MiniMax H3 Image testers report 16 reference images and 4K output

• Early-access testers report up to 16 reference images and 4K output in AIwithSynthia's H3 Image test.
• Reference reuse carries designs between images: Artedeingenio changed a vehicle's camera angle, while DrSadek_ turned a dungeon into matching asset concepts.
• One generated face anchors skincare, catalogue and magazine concepts in DrSadek_'s commercial tests.
• Long portrait briefs fit too: underwoodxie96 reports a 32K-character prompt limit and shares detailed JSON-formatted prompts.

A Caravaggio-style painting becomes a five-tradition repainting experiment. A portrait prompt specifies lens equivalents, camera angles and separate must-keep and avoid lists. Meanwhile, a 269-image review tests Pruna's P-H3-Image, a separate image model built on H3.

The early-access reports describe three limits:

• Up to 16 reference images: MayorKingAI says they used all 16 in a style experiment.
• Up to 4K output: the resolution claim appears in AIwithSynthia's testing thread.
• Up to 32K characters per prompt: underwoodxie96 says the allowance accommodates their full JSON portrait brief.

Artedeingenio used the first generated image as a reference to request additional views of the same sci-fi walker, reporting consistent vehicle design across perspectives.

DrSadek_ generated an isometric dungeon, then reused it as the reference for three follow-up concept sheets. This is the keeper: one environment supplies the visual brief for its constituent parts.

The prompts published in the images' alt text specify:

1. An architectural kit, preserving masonry style, proportions, stone colour and restrained moss:
2. A prop sheet, matching the reference's forged-iron treatment:
3. An isolated sarcophagus, retaining its silhouette, carved lid, relief decorations and stepped base while removing the surrounding room.

The architectural prompt also fixes presentation: separate objects, clear spacing, matching scale, elevated three-quarter views and neutral lighting. Those instructions turn the scene into a coordinated set of concept images.

DrSadek_ used the first generated model image as a reference for three commercial treatments:

• A skincare advertisement.
• A fashion catalogue shot.
• A magazine cover.

The skincare result also exposes a limitation: the bottle's smaller label copy is garbled.

pzf_ai started with a Caravaggio-style painting, then asked H3 Image to repaint it in five different art traditions.

In a follow-up, pzf_ai singled out the canvas detail. MayorKingAI took a different route, describing the use of 16 reference images to explore artistic styles in their testing thread.

underwoodxie96's published portrait prompts separate camera instructions, identity constraints and unwanted results into explicit fields:

• First portrait: a vertical 3:4 frame, high overhead angle and approximately 24–35mm-equivalent perspective, with deliberate grain and imperfect overexposure.
• Second portrait: a roughly 2:3 frame, approximately 35–50mm-equivalent perspective and a camera tilted downward by 25–40 degrees.
• Identity preservation: the second prompt requests the reference's facial proportions and natural asymmetry, and explicitly prohibits horizontally flipping the composition.
• Constraints: the first prompt separates must_keep, avoid and negative_prompt lists, including instructions against plastic skin, distorted fingers and an excessively polished studio look.

MiniMax's H3 Image preview still describes the model as upcoming. It leaves the public model identifier, pricing and final licence terms pending, and identifies the intended output as still-image concepts.

Three related names appear in these demonstrations:

• MiniMax H3 Image: the early-access image model behind the reference and repainting tests above.
• MiniMax H3: the open-weight video model, which MiniMax says generates clips up to 15 seconds at 2K with native stereo audio.
• Pruna P-H3-Image: a separate image-generation and editing model based on H3. Its model card lists eight input images, 8-megapixel output and eight sampling steps, with ComfyUI files also available.

A tester whose 269-image review was shared in kaigani's thread reports averaging 25 seconds per native 8-megapixel image on one RTX 4090. Those measurements concern Pruna's model.

GlennHasABeard's 15-second paw clip follows a tightly specified beat: press once into frost, lift, then leave the frame. The camera remains locked in a top-down view.

The creator describes this production chain:

1. Claude and MCP: the creator supplied the desired clip and references, then Claude used MCP to handle the remaining work, according to their workflow reply.
2. ComfyUI: local generation with open-weight MiniMax H3.
3. Topaz: upscaling the rendered video to 4K.

The single-press paw shot

To avoid inventing a cat's coat colour without a reference, the creator limits cats to partial views in these clips.

GlennHasABeard also shared The Ooze from Outer Space Ep.3: The Water Plant, a roughly two-minute episode with an attached comparison pairing plant reference footage with its stylized animated counterpart.

Reference footage alongside the animated water-plant sequence