Hailuo releases MiniMax H3 Image weights and API access

• MiniMax H3-Image ships as downloadable weights, an API and hosted tools through Hailuo_AI's release links.
• Native 4K took seven seconds using six steps on an RTX PRO 6000 in hellorob's reported test.
• Up to 16 references can feed one scene; koldo2k's timed demonstration completed a multi-reference composition in about a minute.
• Open weights carry territorial and commercial restrictions, despite Hailuo_AI's “open for everyone” framing.

The restaurant branding experiment started with a fake phone photograph made in GPT Image 2.6. A community two-step LoRA already produces roughly 1 MP previews in 1–1.5 seconds. The weights license restricts where generated images can be displayed, too.

MiniMax's model card describes one checkpoint for text-to-image generation and reference-based editing. It generates roughly 16-megapixel images directly, without an upscaler:

• Square: 4096 × 4096.
• Widescreen: 5472 × 3072 at 16:9.
• Ultrawide: 6272 × 2688 at 21:9.

MiniMax promises under ten seconds for a square 4K image in its announcement. The seven-second result in hellorob's local-generation thread specifically names an RTX PRO 6000.

Hosted early-access timings were different: koldo2k reported roughly 21 seconds per 2K image across aspect ratios, with timers included in the demonstration.

DrSadek_ uses a neat art-direction trick in his seven published prompts: the identity sheet controls branding, while the restaurant photograph supplies location content.

The sequence specifies:

1. Identity sheet, 4:3: One stacked wordmark, an SH monogram, four labelled colour swatches, typography samples and a repeating pattern.
2. Homepage, 3:4: A mustard-yellow burger hero, product listings, one restaurant photograph and ordering buttons. The requested output is a flat webpage screenshot.
3. Printed menu, 2:3: Food photography, exact descriptions and prices, with brand-guide annotations excluded.
4. Staff uniform, 4:5: A cream T-shirt, brown apron and restrained branding, photographed inside the reference restaurant.
5. Packaging, 4:3: A burger box, fries carton, drink cup and takeaway bag, each with a specified treatment.
6. Loyalty card, 3:2: Front and back artwork, including exactly nine stamp circles in a three-by-three grid.
7. Storefront, 4:3: An exterior inferred from the interior, with matching signage, materials and window graphics.

The shared results show the identity sheet, menu and homepage; the prompt sequence extends the concept into the other formats.

Artedeingenio first generated a sci-fi exploration walker, then reused the original image to produce multiple camera angles. His follow-up prompts specify four additional presentations:

1. CAD-style blueprint: Front, side, top and isometric views, with hidden lines and mechanical centre lines.
2. Exploded view: Separated cockpit, chassis, legs, hydraulics and armour, connected by alignment guides.
3. Mechanical cutaway: Removed panels revealing the cockpit, internal frame and one leg's mechanisms.
4. Design board: A hero render, orthographic drawings, component close-ups and material studies.

All four briefs request vertical 9:16 images, including the engineering-style drawings.

DrSadek_ generated a dungeon, then used it as the reference for three separate concept presentations:

• Architecture: A wall section, framed iron gate, intact pillar, broken pillar and floor section.
• Props: A brazier, chain, lantern and shackles.
• Sarcophagus: An isolated presentation retaining the original silhouette, stonework and iron restraints.

The prompts specify matching materials, consistent elevated viewpoints and separately spaced objects on charcoal-grey backgrounds.

ai_artworkgen assigned three reference images distinct jobs:

1. Image 1: Character sheet.
2. Image 2: Environment.
3. Image 3: Lighting reference.

The published prompt explicitly maps those roles: “place [@]Image1 inside the world shown in [@]Image2. candid close-up of her at the control station, use the lighting from [@]Image3”.

