Skip to content
Images

chroma-h3

видео-диффузия, обученная на статичных кадрах

2 min read

chroma-h3 — image- («stills») fine-tune модели MiniMaxAI/MiniMax-H3 от lodestones. Это полный H3 diffusers-пайплайн, где fine-tune влит в transformer/ в bf16, а Qwen3-VL text encoder, tokenizer, processor, video/audio VAEs и scheduler оставлены без изменений. В метаданных указано около 33,12 млрд параметров.

Ключевые особенности

  • Корневая модель — готовый merged-пайплайн для генерации одного кадра примерно на 1 MP.
  • lora/lora.safetensors — тот же fine-tune как LoRA на base H3 transformer: fp32, rank 256, alpha 256, strength 1; затрагивает FFN в блоках transformer и token-refiner.
  • lora/lora_dmd8_student.safetensors — few-step DMD-студент (12 000 обновлений, дистилляция при CFG 3) для merged-модели этого репозитория: plain Euler, CFG 1, 4 или 8 шагов; shift 3 для stills и shift 12 для видео.
  • lora/lora_vspace_dmd8.safetensors — та же few-step-модель одной LoRA, но для стоковой MiniMax-H3, не для merged-трансформера: внутри уже есть fine-tune и DMD-студент.
  • LoRA rank 256 / alpha 256 обучена только на feed-forward Linears всех transformer-блоков и token-refiner; attention, AdaLN и embeddings не трогали.
  • Во время обучения в image-режиме audio AdaLN срез отбрасывался; в merged-модели он сохранён без изменений.

Пример

Сценарий из карточки: сгенерировать один кадр (1 latent frame) примерно на 1 MP. Для merged-модели нужен classifier-free guidance: CFG ~3 с пустым negative prompt и около 30 шагов Euler при shift 3. Для few-step LoRA — plain Euler, CFG 1, 4 или 8 шагов; shift 3 для stills, shift 12 для видео.

ПодробнееБенчмарков в карточке нет.
Обучение: 13 605 обновлений, глобальный batch 48 изображений, lr 0.0001, flow matching, timestep shift 3.0, aspect-ratio buckets около 1024², 10% empty-prompt dropout.
Ограничение: merged fine-tune больше не ведёт себя как guidance-distilled — нужен CFG ~3 и ~30 шагов Euler; few-step LoRA обучены только на stills, но переносятся на видео.
VRAM и требования к железу в карточке не указаны.
Discuss Telegram 45 65

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # lodestones/chroma-h3
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download lodestones/chroma-h3 --local-dir ~/models/chroma-h3
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to chroma-h3
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • SageAttention 3v2.2.0from October 28, 2025

    even faster on RTX 50 / Blackwell

    FP4 attention for Blackwell. Faster than SageAttention 2, but the authors suggest v2 when accuracy matters.

    How to enable: Build from the sageattention3_blackwell folder of the repo; in ComfyUI via KJNodes.

    Earlier generations (2)
    • SageAttention 2++ — Quantised (INT8/FP8) attention with no visible quality loss. The main speed-up for image and video generation.
    • SageAttention 1 — First version: INT8 for Q·K. Still works, but 2++ is faster and more accurate.
  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (3)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
    • xFormers — The older way to speed up attention, works almost everywhere. Sage or Flash are faster where available.
  • cache-dit (DBCache)v1.5.6from November 20, 2026new release

    1.5–3× at nearly the same quality

    Reuses block outputs between diffusion steps. Supports most DiT models.

    How to enable: pip install cache-dit; one line cache_dit.enable_cache(pipe) for diffusers.

    Earlier generations (2)
    • MagCache — Skips steps by magnitude of change. More accurate than TeaCache, no calibration.
    • TeaCache — The first popular diffusion step cache. Works, newer methods are more accurate.
  • Nunchaku (SVDQuant 4-bit)v1.2.1from January 25, 2026

    ≈3× faster and 3.5× less memory

    4-bit versions of popular image models with almost no quality loss (FP4 on RTX 50).

    How to enable: Ready weights for supported models; in ComfyUI use ComfyUI-nunchaku.

  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

MiniMax H3 prompts

All prompts

Ready prompts with the resulting video: copy and try.

Model family

New posts about versions of chroma-h3 (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

lodestones/chroma-h3
Author
lodestones
Category
Images
Released
Posted

On NeyroLearn

Hugging Face

6545

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments