MiniMax-H3-Image Bench

600 prompts in two rounds. Round 1 (sets A-F): 300 prompts reused from a MOSS-Image benchmark, across people, groups, sci-fi, art styles, text and design, and lighting. Round 2 (sets G-L, tabs marked R2): 300 new prompts written for this model in MiniMax's own prompt style, covering commercial, places and nature, non-Latin scripts and long text, information design, dense scenes and extreme aspect ratios, and physics, history and exclusions. Each has a checklist graded pass, partial or fail, rendered at the model's native ~16 MP. Click any image for its grading sheet; click the print to view it at 100%.

minimax_h3_image_pruned_int8_convrotQwen3-VL-32B NVFP4 encoder6 steps · euler / simple cfg 1, no negative~16 MPRTX 5080 16 GBComfyUI v0.45.02026-11-10
How this was made

Hardware

  • One NVIDIA RTX 5080 (16 GB VRAM) in a home workstation with 192 GB of system RAM, running ComfyUI v0.45.0 as a service. ComfyUI was the only process on the GPU when checked during the run.
  • The int8 DiT is about 21 GB and the text encoder about 15 GB, so neither fits in VRAM. ComfyUI's dynamic VRAM loading streams weights from system RAM, and the text encoder is evicted before sampling. No out-of-memory errors at 16 MP.

Workflow

  • The stock ComfyUI text-to-image template for MiniMax-H3-Image (image_minimax_h3_image_t2i), native nodes only, no custom nodes: UNETLoader → MiniMaxH3ImageConditioning (which also makes the empty latent) → KSampler → VAEDecode. The negative is a ConditioningZeroOut of the positive.
  • Sampler fixed by the model's distillation: 6 steps, cfg 1, euler / simple, denoise 1. No negative prompt. Batch 1.
  • The template's optional prompt enhancer (Qwen3.5-9B rewriting the prompt first) was off: every image saw the prompt exactly as written.
  • Seeds are fixed per prompt (CRC32 of its id) and match the earlier MOSS-Image run.

Model files

  • Diffusion model: minimax_h3_image_pruned_int8_convrot (21 GB, int8), the build the template uses.
  • Text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq (Qwen3-VL-32B, NVFP4, 15.7 GB), loaded with CLIPLoader type minimax. This differs from the stock template, which names the int8 build of the same encoder (qwen3vl_32b_minimax_h3_int8_convrot, 27 GB). The NVFP4 build was already on the machine for MiniMax-H3 video. The two were not A/B tested.
  • VAE: minimax_h3_image_vae_fp16.
  • All from Comfy-Org/MiniMax-H3-Image and Comfy-Org/MiniMax-H3.

Resolution

  • About 16 MP per image, the size the model card recommends (working range 4 to 20 MP): 4096×4096 for 1:1, 4736×3552 for 4:3, 5024×3360 for 3:2, 5472×3072 for 16:9, 6272×2688 for 21:9, 4576×3648 for 5:4, and the portrait equivalents. Round 2 adds 5792×2880 for 2:1, 6912×2304 for 3:1 and their portrait equivalents. Every edge is a multiple of 32.
  • The viewer loads a WebP at full resolution (quality 90). The lossless PNG straight out of ComfyUI is linked on each image.

Prompts

  • 300 prompts in six sets of 50, each with a checklist of 4 to 8 visible requirements (1,876 checks in all). They were written for an earlier benchmark of MOSS-Image and reused unchanged here.
  • 282 are prose captions, close to the style MiniMax recommends. 10 are anime tag lists (set D), sent as-is.
  • 8 design prompts in set E were written with a bounding-box layout for MOSS-Image. This model has no layout input, so it received only the caption, which still states every string and its position.
  • Round 2 (sets G-L, 300 prompts) was written from scratch for MiniMax-H3-Image, following MiniMax's published prompt-writing guide: one English description of the finished image, 150 to 300 words, positions anchored to the frame, every visible string in quotes, unscripted text surfaces stated as plain, exclusions written into the prompt. Each prompt carries 5 to 8 checks and tags naming what it probes. Round 2 targets what round 1 never covered: commercial and food photography, architecture and nature, eight non-Latin scripts and paragraphs of 30 to 90 words, charts, maps, UI and logic artifacts (calendars, clocks, chess), scenes with 10 to 20 placed elements, 2:1, 1:2, 3:1 and 1:3 frames, physics, machines, historical periods, expressions and exclusions. Seeds are CRC32 of the id, as in round 1.

Grading

  • Each image was graded by a Claude reviewer, check by check: pass, partial or fail, with a note on what is actually visible. Then a 1 to 5 score against the whole prompt (5: everything met, no visible defects; 4: minor misses, still usable; 3: one key requirement missed; 2: several missed; 1: off-brief). Plus anatomy defects, a transcription of any requested text, and a standout flag for roughly the best 15% of each set.
  • Reviewers looked at a 1280 px copy, and cropped the full-resolution file for small text, finger counts and counting.
  • Each set's findings (failure modes, what correlated with success, prompting advice) are in the panel that opens when you pick a category.

Timing

  • Measured execution time per image is shown in each image's details.
click to zoom 100%