ComfyUI-AMD-MiniMax-H3-Image

ComfyUI-AMD-MiniMax-H3-Image

ComfyUI node for MiniMax H3 Image text-to-image and multi-image editing through the AMD gateway. It follows ComfyUI's built-in MiniMax H3 Image partner node, limited to what the gateway serves.

Install

Search for ComfyUI-AMD-MiniMax-H3-Image in ComfyUI Manager, or install it from the registry:

comfy node install amd-minimax-h3-image

To install from source instead:

cd ComfyUI/custom_nodes
git clone https://github.com/wangxunx/ComfyUI-AMD-MiniMax-H3-Image.git

Restart ComfyUI. No extra Python dependencies are required. ComfyUI 0.44.0 or newer is needed.

Configure your API key

Register at https://h3.oneclickamd.ai to get a key, then use either method below. This is a one-time setup per machine and applies to every workflow:

• Save the key to h3_gateway_api_key.txt in the ComfyUI user directory
• Or set the H3_GATEWAY_API_KEY environment variable

MiniMax H3 Image needs a personal API key from the Account page. The gateway's shared key is refused with a permission error for this model.

The file is read on every run, so dropping it in takes effect without restarting. The environment variable is read from the server process, so changing it needs a restart.

If the key is missing, the node fails immediately and tells you the registration URL and the exact file path it looked at.

The gateway serves its image and video models from one account, so a personal key already configured for ComfyUI-AMD-MiniMaxH3 or ComfyUI-AMD-Qwen-Image-2.1 works here with no further setup.

To point the node at a different gateway, set H3_GATEWAY_BASE_URL.

Windows, with the ComfyUI Desktop app

Use the file rather than the environment variable. Desktop launches from the Start menu and does not inherit variables you set in a PowerShell or Command Prompt session, so setting one there has no effect. Making it stick would mean setx or the System Properties dialog plus a full restart of Desktop.

Desktop keeps the user directory at user inside its install folder. That folder moves between Desktop versions and depends on what you picked during setup, so the reliable way to find it is to run the node once with no key: the error names the exact file it looked for. On one Windows install it is

C:\Users\<you>\AppData\Local\Comfy-Desktop\ComfyUI-Installs\ComfyUI\ComfyUI\user\h3_gateway_api_key.txt

Create the file there, paste the key, and save. Notepad's default encoding is fine.

Two Windows details: if File Explorer hides known extensions, New > Text Document already ends in .txt, so do not type the extension again. And do not create the file with echo your-key > h3_gateway_api_key.txt in Windows PowerShell, which writes UTF-16 that the node cannot read.

Node

| Node | Inputs | |---|---| | AMD MiniMax H3 Image | prompt, aspect_ratio (six ratios), seed, images (0-3 reference images) |

One node covers both tasks. Connect nothing to images and it generates from the prompt alone; connect one to three reference images and the same node edits or combines them. A prompt is always required, up to 8,000 characters. Extra image_N sockets appear as you fill the existing ones.

Refer to the references in the prompt by connection order: @Image1, @Image2, @Image3, the same convention as ComfyUI's built-in MiniMax H3 nodes. Referencing an image that is not connected fails before anything is sent.

Every aspect_ratio renders at about one megapixel, for example 1024x1024 or 1344x768 for 16:9. It applies to edits too: the output is reframed to it rather than keeping the shape of the reference images. Steps and guidance are fixed by the gateway, and negative prompts are not supported, so the node has no inputs for them.

The same prompt, settings and seed give the same image. seed changes after every run, like KSampler's; set its control to fixed to keep a seed. When editing an image you generated, use a different seed from the one that made it: the same noise sharpens the image on every pass.

Each image costs 4 credits. Text to image and single-image edits usually take 15 to 30 seconds. The first request at a resolution the backend has not served before takes noticeably longer because it compiles convolution kernels on the spot; that path is slow, not failing.

Differences from the built-in MiniMax H3 Image node

| | Built-in partner node | This node | |---|---|---| | Models | MiniMax-H3-Image, MiniMax-H3-Image-Highspeed | MiniMax-H3-Image | | Size | 2K, 4K or custom, ratio can be auto | six fixed ratios, about 1MP | | Reference images | up to 16 | up to 3 | | Prompt enhancement, revised_prompt output | yes | no | | Seed | re-run trigger only | reproducible |

Example workflows

In ComfyUI, open the template browser from the left sidebar, scroll to the EXTENSIONS section, and pick ComfyUI-AMD-MiniMax-H3-Image. Both workflows run as-is once your key is configured, and each includes a note explaining the key setup.

| Workflow | What it does | |---|---| | AMD MiniMax H3 Image Text to Image | Generates from a prompt | | AMD MiniMax H3 Image Edit | Generates a street portrait, then restages it as a snowy neon night by referring to it as @Image1. Replace the first node with Load Image to edit your own photo. One run costs 8 credits |

Your API key never enters a workflow

The node deliberately has no api_key widget. Widget values are serialized into workflow JSON and embedded into output image metadata, which would leak the key as soon as a workflow or a generated image is shared. Reading the key from the environment or from a file outside the graph keeps it out of both.

License

MIT. See LICENSE.