Creators test MiniMax H3 Image for game concept art

• Character, setting and lighting can come from separate images, as ai_artworkgen demonstrated in a three-reference test.
• Creators reused an initial image for new camera angles and matching concept sheets in tests by Artedeingenio and DrSadek_.
• H3 Image has moved from early-access 4K tests to an official November 11 release, with open weights, Hailuo access and an API.

DrSadek_'s dungeon prompts get wonderfully fussy about stone colour, proportions and matching moss. MiniMax's model card describes roughly 16-megapixel output generated directly, without an upscaler.

In a test by ai_artworkgen, three reference images each had a separate job:

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

The prompt assigned those roles explicitly:

place [@]Image1 inside the world shown in [@]Image2. candid close-up of her at the control station, use the lighting from [@]Image3

The follow-up compares the lighting-reference version with the raw result.

Artedeingenio generated a sci-fi walker, then reused the original image as a reference for different camera angles. The creator reported maintaining the vehicle's design across perspectives in the early-access test.

An isometric dungeon became the reference for three follow-up concept sheets in DrSadek_'s experiment:

• Architecture: A wall section, iron-gate doorway, intact pillar, broken pillar and flagstone floor section.
• Props: A burning tripod brazier, heavy chain, floor lantern and iron shackles.
• Sarcophagus: The central object isolated from its surroundings, with instructions to preserve its silhouette, carved lid, relief decorations and stepped base.

The architectural-kit prompt specifies separate pieces, matching scale and elevated three-quarter views against a charcoal-grey background.

Architectural concept pieces generated from the dungeon reference.

MayorKingAI explored game environments and character design in an early-access Hailuo test; a Western series from AIwithSynthia focused on desert camps, character designs and sunset lighting.

An armored-character concept sheet from MayorKingAI's test.

A Western camp concept from AIwithSynthia.

For a style-transfer experiment, pzf_ai made a Caravaggio-style painting, then asked H3 Image to repaint it in five different art traditions.

A broader sweep from kaigani covers 363 animation styles in an early-access video, with examples generated through the model's website.

MiniMax says H3 Image produces 4096 × 4096 stills in under 10 seconds in its launch announcement.

A separate H3 video experiment from GlennHasABeard used local ComfyUI generation for a 15-second soup clip, then Topaz for the 4K upscale.

• Hosted: MiniMax's release post lists Hailuo AI and MiniMax Code.
• API: The image-generation guide supports 1 to 16 reference images and outputs up to 20 million pixels.
• Local: Model weights are published on Hugging Face.
• Launch pricing: MiniMax advertises 50% off image-generation requests at 4K resolution and above during launch week in the announcement.

The API does not support masked edits or transparent backgrounds, according to the guide. Image-edit requests always apply prompt enhancement, and responses expose the rewritten prompt as revised_prompt.