tune LLMs in AI Toolkit with a focus on Omni models. First one added is Qwen2.5-Omni-7B and I trained a LoRA for it to better caption and transcribe music files. You can use these LoRAs in the dataset captioner along with default cloud ones.

AI Toolkit now supports training LoRAs for YuE2-3B, an amazing music model. Support would not be possible without the tokenizer and help from @sin_ceriously . Thank you! Also special thanks to @Machinedelusion. More in 🧵

If I have a car and I am reckless with it and I kill someone because of it. I have two options. 1. Go to jail. 2. Say cars are too dangerous and we need to slow the production of cars down! If you have a model that you were reckless with and it illegally hacked servers.....

Hey, I just released a Style Transfer LoRA for MiniMax H3 (ref2va) that restyles a whole video from a single reference image. Totally experimental. Two ways to use it: drop in a style image and point at it with , or skip the image and just name a style in the prompt for looks the base model already knows. Don't do both in the same prompt the model follows the text and ignores the image. The source video conditions on guide latents, so it locks frame-for-frame instead of "reinterpreting" the motion. The style image rides on the reference slot, which is what lets you transfer a look you can't name. Heads up: a reference with a person in it can leak that identity into the video. Pick one without a prominent figure. Rank 64, ComfyUI workflow included. #MinimaxH3# #H3# #StyleTransfer# #ComfyUI# #Workflow#

I decided to have fun this weekend and train my own image diffusion model from scratch. Training on a single local GPU. This is 2 days in. It is a purely experimental arch that is working surprisingly well for its size and how it works. There is no text encoder. I am using the qwen3 tokenizer and dumping those token embeddings straight in next to image tokens with a time 0. Only self attention. It is pixel space. I am just doing patch 16 straight into the transformer. I am not doing any projection layers. The hidden size is 768 so the blocks work directly on the pixels not a projected image embedding. There were visible patches, but I added a shuffler to fix it. It does an alternating unpatching, mirror padding 8 on all sides and then re patching so the cross seams land in the center of patches, runs through a block, then depatch, crop and repatch every other block. It cleaned up immediately when I added it, something only possible without projection. There is no VAE, no text encoder, and it is tiny and super fast. Really tiny. It is 441 million parameters total. 116 (1/4) of those is the token embedding table. To put that in perspective, SD 1.5 was 983 million (te and unet) Will it be able to make coherent images? Not sure. I will keep training it for a bit though to find out. It seems to be understanding the basic concept of the prompts and if you squint really hard, it kind of looks like the other AITK samples with the same prompts.

I removed all the critics from my turbo time continuous training method and switched to using the model itself as an evolving semi frozen critic. I also switched to Video Sparse Attention for a speed boost. This is the result from training over night. These are all 1 step.

At Reactor we ❤️ open-source. We teamed up with @haoailab to ship an infinite live stream powered by FastVideo’s FastH3: And we’re open-sourcing everything!

How to Train a MiniMax H3 Ref2VA LoRA with AI Toolkit In this tutorial, I train a vid2vid LoRA on @MiniMax_AI MiniMax H3 that makes everyone super jacked 💪 Links and more in 🧵