Introduction

What if your favorite video model could also be a really good image model? lodestone, the fine-tuner behind Chroma, Kroma and Kromoss, just released chroma-h3 on Hugging Face. It's an open-weights text-to-image fine-tune of MiniMax H3: the H3 video DiT runs with a single latent frame and is trained on text-image pairs. You can run it locally today with diffusers. For ComfyUI, there's already a community conversion of the LoRA.

The repo went up on October 31 at 01:41 UTC. An 8-step DMD LoRA followed at 16:08 UTC. The license is the MiniMax H3 Community License, same as the base model.

What's in the repo

lodestone ships the fine-tune in three forms, so you can pick whatever fits your setup:

  • Merged model (root): the full H3 diffusers pipeline with the fine-tune merged into transformer/ in bf16, about 66 GB of transformer shards. Everything else is the stock H3 base: Qwen3-VL text encoder, tokenizer, processor, video and audio VAEs, and scheduler
  • lora/lora.safetensors (about 2.9 GB, fp32): the same fine-tune as a rank-256 LoRA on the base H3 transformer. Apply it at strength 1
  • lora/lora_vspace_dmd8.safetensors (about 5.5 GB, fp32): an experimental few-step version, DMD update 8,700 of 12,000. It already contains the fine-tune plus a distilled student

The fast LoRA has one important quirk. Apply it to the stock MiniMax H3 transformer, not to the merged chroma-h3 one, because the fine-tune is already baked in.

Settings that work

The full fine-tune isn't guidance-distilled anymore, so treat it like a classic CFG model:

  • Full model or main LoRA: one latent frame at about 1 MP, CFG around 3 with an empty negative prompt, about 30 Euler steps, shift 3
  • DMD 8-step LoRA: plain Euler at CFG 1, 4 or 8 steps, shift 3 for stills
  • Video bonus: the DMD LoRA was trained on stills only, but lodestone reports that it carries over to H3 video. Use 4 or 8 steps with H3's usual shift 12, and motion stays coherent

How it was trained

  • LoRA rank 256 / alpha 256 on the feed-forward layers of every transformer block and the token-refiner blocks. Attention, AdaLN and embeddings were left alone
  • 13,605 updates at a global batch of 48 images, one pass over the data, learning rate 0.0001, flow matching with timestep shift 3
  • Aspect-ratio buckets around 1024², with 10% empty-prompt dropout
  • Captions are an even mix of booru tags, short Midjourney-style summaries, long captions and scene graphs, so you can write prompts in whichever style feels natural

Here's a starter prompt in the long-caption style:

A quiet night market street after rain, warm paper lanterns reflecting in shallow puddles, a young woman in a lilac raincoat holding a steaming cup of tea under a striped awning, soft aqua neon signs blurred in the background, gentle film grain, 35mm photograph, shallow depth of field

Running it in ComfyUI

Community user crombobularBoop published chroma-H3-comfyui on October 31 at 19:02 UTC. It's a key-renamed bf16 LoRA (about 1.4 GB) in ComfyUI's format, along with the convert_chroma_h3_lora.py script that made it. Load it on top of your existing MiniMax H3 setup at strength 1 and render a one-frame clip to get a still. At the time of writing, the repo's notes also list a converted DMD8 file, but only the main LoRA has been uploaded so far.

Worth knowing

  • The training credit goes entirely to lodestone. The ComfyUI repo only converts key names and layout
  • The MiniMax H3 Community License carries over, including its territory limits, so check whether it covers your region and use case before you ship anything commercial
  • The DMD LoRA is labeled experimental, from a run that hadn't finished yet (8,700 of 12,000 updates), so expect updates
  • This is early: the repo had only a handful of likes when we found it, so samples and community recipes are still thin

Conclusion

chroma-h3 is a playful reminder that a video DiT is also an image model if you squint at one frame. Fast 8-step stills that can slip right back into motion make it a lovely experiment for anyone already running H3 locally. Grab the LoRA, set CFG to 3 and see what H3 dreams about when it holds still.

Curious how the base H3 moves before you go stills-only? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=chroma-h3-lodestone-minimax-h3-image-fine-tune-dmd-8-step

Original Source

Hold that frame and have fun!

—Aurelia ♡