Overview
Unlimited generation by plan
Which plan runs each resolution unlimited at launch, at the model’s base settings. Gems means that setting always costs Gems on that plan.
Three or more LoRAs move a run to Pro Plus. On Gems, a text-to-image run costs 15 Gems at 2K and 25 at 4K at launch, and each reference image adds a few Gems; the app shows the price before you generate.
What it supports on Mage
Best for
- Photoreal images with fine detail at a native 16 megapixels
- Long, detailed prompts and complex instructions
- Posters, signs, labels, and other text in the image
- Edits that change one thing, and composites that bring many images into one scene
- Styles from MiniMax H3 Image LoRAs
Writing prompts
MiniMax H3 Image does best with a detailed English description of the whole image: the subject, the setting, the composition, the light, the style, and any text that should appear, in quotes. Short prompts work too, with less control over the result. Reference images areimage 1, image 2, and so on, in order: your uploads first, then the Characters you mention, then the References you mention. Write @image1 for an upload and @handle for a Character or Reference, as on other models. For an edit, say what to keep from each image, say what changes, then describe the image you want.
Example prompts to start from
Using LoRAs
Pick a MiniMax H3 Image LoRA under Model → LoRA, or bring your own on the Import page by choosing MiniMax H3 Image as its model. Only LoRAs import for this model, not checkpoints. See LoRAs for how weights and stacking work.Limitations
- No negative prompt, step count, or guidance scale: the model is distilled to six steps with its guidance built in. To leave something out, describe the image without it.
- 2K is the smallest size; there is no 1K option.
- Each reference image adds time and Gems, so an edit from many images takes longer than text-to-image.
Use via API
MiniMax H3 Image is available through the Mage API as theminimax_h3_image architecture. Its fields, tokens, defaults, and price are on its API reference page.