Physix Frontier · News Briefing Card (TMTPost · Nov 21, 2026)
Three AI image models tested: Doubao and H3 pass continuous editing, Qwen falls behind
KEY FACTS
- On November 9, ByteDance released Seedream 5.5; on November 10, MiniMax released H3-Image and open-sourced its weights.
- The test used the same prompt to generate images in five styles, then made three rounds of continuous edits to the cyberpunk image.
- Both Doubao and H3 completed the three steps of swapping the umbrella, changing the raincoat color, and adding sign text, with no obvious drift in the original image's main subject.
- Qwen lost the original scene in the first round of editing, the second round turned the image into a hallway coat rack, and the test was abandoned in the third round.
- H3 output 4K images throughout, while Doubao and Qwen produced noticeably smaller images.
KEY DATA
4KH3-Image output resolution
6Qwen max images per generation
A dozen-plus secondsDoubao generation time
About one minuteQwen generation time
PHYSIX OBSERVATION
Good-looking generated images are only the entry ticket; continuous local editing is the real workflow necessity. Doubao and H3 can preserve the original image structure, showing that multi-round editing is moving from demo to usable; Qwen has visual highlights in single images, but losing the scene during editing exposes the weakness of conversational editing. When users choose a model, they should score "can it be edited" and "does it look good" on the same table.
Source: TMTPost report
Physix Frontier