Notes on film, new tools and the decisions behind the work.

Tools / Analysis

MiniMax H3-Image can combine 16 references. The licence is another part of the brief

MiniMax’s image model takes several visual references into one generation. Its API documentation and downloadable-weights licence tell a more useful story than speed alone.

A cobalt vase with a red flower beside stone and colour samples. Conceptual Google Flow illustration, not an H3-Image output.
DU / SIGNALS · S/111Conceptual illustration · Google Flow; not H3-Image output

A bare room, nine pieces of furniture and one instruction: put them together. In a selected example on MiniMax’s model page, those ten inputs become a furnished interior. The interesting part is the handoff. A creator can bring several specific objects to a single composition instead of describing each one from memory.

MiniMax introduced H3-Image in an official X post dated November 10 and a blog announcement dated November 11. It generates images from text and edits images using up to 16 visual references. The company advertises a 4096 × 4096 image in under ten seconds; the published statement does not establish the hardware, queue conditions or repeatability of that number. The model’s downloadable weights and MiniMax’s hosted API also come with different questions about access.

A room built from ten inputs

The image on the model card places a small view of the empty room beside nine furniture references. Below, a sofa, striped ottoman, wooden cabinet, two tables, chairs and a blue shelf appear in a single room with matching sunlight. It is a compact example of a familiar art-direction request: retain the setting, bring in chosen objects and make them feel as if they belong together.

The image is a selected, downscaled showcase, not a record of rejected attempts. It does not show whether the same object would stay faithful across several variants, or whether the furniture is available as separate editable layers. Those distinctions matter when the picture has to survive client revisions rather than simply work as a still.

The API exposes generation and editing as separate requests. Editing accepts one to 16 input images; the number is an input limit, not a promise that every object or instruction will be preserved. The same model checkpoint is described as supporting both text-to-image and reference-based editing.

Sources: MiniMax [2] · MiniMax [3]

X / DEMO

MiniMax introduces H3-Image

Loading post from X…

Credit / MiniMaxOpen original 

Official product announcement with media. Company claims have not been independently tested.

What the 4K claim actually measures

A 4096 × 4096 image contains about 16.8 million pixels. That is a square, not a standard widescreen frame. In the API’s 4K area setting, MiniMax lists 5472 × 3072 for 16:9; the area is roughly the same even though the width and height change. The distinction matters when comparing generation time or cost across differently shaped outputs.

MiniMax’s model card says these images are generated directly, without an upscaler, in six sampling steps. The card describes a diffusion transformer derived from its H3 family. Those specifications are useful for understanding the approach, but they do not by themselves show how much fine detail remains usable in a large print or how quickly a request finishes under a particular workload.

The sub-ten-second figure comes from the company’s announcement. It does not identify a full timing protocol in the statement we reviewed: hardware, number of concurrent requests, reference inputs and queue time all affect what a creator experiences. We have not run the model or timed a request.

Sources: MiniMax [1] · MiniMax [2] · MiniMax [3]

The edit has a second author: the prompt expander

An editing instruction is not necessarily the exact instruction the image model receives. MiniMax’s API documentation says prompt enhancement is always applied to image edits, and the revised prompt is returned with each result. For text-to-image it is enabled by default but can be disabled when the request specifies its dimensions or aspect ratio.

That returned text is worth keeping with the images. When an edit changes an unexpected detail, it gives a reviewer another piece of evidence: what was requested, what the service rewrote and what the image actually contains. The documentation provides the mechanism; we have not submitted an edit and cannot say how a specific prompt would be rewritten.

There are also practical boundaries. The documented API accepts JPEG, PNG and WebP output, but lists masked editing and transparent backgrounds as unsupported. A PNG file extension alone does not make an isolated object transparent. A designer who needs a precise cutout or a local mask should plan another step or verify whether a different tool provides it. These limitations refer to MiniMax’s documented API, not every possible implementation of the weights.

Sources: MiniMax [3]

Downloadable does not mean unrestricted

MiniMax has made the weights available under its own community licence. The agreement defines an applicable territory that excludes the European Union, the United Kingdom, the United States and the Republic of Korea. It says interested users in excluded territories can contact MiniMax about a separate licence. Section V.4 also prohibits using, reproducing, modifying, distributing or displaying the covered works or their outputs outside that territory. This is a term of the community agreement; it does not by itself establish the terms of the separately operated API.

The licence requires prior written authorisation if the annualized revenue of the licensee and/or its affiliates exceeds US $1 million, or if the licensee provides a hosted service to third parties. It does not define how affiliate revenues are combined or how annualized revenue is calculated. These terms concern the works covered by the agreement; they do not, on their own, establish where MiniMax’s separately operated API can be purchased or used. Its service terms and regional access need their own review.

The model card lists the BF16 model package at 68.65 GB and identifies a separate 66.7 GB text encoder from the H3 repository. These are listed component sizes, not measured GPU-memory requirements. Loading the model can involve other components and runtime overhead; without a measured setup, neither number promises it will fit a particular workstation.

Sources: MiniMax [4] · MiniMax [2]

Count the iterations, not just the pixels

The hosted API bills image work using input text, input image and output image tokens. MiniMax supplies examples rather than one universal per-image cost. On its pricing page, a standard 4096 × 4096 text-to-image request is illustrated at $0.03439 during a 50% promotion, versus $0.06879 before the discount; actual cost can vary with input tokens. The offer covers qualifying requests with a final delivered image of at least 16.6 million pixels and ends before 17 November at 16:00 UTC. Check current pricing and availability before committing a project.

For a meaningful creative test, choose a brief with references that can conflict: a room, a chair with a distinctive silhouette and a material that must remain recognisable. Save the inputs and returned prompts, generate several variants, then compare placement, shape and unintended changes. Record the usable results, not only the fastest request. This is a proposed workflow, not a test performed for this article.

H3-Image puts a large reference set and a high-resolution output into one documented workflow. Whether that saves work depends on the details after generation: how faithfully it resolves the references, how many attempts are discarded and what must be rebuilt for the final delivery.

Sources: MiniMax [5] · MiniMax [3]

SOURCES & METHOD

The record behind the article

Sources checked on 11 November 2026. Product claims are attributed to MiniMax’s announcement and documentation. We did not run the model or conduct comparative tests for this article. Cover: A cobalt vase with a red flower beside stone and colour samples. Conceptual Google Flow illustration, not an H3-Image output.

  1. MiniMax H3-Image announcement MiniMax · 2026-11-11
  2. MiniMax H3-Image model card and showcase MiniMax
  3. Image generation and editing API guide MiniMax
  4. MiniMax H3-Image Community License Agreement MiniMax · 2026-11-10
  5. MiniMax API pricing overview MiniMax
  6. MiniMax H3-Image announcement on X MiniMax · 2026-11-10