AIToday
Image GenerationQiita 機械学習Published: Nov 12, 2026, 13:00 JST

P-H3-Image makes 8-megapixel posters; crowd faces break

Send on LINE
P-H3-Image makes 8-megapixel posters; crowd faces break

3 Key Points

  1. What happened

    A writer used Claude Code with P-H3-Image to generate 8-megapixel posters from text in one pass each; the portrait's text held up, but the Shibuya crowd's central face broke and needed cropping.

  2. Why it matters

    The result suggests single-pass generation can carry a large figure and big simple text, but dense small faces are where it fails, so crowd scenes may need manual repair.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The experiment pairs a single-subject portrait poster with a dense Shibuya scramble crossing to test how P-H3-Image handles scale and detail at roughly 8 megapixels.

The repair process is the most concrete finding. Editing the full 8-megapixel image changed the composition and rounded the output to 3552×2368. A 640-pixel crop around the face barely changed anything. Cropping just the head to a smaller square, upscaling it, and editing it with a portrait-redraw instruction fixed the face, then the writer composited it back with a blurred oval mask. The writer says the reason was not verified but suspects that a larger share of the frame lets the model spend resolution on facial detail.

The write-up is candid about limits: each version used a single seed, few repeat-seed comparisons were run, and the relationship between face size and breakage rate was not counted. Separately, the license notes that the MiniMax H3 Community License excludes the EU, UK, South Korea, and US from usable regions, and that marking generated output as machine-made appears in the prohibited list.

FAQ
What model and tools were used to make the posters?
The writer used P-H3-Image, a still-image model based on MiniMax H3, in its int8_convrot version with a bf16 VAE, run in ComfyUI on one GPU. Claude Code handled execution, logging, and drafting, while the writer directed what to make and selected images.
How long did generation take?
The Shibuya image took 69.7 seconds. The first portrait image took 128.5 seconds, including model loading.
What license applies to the model's weights?
P-H3-Image weights are distributed under the MiniMax H3 Community License. The license excludes the EU, UK, South Korea, and the US from usable regions, and the images were generated in Japan.
Qiita 機械学習Read Original Article
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleApple's Harbor app turns Macs into local AI servers