For a four-panel Pop Art story, pzf_ai created both characters first and reused them as references for the remaining panels.

egeberkina reported changing an uploaded image's textures while retaining its composition. The instruction was “retexture the image attached based on the JSON aesthetic below”, followed by these controls:

• Style: Photorealistic 3D render.
• Material: Transparent, iridescent glass.
• Surface: Smooth and polished, with reflections and refraction.
• Lighting: Studio HDRI, a top-left key and ambient fill.
• Palette: Blue, green and purple iridescence.
• Background: Black, with a vignette and no texture.
• Post-processing: Chromatic aberration, glow, high contrast and sharp detail.

techhalla published film-and-print tests with prompts in the image alt text, including handwritten captions, comic lettering and text printed on clothing.

Game-interface concepts also feature dense lettering and recurring characters:

Fantasy quest-interface concept from gizakdag's tests

The comparisons and commercial mockups expose some rough edges:

• Game scenes: H3-Image looked flatter than GPT Image 2.5 to underwoodxie96 in a same-prompt comparison.
• Portraits: The same tester preferred GPT Image 2.5's body proportions but H3-Image's photorealistic style in a portrait comparison.
• Ingredient copy: The printed-menu prompt explicitly excludes beef from the Mushroom Burger, while the homepage mockup describes it with a beef patty.

ComfyUI supports both local weights and API-backed Partner Nodes, according to its launch post. Its native workflow has several specific behaviours:

• Fixed sampling: Six steps at guidance 1, with no negative prompt.
• Default size: Automatic text-to-image sizing produces 2048 × 2048; editing follows the first image.
• Reference order: Image slots determine the reference numbers used in the prompt.
• VAE connection: Connected references condition both image latents and the text encoder; without the VAE, only text-encoder conditioning remains.

The unquantized downloads are substantial. MiniMax lists 68.65 GB for the image checkpoint and another 66.7 GB for the separately downloaded text encoder in its model card.

That encoder is shared with the separate H3 video model. GlennHasABeard's 15-second soup demonstration uses a text-to-video workflow in ComfyUI, followed by Topaz upscaling to 4K.

_akhaliq released a community pDMD LoRA that reduces denoising from six evaluations to two. The package includes both a standalone LoRA and a pre-merged ComfyUI checkpoint, alongside a ZeroGPU demo.

The project's measurements put a full 4096 × 4096 render at roughly 17.4 seconds versus 25.1 seconds for stock six-step sampling on one H200. Its optional heavyweight refinement pass pushes the total to about 50 seconds; the roughly one-second interactive previews run at a much smaller resolution.

Hailuo's launch links cover Hugging Face weights, the API, Hailuo and MiniMax Code.

MiniMax's API guide documents:

• Generation: POST /v1/images/generations, with up to ten images per request.
• Editing: POST /v1/images/edits, accepting one to sixteen references.
• Prompt limit: 32,000 characters, but only about the first 4,096 tokens take effect when prompt enhancement is disabled.
• Unsupported features: Masked edits, transparent backgrounds and preview images.

Text-to-image applies prompt enhancement by default; image editing always applies it. The actual expanded instruction is returned as data[].revised_prompt.

MiniMax's pricing examples list these estimated costs for one 4096 × 4096 text-to-image result:

| API model | Regular example | Launch-discount example |
| --- | --- | --- |
| Standard | $0.06879 | $0.03439 |
| Highspeed | $0.13001 | $0.06501 |

The 50% discount applies to delivered images containing at least 16.6 million pixels and ends November 17, 2026, at 16:00 UTC. Actual charges vary with text tokens, including prompt enhancement, and reference-image inputs.

Hailuo calls the model “open for everyone to build on”, but the Community License imposes specific limits:

• Excluded territories: The EU, UK, South Korea and US are excluded from the default license grant.
• Generated outputs: The territorial restriction also covers using, distributing or displaying outputs outside the applicable territory.
• Commercial authorization: Annualized revenue above $1 million, or providing a hosted service to third parties, requires separate prior written authorization.
• Product attribution: Commercial products or services using the model must prominently display “MiniMax-H3-Image” in their interface.

MiniMax says in its licensing FAQ that its API remains globally available. Organizations in excluded territories can apply for a separate license, with authorization dependent on the deployment scenario and compliance safeguards.