MiniMax CN I2I Config

The interesting part of MiniMaxCNI2IConfig is what its image input is called: subject_image. This isn't a generic img2img "start from this picture" node - MiniMax treats your input as a subject reference, something to keep consistent across the edit, which is the single hardest problem in AI image work. If you've ever fought to keep a character's face stable between generations, you know why this matters.

It's the I2I sibling of MiniMaxCNT2IConfig: same model choices, same style system, same seed and watermark controls - but instead of size_mode/width/height, you get a subject_image input. Output is a MEDIAHUB_IMAGE_REQUEST for MediaHubGenerateImage to execute.

How the reference works

Under the hood the node does the familiar conversion: your IMAGE tensor becomes base64 JPEG data URLs. But note the shape of what it builds - the payload frames each image as a {"type": "character", "image_file": ...} reference. MiniMax's API consumes that as a character reference for the generation. So this isn't "paint over my picture," it's "generate in this character's likeness," and the prompt steers the scene while the reference anchors identity.

Practical implications: one clean, well-lit reference of the subject works better than a cluttered collage, and the model will chase the character rather than reproduce your exact pixels. If you want pixel-level img2img, this isn't the node; if you want consistent identity, it is.

Inputs that matter

• subject_image - the reference IMAGE. The star of the show.
• prompt - the scene, pose, or change you want.
• model - image-01 or image-01-live.
• use_style + style_type (漫画/元气/中世纪/水彩) + style_weight - same style system as T2I, and it composes with the subject reference nicely: identity from the image, look from the style.
• use_seed / seed, n, prompt_optimizer, aigc_watermark, aspect_ratio, response_format - the shared MiniMax set.

Output: request → MediaHubGenerateImage.

Installing it

Same pack, same drill:

cd ComfyUI/custom_nodes
git clone https://github.com/vantang/ComfyUI-MediaHub.git
/path/to/ComfyUI/python -m pip install -r ComfyUI-MediaHub/requirements.txt

Restart ComfyUI; node under MediaHub/Provider/MiniMax CN/Image. Requirements requests/Pillow/numpy, and a MiniMax CN key (mainland endpoint, separate account from MiniMax Global).

Common gotchas

Quality of the subject image is the whole game - a small or heavily compressed reference gives you a small or heavily compressed identity. And because the input is JPEG-encoded at quality 95 before upload, extreme fine detail in the reference degrades slightly; for a face it won't matter. If results look nothing like the reference, check that you're on the I2I node and not accidentally wired to the T2I one - both output the same MEDIAHUB_IMAGE_REQUEST type, so ComfyUI will happily let you connect the wrong config and produce exactly zero reference images.