MiniMax-H3-Image · 16MP
Official 33.1B DiT architecture for 6-step native 4096×4096 (16MP) synthesis and multi-reference editing.
Inputs and settings
- Aspect ratio
- 1:116:99:164:33:43:22:321:9
- Output size
- 2K / 4K
- Reference
- Up to 4 images
Examples
Character Reference Sheet (3 Views + 6 Expressions)
Character reference sheet of one young woman with short silver hair, an olive green bomber jacket, black cargo pants and white sneakers. Top row: full-body front view, side view and back view in the same neutral pose, evenly spaced on a plain light gray background. Bottom row: six head-and-shoulder close-ups with different expressions labeled "HAPPY", "ANGRY", "CALM", "SURPRISED", "THINKING", "LAUGHING". Clean digital concept illustration, flat even studio lighting, identical proportions in every view, no other text.
Source · MiniMax H3 Community License · 2026-11-10
Multi-Reference Living Room Composition
Edit <Picture 1>. Furnish the sunlit mid-century living room with the cognac leather sofa from <Picture 2>, the sculpted walnut coffee table from <Picture 3>, and the arched brushed-brass floor lamp from <Picture 4>. Keep the tall factory windows, herringbone oak floor and morning sunlight direction exactly as they are.
Source · MiniMax H3 Community License · 2026-11-10Architecture, Native Resolution Table & Hardware Comparison
33.1B Single-Stream Transformer
Built on the multimodal H3 backbone with 50 layers, 56 attention heads and 3D multimodal RoPE. Text prompts are encoded through a 50-layer Qwen3-VL encoder while uploaded reference frames pass through a 24-channel 2D ViT spatial VAE into one unified sequence.
6-Step & 8-Step CFG-Free Sampling
Distilled with per-step PDD output heads at fixed guidance 1.0. Because classifier-free guidance and negative prompts are eliminated, every request completes in 6 to 8 deterministic steps with natural tonal roll-off, authentic skin pores and crisp typography.
Cloud Browser vs Local VRAM
Self-hosting the full BF16 checkpoint requires 66.28 GB for the diffusion transformer, 66.7 GB for the text encoder, and 2.34 GB for the VAE (~135 GB total on H200/B200 accelerators). Even 4-bit ComfyUI quants require 21 GB of VRAM. Here you run both checkpoints instantly in your browser.