MiniMax H3 character consistency with a character sheet

From turnaround sheet to MiniMax H3 in five steps

1. Get a character sheet: download a pack from the library or a free character. The front, side and back are already separate files.
2. Upload the five images below as reference images, in this order. Five keeps you inside the API's free image count.
3. Define the character once as <Subject 1> and say which picture supplies what: face, profile, back, proportions.
4. Describe the shots under detailed_description, cutting with timestamps such as "At 00:05.000".
5. Generate at 768P. Regenerate only the take you keep to 2K: 768P plus regeneration costs the same per second as generating at 2K directly, so the rejected takes come out cheaper.

Browse character turnaround sheets

1. <Picture 1>02_turnaround_views/view_front.pngFace and front of the outfit
2. <Picture 2>02_turnaround_views/view_side.pngProfile and hair from the side
3. <Picture 3>02_turnaround_views/view_back.pngHair and outfit from behind
4. <Picture 4>01_character_reference/full_body.pngHeight and proportions
5. <Picture 5>05_scenes/scene_workspace.pngThe location

MiniMax's reference guide says one subject can be defined by several images, with each image's contribution stated. That is exactly a turnaround: one <Subject> built from front, side and back.

Prompt template: one character

This is the structured format from MiniMax's reference-mode prompt guide, the same structure its hosted H3-Context-IR step writes. On the API and in the Hailuo app a plain description that names the reference images also works; MiniMax's own API example is one plain sentence. The open weights have no Context-IR, and MiniMax recommends following its prompt guide instead, so this format matters most when you run H3 locally.

Prompt template: two characters

With two characters, front and side for each plus a scene uses five images. Fix who stands where, and keep each character's description identical in every clip.

Character sheet plus a first frame

The API will not combine a first_frame image with reference images. Upload the frame as another reference image instead and declare it in the prompt: "<Picture 6> is the first frame of [Shot 1]". MiniMax's guide uses a standalone <Picture N> for exactly this case.

Running the open weights in ComfyUI

• Reference-to-video needs the ref2va weights, not the fl2va weights used for text and image to video.
• Images become <Picture 1>, <Picture 2>… in node order (ref_image_0 is <Picture 1>), so wire the pack files in the order above.
• The 4-step Turbo LoRA is faster but holds references more weakly. When the face must hold, use the base schedule at 20–25 steps.
• Style LoRAs compete with the references; ComfyUI suggests a low strength (about 0.5) when adherence matters.

Community workflows that turn 1–9 images into one H3 character sheet are third-party, not part of MiniMax's release.

What one attempt costs on the API

From MiniMax's official price list. Five reference images are free; video inputs are billed by their length.

• 10 seconds at 768P: $0.80
• 15 seconds at 768P: $1.20
• 15 seconds at 2K: $1.95
• Regenerating a kept 15-second take to 2K: $0.75, plus any paid input images again

The Hailuo app has its own plans and credits. MiniMax's video packages do not cover H3 yet.

H3, Highspeed, Lite or Max: which version to use

MiniMax's API now offers four H3 versions. All of them accept reference images and use the same request format, so the same pack files, picture order and prompt work if you only change the model name.

• MiniMax-H3: the quality version. 768P or 2K, 4–15 seconds, $0.08 per second at 768P and $0.13 at 2K. The first 5 images are free, then $0.04 each. It is the only version that supports the H3-Context-IR prompt step and 2K regeneration.
• MiniMax-H3-Highspeed: balances quality and speed. 768P only (no 2K), 4–15 seconds, $0.06 per second. The first 5 images are free, then $0.03 each.
• MiniMax-H3-Lite: the low-cost version for exploration and batches. 480P or 768P, 4–15 seconds, $0.025 per second at 480P and $0.04 at 768P. The first 5 images are free, then $0.02 each. It does not accept the adaptive aspect ratio yet, so set one such as 16:9 or 9:16.
• MiniMax-H3-Max: a fast version post-trained by fal.ai on H3. 480P or 768P, 5–15 seconds (no 4-second clips), $0.05 per second at 480P and $0.08 at 768P. Only the first 2 images are free, then $0.074 each, so a five-image pack adds about $0.22 per clip.

A cheap way to test a character: draft the shots on Lite at 480P with the five pack images (10 seconds costs $0.25), then render the take you keep on MiniMax-H3 at 768P with the same pictures and the same subject_definitions. We have not compared how well each version holds a face, so check the final take on H3 before you build a longer sequence on a cheaper version.

Some third-party listings describe Highspeed as going up to 2K. MiniMax's API reference lists 768P only for Highspeed, and regeneration to 2K only takes 768P videos made by MiniMax-H3.

Face drift and other problems

The face changes from shot to shot.

Use the same pictures in the same order and paste the same subject_definitions line into every clip. Change only the shots, never the identity text.

Faces got softer after switching on Turbo.

The 4-step Turbo LoRA trades reference strength for speed. Go back to 20–25 steps for the final render.

The request fails when I add a first frame.

First/last-frame and reference images are mutually exclusive. Send the frame as a reference image and name it as the first frame in the prompt.

Two characters swap features.

Give each character its own <Subject> with its own pictures, fix their positions, and list both as fully_preserved in retention_analysis.

The side profile is invented.

Make sure the side view is uploaded and named as the profile in the subject definition. A front view alone shows the model nothing to copy.

My draft looked fine on Lite, but the face changed on H3.

The versions are different models, so treat the Lite draft as a test of the shots, not of the final face. Keep the same pictures in the same order and the same subject_definitions, and judge identity only on the version you will publish.