Skip to content
Images

MiniMax-H3-Image-Exp-Comfy

full MiniMax-H3-Image port and experimental LoRA for ComfyUI

3 min readupdated

Creator t8star has put together a complete, unpruned ComfyUI conversion of the MiniMax-H3-Image image generation and editing model, bundled with a toolkit for experimenting with multi-step sampling. The main 66.29 GB checkpoint is converted from the native format to ComfyUI while preserving its original mixed BF16/FP32 precision and all AdaLN layers (without dequantizing from INT8). On top of that, the author trained an experimental LoRA (a lightweight add-on for the base model) called direct MLP100 and built a custom H3ImageSingleHead node to smooth out harsh texture artifacts when switching from the standard 6-step pipeline to 50 steps at CFG 3. This release is built for ComfyUI users who want to run base MiniMax-H3-Image at full quality or test an alternative sampling mode.

Key features

  • Full, unpruned checkpoint: minimax_h3_image_full_bf16_comfy_qkvfix.safetensors (66.29 GB) keeps the original BF16 + FP32 weights and supports the native 6-step pipeline.
  • Experimental texture LoRA: the experimental_direct_mlp100_rank16.safetensors adapter (172.07 MB, rank 16) is trained on 100 MLP modules across 50 layers to tone down harsh, "carved" artifacts on skin, fabric, fur, and watercolors.
  • Dedicated node for multi-step sampling: the custom_nodes/h3_image_single_head folder merges the 6 PDD weight heads into a single weighted head directly in memory without modifying the base checkpoint file.
  • Ready-to-run bundle: the repo also includes pre-built JSON workflows (for both 6-step and 50-step runs), the Qwen3-VL-32B text encoder in INT8 (27.14 GB), and an FP16 VAE (1.17 GB).

What you can do

  • Generate and edit images natively: load the workflow from workflows/native_pdd6 to run the original 6-step generation (2048×2048) and photo editing (2048×1344) in ComfyUI with no extra add-ons.
  • Test softer textures at 50 steps: wire up H3ImageSingleHead and direct MLP100 (strength +1, Euler sampler, 50 steps, CFG 3) for portraits, fabrics, and illustrations with smoother surface rendering.

Example

Node connection order in ComfyUI for the experimental 50-step mode (you must use LoraLoaderBypassModelOnly instead of the standard LoRA loader, and place H3ImageSingleHead right after loading the model):

UNETLoader -> H3ImageSingleHead -> LoraLoaderBypassModelOnly -> CFGGuider
Read moreDetails
This repository is an unofficial research port based on MiniMaxAI/MiniMax-H3-Image and does not completely remove PDD distillation.
Author's test rig: RTX 4060 Ti 16 GB VRAM and 128 GB RAM — running the 66 GB weights requires high system RAM and offloading.
In manual testing across 10 fixed prompts, direct MLP100 passed 8/10 (improving textures, though it can shift composition and fine details); using the newer checkpoint inside the rejected folder is not recommended.
The MiniMax-H3-Image Community License restricts usage in the EU, UK, South Korea, and the US, and sets separate rules for commercial services.
Examples
Discuss Telegram 11 4.1K

Will it run?

Only with RAM offload

Even 4-bit is larger than your VRAM. Some layers can stay in system RAM (llama.cpp, ComfyUI), but it will be several times slower.

  • BF16weights 90 GB · ready files100 GB
  • FP8 / INT8recommendedweights 45 GB51 GB
  • 4-bitweights 25 GB29 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

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.

    # t8star/Minimax-H3-Image-Exp-Comfy
    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 t8star/Minimax-H3-Image-Exp-Comfy --local-dir ~/models/Minimax-H3-Image-Exp-Comfy
    # The model is in ~/models. How to run it is in the model card (link above).
  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

  1. MiniMax-H3family root
  2. MiniMax-H3-Image
  3. Minimax-H3-Image-Exp-Comfy

New posts about versions of MiniMax-H3 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 12All on Hugging Face
Merges 3All on Hugging Face

About the model

t8star/Minimax-H3-Image-Exp-Comfy
Author
t8star
Category
Images
Released
Posted

On NeyroLearn

Hugging Face

4.1K11

Ask AI about this model

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

Comments