MiniMax H3 Image Evening Report 2026-11-22

This issue's H3 Image section features only 1 new versions: a style LoRA trained on images generated by GPT-image-2. The author labels it as a beta version, trained using only 23 images. The page type is incorrectly marked as a base model. It retains a B rating, and no top recommendation is issued for this issue.

Zhenzhen's AI Radar · MiniMax H3 Image · 2026-11-22 Evening Report

This issue features 1 new versions: Strongly Recommended 0 · Recommended 0 · Worth Watching 1 · Included Only 0

Editor's Note This issue's H3 Image section features only 1 new versions: a style LoRA trained on images generated by GPT-image-2. The author labels it as a beta version, trained using only 23 images. The page type is incorrectly marked as a base model. It retains a B rating, and no top recommendation is issued for this issue.

Worth Watching (1)

[B] Ps GPT-image-2 style for H3 image · v1

Base Model / Checkpoint · New Model · Author Pavers · Published 11-22 19:28 · Downloads 9

A style LoRA that distills a specific artistic style, which the author extracted from GPT-image-2, into H3 Image. All training images were generated by the author using GPT-image-2.

Images

Example 1280x1920

Reasons for Recommendation

  • New model: Serves text-to-image users who want to reproduce a specific GPT-image-2 output style locally on H3 Image; the author provides no textual description of the specific style, so it can only be judged by viewing the example images
  • The author has disclosed the training configuration: H3 Image pruned BF16 base model, qwen3vl_32b text encoder int8 version, 23 images, rank 16 / alpha 16, total 805 steps; recommended output resolution is no lower than 8MP
  • 3 example images are visible (another 3 sensitive-level images are hidden), free download, 9 downloads within 2 hours of publication

How to Use

  • The page type is listed as a base model, but the file is only about 148 MB and was trained with rank 16. It should be loaded as a LoRA stacked on top of the H3 Image base model, not as a replacement for the base model
  • No trigger words; the author did not specify strength, steps, or CFG (805 refers to training steps, not generation steps); the author recommends an output resolution of 8MP or higher

Risks and Limitations

  • The author self-labels this as a beta version, with a training set of only 23 images
  • All training images are generated results from GPT-image-2, which may be subject to that service's terms of use; however, the license permits rental and commercial use of generated images, so readers must judge the compatibility themselves
  • The page type is incorrectly marked as a base model (Checkpoint); the "replace base model loading" usage suggested by automated workflows based on this classification is incorrect

Data: Civitai Public API (Base Model MiniMax H3), inclusion window 11-22 08:31 → 11-22 21:01 (Beijing Time). Recommendation reasons have been reviewed by editors; download counts reflect attention levels only.

Civitai

Explore the latest AI Radar updates