MiniMax H3-Image: API Pricing, Editing and Local Setup

Use MiniMax-H3-Image for text-to-image or image editing. It is a separate checkpoint from the MiniMax H3 video model. Pruna P-H3-Image is another image pipeline with its own eight-step workflow and reference limits; its measurements do not describe the new official six-step release.

The hosted API supports up to 10 output images per request and up to 16 reference images for edits. It accepts PNG, JPEG and WebP output, but does not support transparent backgrounds or masked edits. For a cutout or a localized mask operation, choose a model and workflow that provide that feature.

Choose either an explicit WIDTHxHEIGHT size or an aspect ratio plus area level. The API’s 4K area means about 4096 × 4096 pixels. At 16:9, its documented 4K example is 5472 × 3072. Custom output area ranges from 655,360 to 20,000,000 pixels.

Create a MiniMax platform account with an API key and usable balance. Send text-to-image requests to /v1/images/generations and edits to /v1/images/edits. The example requests one 2K-area image. To edit, add images with accessible URLs, uploaded MiniMax file IDs or supported inline data, and describe which numbered reference supplies each feature.

Prompt enhancement is on by default for text-to-image and always on for hosted image editing. Read revised_prompt when you need to understand what the service used. For repeatable text prompts, reuse the revised prompt with enhancement disabled and specify a fixed size or aspect ratio.

Inspect each returned data item: a request can return HTTP 200 when only some images succeed. Save successful outputs to your own storage and handle per-image errors separately. Automatic retries can generate a second paid result, so inspect the original request outcome before repeating it.

Standard MiniMax-H3-Image lists $2.30 per million input text tokens, $1.50 per million input image tokens and $3.80 per million output image tokens. MiniMax-H3-Image-Highspeed uses the same text rate with $3.00 for image input and $7.60 for image output. These rates are token prices, not flat per-image fees.

The pricing page’s standard examples list $0.02333 for one 2K square text-to-image result and $0.06879 for a 4K square result before promotion. One 1K reference plus a 4K output has a listed example cost of $0.07312. Actual totals depend on text tokens, including prompt enhancement, as well as reference count and image sizes.

The platform’s 50% discount applies to the total request cost when the final delivered image has at least 16,600,000 pixels. The offer ends at 16:00Z on November 17, 2026. Its 4096 × 4096 example qualifies: the displayed standard text-to-image estimate becomes $0.03439. Smaller 2K results do not qualify. For 100 matching 4K text-to-image requests, the displayed examples imply about $3.439 during the offer or $6.879 at the regular estimate.

The model card lists a 33.1B-parameter diffusion transformer. Its 68.65GB image repository includes the transformer and image VAE, but the full 66.7GB H3 text encoder is downloaded from the separate MiniMax-H3 repository. Together those listed downloads are roughly 135GB before the optional prompt expander and runtime files.

The optional MiniMax-H3-Image-PE prompt expander adds a listed 17.9GB NVFP4 download. Disk sizes are not a complete VRAM minimum: precision, offloading, resolution, references and concurrency affect memory at runtime.

For the vLLM-Omni profile in the official README, a B200 or B300 can hold the full set of components at about 135 GiB; an H200 setup offloads the text encoder; a four-GPU Ulysses setup is another listed route. These are runtime-specific examples, not a universal minimum for other implementations.

The official card links SGLang, vLLM-Omni, Diffusers, LightX2V and ComfyUI recipes. Follow the selected runtime’s current recipe and matching components. The local server contract differs from the hosted API: it fixes six inference steps and guidance scale 1, rejects negative prompts and documents local seed handling.

Use the official image-generation or editing template, update ComfyUI, and download the image-specific transformer and VAE plus the matching text encoder. The template starts at a reduced 4MP resolution to limit memory use; raise it toward the documented 16MP sizes after checking the complete workflow on your hardware.

Quantizing or replacing the text encoder changes the pipeline as well as its memory budget. In a November 11 Hugging Face discussion, fengjunwei favored the BF16 model with the full encoder. Pvok’s separate comparison found better image detail but weaker character likeness than the H3 video route. Those are qualitative user reports without a matched hardware-and-latency test.

For reference editing, check the features you need to preserve: face, product geometry, lettering and camera position. Keep the reference images, prompt, precision and output size with the result. A faster small preview is useful for composition, but it does not establish quality or memory use at the final resolution.

The weights use the MiniMax-H3-Image Community License Agreement, not Apache 2.0. Its applicable territory excludes the European Union, United Kingdom, South Korea and United States. The license invites people in excluded territories to contact MiniMax for a separate license.

Separate prior written authorization is required if you or your affiliates exceed $1 million in annualized revenue, or if you provide a hosted service to third parties. The hosted-service condition applies independently of the revenue threshold. Commercial products also have a model-name display requirement, with additional distribution and acceptable-use conditions in the agreement.

For a hosted API or Hailuo application workflow, evaluate that service’s account terms and access separately from downloading and operating the weights. A third-party quantization or workflow does not by itself remove the original model’s licensing conditions.

Choose the official image checkpoint when its six-step, up-to-16-reference workflow fits your task and license. Pruna P-H3-Image has documented quantized desktop configurations and separate eight-step measurements; compare complete memory and retained detail rather than substituting one pipeline’s speed into another’s budget.

For video and audio generation, use the MiniMax H3 page. For transparent image assets and repeated targeted edits, compare Seedream 5.5 and its available service routes. Start with the output format and editing behavior you need, then compare the cost of accepted results.