ByteDance released Seedream 5.5 on November 9, 2026, one day before MiniMax-H3-Image. Both make and edit images. This page lists only what each company has published; we have not run our own tests.
At a glance
| MiniMax-H3-Image | Seedream 5.5 | |
|---|---|---|
| Released | Nov 10, 2026 | Nov 9, 2026 |
| Weights | Open weights on Hugging Face (community license, territory-restricted) | ByteDance's announcement does not mention open weights |
| Where to use | MiniMax API (global and China), Hailuo AI, MiniMax Code, self-hosted | Doubao, Doubao Work, Jimeng; API on Volcano Engine Ark announced as coming |
| Output size | ~16 MP native (4096×4096 to 6272×2688) | Not stated in the announcement |
| Transparent background | Not supported (per MiniMax docs) | Native RGBA with full and partial transparency |
| Reference images | Up to 16 | Not stated in the announcement |
| Editing headline | Multi-reference editing in one checkpoint | Stable multi-turn editing, up to 100 rounds, unedited regions kept pixel-level |
| Vendor's quality claim | Strongest instruction following among open-weights models | Leads GPT Image 2.5 Sunburst and Nano Banana 2.1 on text structure and text-to-image realism in ByteDance's tests |
| Example price | 2K edit, 1 ref: ¥0.1747 (Standard, China) | 2K square edit: about ¥0.28 per image (reported); text tokens free |
Both quality claims are each company's own evaluations, not independent benchmarks. The Seedream 5.5 price is from Science and Technology Daily's report, not a price list we could open.
Speed
MiniMax's launch chart puts H3-Image at 10.6 s (Standard) for 1K text-to-image and lists Seedream 5 Pro at 34.4 s. That comparison is to the previous Seedream generation, not 5.5. Science and Technology Daily reports that 5.5 is much faster than 5.0 Pro, without a figure. A hands-on by ZhiDongXi reported about 5–10 seconds per image on Doubao Work.
Which to use when
- You need to self-host, fine-tune or run offline: H3-Image, because its weights are public. Check the license territory and hosted-service rules first.
- You need very large outputs: H3-Image documents ~16 MP native output. Seedream 5.5's maximum resolution is not stated.
- You need transparent PNG assets: Seedream 5.5 claims native RGBA. H3-Image's API does not support transparent backgrounds.
- You edit one image over many rounds: that is Seedream 5.5's headline claim (100 rounds).
- You combine many references in one shot: H3-Image documents up to 16.
Launch-day side-by-side
An X user posted a three-stage edit test (Halloween clown → makeup removed → modern look) comparing H3 Image, Seedream and GPT Image. It is a single anecdote, but a real one: post by @One2xDatalab.