MiniMax H3-Image: native 4K at 16.8 megapixels in six sampling steps (three ComfyUI workflows)

MiniMax's H3-Image is unusual for one reason: it does not upscale its way to a big picture. It generates natively at 4K-class resolution - about 16.8 megapixels - and it does it in six sampling steps. This episode tests that claim with three separate ComfyUI workflows, each built for a different job.

Who this is for

  • People who need large prints or crops out of a generated image, not just a 1024px square
  • Anyone doing image editing who is tired of losing resolution when they re-render
  • Teams evaluating H3-Image before committing to the weights

The three workflows, and when to use each

  1. Text to image - AIKSK_H3_Image_文生图.json
    The simplest chain: prompt in, 6 steps of sampling, 4K image out. Six nodes total, so you can read the whole graph in a minute. Use it when the image does not exist yet.
    Online: https://www.runninghub.ai/zh-cn/post/2120808839725920258/?inviteCode=rh-v1111

  2. Single image edit - H3_Image_单图像编辑.json
    48 nodes. One reference image goes in with an instruction, and the model re-renders while keeping the parts you did not ask it to change. This is the workflow for "same photo, different weather / different light / fix this corner".
    Online: https://www.runninghub.ai/zh-cn/post/2120808831056293890/?inviteCode=rh-v1111

  3. 16-image reference edit - H3_Image_16图编辑.json
    Also 48 nodes, but the reference side is a bank of up to sixteen images. That is what you want for character consistency across angles, or for feeding a whole product set into one edit so the result matches all of them.
    Online: https://www.runninghub.ai/zh-cn/post/2120808834377023490/?inviteCode=rh-v1111

What is inside the JSON

  • Text-to-image: 6 nodes, 5 weight references, including qwen3.8_27b_w4a8 for the prompt path.
  • Single and 16-image edit: 48 nodes each, 4 weights: minimax_h3_image_pruned_int8_convrot, minimax_h3_image_vae_fp16, qwen3.5_9b_w4a8_gptq, qwen3vl_32b_minimax_h3_int8_convrot.

No model weights are bundled in the repos. You bring the weights; the workflow carries the wiring, the sampling settings and the conditioning structure.

Known limits

  • Six steps is fast, but it is not a free lunch: hard prompts still need more than one roll, and 4K output means 4K-sized mistakes are visible.
  • The 16-image branch costs memory and time roughly with the number of references. If you only have two images, use the single-image branch.
  • These are image workflows, not video. For motion you need a different H3 pipeline.
  • Resolution is native, so do not stack a second upscaler on top by default; judge the 4K output first.

How to try it without a local install

All three run as online workflows on RunningHub, which is what we used for the tests in the episode - same graph, no GPU required on your side:

Episode links

Where this goes next

The interesting question is not whether 4K native generation looks good in a hero shot - it does. It is whether the 16-image branch holds identity when the references disagree with each other. That is the test worth running yourself.