GPT-Image-2.6 Sunburst
Create detailed images and reference-based edits through an API with independent planning controls and vector output.
MiniMax H3 Image
design.minimax.io
An open-weight image model generates and edits still images using text instructions and multiple visual references.
Visit website
Pricing
Paid
Available on
Web, API, Desktop app, Local deployment
Category
Ai Image Tools
What is MiniMax H3 Image?
MiniMax H3 Image generates and edits still images from written instructions and visual references. Designers can use it for concepts, product compositions and scene exploration through MiniMax Design. Developers can integrate the hosted API or download the model weights under its community license.
Generation and editing share one checkpoint. The downloadable model produces approximately 16-megapixel images directly without an upscaler, while hosted workflows offer their own resolution and export settings. The model card also documents Hailuo web access, desktop application access and several local inference options.
Key features
Text generation and reference editing
Create a new image from a brief or combine visual references for subject, color and layout. Editing requests support up to 16 reference images.
Resolution and export controls
MiniMax Design lists 1K, 2K and 4K output, aspect ratios from 1:3 to 3:1, batches of one to ten images, and PNG, JPEG or WebP export.
Long text instructions
The generation API accepts prompts up to 32,000 Unicode characters. Without prompt enhancement, effective reading is limited to approximately the first 4,096 tokens.
Standard and Highspeed API models
Developers can select MiniMax-H3-Image or MiniMax-H3-Image-Highspeed with the same documented parameters and limits.
Downloadable model weights
Local inference options include Diffusers, SGLang, vLLM-Omni, LightX2V and ComfyUI. Setup also requires the separate MiniMax-H3 text encoder.
Who is it for?
Designers exploring product compositions
Combine product, palette and layout references to explore several visual directions.
Game artists developing scene concepts
Use a written brief and visual references to investigate environments or character compositions.
Developers adding image workflows
Evaluate hosted generation and editing endpoints for an existing creative application.
How to use MiniMax H3 Image
- Open the MiniMax Design product page and choose its image creation workflow.
- Write a brief describing the desired subject, style and composition.
- For an edit, add reference images and explain the role of each.
- Choose available resolution, aspect ratio and output settings, then generate.
- Review the result and refine the instructions or references before exporting.
MiniMax H3 Image pricing
Paid
Hosted API access uses token-based billing deducted from a prepaid balance. Standard and Highspeed have different rates. Actual request cost depends on text, reference images, output and prompt enhancement. Check the official pricing page for current rates and offers.
MiniMax-H3-Image Standard API
USD 2.30 per million input-text tokens; USD 1.50 per million input-image tokens; USD 3.80 per million output-image tokens
Usage-based token charges deducted from prepaid balance; no monthly subscription interval
Base Standard API rates. Prompt enhancement contributes input tokens; eligible temporary offers can change the final request charge.
What to know before you use it
- The published community license excludes the European Union, United Kingdom, Republic of Korea and United States from its applicable territory. Its territorial restrictions apply to model works and generated outputs and results. Hosted services have separate terms that must be checked for the chosen service.
- The community license requires prior written authorization for third-party hosted services or when the user and affiliates exceed USD 1 million in annualized revenue. Commercial interfaces must display MiniMax-H3-Image. Model works and outputs cannot be used to improve other AI models, except MiniMax-H3-Image or its derivatives.
- Editing references have file-size, dimension and aspect-ratio limits. Each image can be up to 64 MiB, but the complete request body is also limited to 64 MiB.
- Editing requests are not idempotent: retrying after a client timeout can create another billable result. The documentation recommends disabling automatic SDK retries.
Frequently asked questions
Can I run the model locally?
The weights and local setup instructions are available. Installation requires the image model and a separate text encoder, and use remains subject to the community license, including its territorial restrictions.
Can I disable prompt enhancement for editing?
No. The editing API always applies prompt enhancement. Text-to-image generation provides a control for disabling it.
Is API pricing a fixed amount per image?
No. Charges depend on token consumption. Images with the same resolution can have different total costs because their text and reference inputs differ.