chroma-h3

An image ("stills") fine-tune of MiniMaxAI/MiniMax-H3: the video DiT run with a single latent frame, trained on text-image pairs.

• Root: the full H3 diffusers pipeline with the fine-tune merged into transformer/ (bf16). Every other component (Qwen3-VL text encoder, tokenizer, processor, video/audio VAEs, scheduler) is the unchanged base.

• lora/lora.safetensors: the same fine-tune as a LoRA on the base H3 transformer (fp32, rank 256, alpha 256, apply at strength 1). Keys are the H3 transformer's own module names (<module>.lora_A / <module>.lora_B); target regex (transformer_blocks|token_refiner\.refiner_blocks)\.\d+\.ff\.net\.(0\.proj|2).

• lora/lora_dmd8_student.safetensors: few-step version of this repo's model: the data-space DMD student alone, to apply on top of the merged transformer/ here (fp32, rank 128, alpha 128, strength 1; block attention

  • FFNs). Sample with plain Euler at CFG 1, 4 or 8 steps (shift 3 for stills, the usual shift 12 for video). Distilled (DMD, 12000 updates) from this model at CFG 3, on stills.

• lora/lora_vspace_dmd8.safetensors: the same few-step model as ONE LoRA for the stock base (DMD update 12000). Apply it to the stock MiniMaxAI/MiniMax-H3 transformer, not to this repo's merged one: it already contains the fine-tune above plus a data-space DMD student distilled from it at CFG 3. Sample with plain Euler at CFG 1, 4 or 8 steps (shift 3 for stills, the usual shift 12 for video). fp32, scales folded into lora_B (metadata alpha: rank), so apply at strength 1 with scale 1; rank 384 on the block FFNs (fine-tune + student), 128 on block attention (student), 256 on the token-refiner FFNs (fine-tune). Trained on stills only, it also carries over to video (4/8 steps, coherent motion).

Training

• LoRA rank 256 / alpha 256 on the feed-forward Linears of all transformer blocks and the token-refiner blocks; attention, AdaLN and embeddings untouched.
• 13605 updates at a global batch of 48 images (one pass over the data), lr 0.0001, flow matching with timestep shift 3.0, aspect-ratio buckets around 1024^2, 10% empty-prompt dropout; the audio AdaLN slice is dropped during training (image mode), the merged model keeps it unchanged.
• Captions: an even mix of booru tags, short / brief / Midjourney-style summaries, long captions and scene graphs.

Use

Generate a single frame (1 latent frame) at ~1MP. The fine-tune no longer behaves as guidance-distilled: use classifier-free guidance (CFG ~3 with an empty negative prompt) and ~30 Euler steps at shift 3.