AI image models, episode 6. Previously: FLUX, Nano Banana vs GPT Image, xAI Grok Imagine, Midjourney vs Ideogram and Recraft vs Krea. Next: Microsoft MAI-Image.
ByteDance and Alibaba both sell image models through their clouds, and both ship fast. ByteDance’s Seed team released Seedream 5.5 today, 9 November 2026, a day after we compared its video sibling Seedance 2.5 with Wan 3.0. Its developer API documentation still tops out at Seedream 5.0 Pro, so we cover 5.0 Pro as the documented API model and note what ByteDance says 5.5 adds. On the Alibaba side we compare Qwen-Image 3.0 / 3.0 Pro (Model Studio API) with Qwen-Image-2.1, the latest open-weight release.
1. Max resolution
- Seedream 5.0 Pro / Flash: resolution levels 1K, 1.5K and 2K (default 2K), or explicit width by height with total pixels from 1280 by 720 up to about 4.6 megapixels. At 2K, 1:1 is 2048 by 2048 and 16:9 is 2816 by 1584. The older 5.0 Lite and 4.5 offer 2K, 3K and 4K levels, up to 4096 by 4096.
- Qwen-Image 3.0 / 3.0 Pro: total pixels from 512 by 512 to 2048 by 2048, PNG output. If you omit size, the model picks one from the prompt.
- Qwen-Image-2.1 (open): the model card’s presets run up to 2048 by 2048 and 2752 by 1536 (see our Turbo weights post).
2. Aspect ratios
- Seedream: 1:16 to 16:1 when you set pixels; at resolution levels you describe the shape in the prompt. Presets include 1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3 and 21:9.
- Qwen-Image 3.0: 1:8 to 8:1 for both generation and editing.
3. Reference types
- Seedream 5.0 Pro / Flash: text to image, single-image and multi-image editing, layer decomposition (one base image plus up to 16 output layers, with order, bounding box, name and description), and interactive editing where you mark the edit with coordinates, a selection box or an arrow. Sequential batch generation is not supported on Pro and Flash; it is on 5.0 Lite and 4.5.
- Qwen-Image 3.0: text to image and image editing in one model, plus negative prompts (3.0 series only). Qwen-Image-2.1 Pro and Turbo on the API can decide to output an image with a transparent alpha channel from the prompt.
4. References at once
- Seedream: up to 10 reference images on 5.0 Pro and Flash; up to 14 on 5.0 Lite, 4.5 and 4.0.
- Qwen-Image: 1 to 3 input images on the 3.0 series; 1 to 10 on the 2.1 series.
5. Price (read 9 November)
- Qwen-Image 3.0 Pro (Model Studio, Singapore, international): $0.04 per 1K output image, $0.075 per 2K, plus $0.003 per input image. Qwen-Image 3.0: $0.03 at 1K or 2K. Beijing and global regions are slightly lower ($0.0344 and $0.0688 for Pro).
- Seedream 5.0 Pro: BytePlus’s pricing page did not render for us. Artificial Analysis lists it at $90 per 1,000 images. The BytePlus Seedream product page advertises new-user offers “starting at $0.02/image”, which is a promotion, not the list price.
- Seedream 5.5: IT Home reports a 2K square edit costs about 0.28 yuan, under half of 5.0 Pro, with the same input-image plus output-image billing. We have not read ByteDance’s price sheet for 5.5.
6. Speed and leaderboard rank (Artificial Analysis, read 9 November)
| Model | Text to image | Editing |
|---|---|---|
| Seedream 5.0 Pro | #18, Elo 1101 | #11, Elo 1129 |
| Qwen-Image-3.0-Pro | #19, Elo 1095 | #23, Elo 1091 |
| Qwen-Image-3.0 | #20, Elo 1073 | #29, Elo 1069 |
| Qwen-Image-2.1 (open) | #23, Elo 1045 | #26, Elo 1080 |
| Seedream 5.0 Lite | #39, Elo 1019 | #35, Elo 1061 |
Both boards are led by OpenAI’s GPT Image 2.6 models. For reference, xAI’s grok-imagine-image-quality sits at #22 (Elo 1049) on text to image. Seedream 5.5 is not on either board yet. On speed, BytePlus offers a fast mode on 5.0 Pro and recommends Flash for latency-sensitive apps; neither vendor gives a seconds figure in the pages we read.
7. Character consistency
- Seedream: ByteDance says 5.5 keeps unedited regions intact “at the pixel level” across multi-turn edits, and IT Home reports a ByteDance test of more than 100 rounds. That is the vendor’s test.
- Qwen-Image: multi-reference editing (up to 10 images on 2.1) is the consistency tool; we found no specific identity-lock claim for 3.0.
8. Style and scene consistency
- Seedream: layers and region edits let you change one element without regenerating the scene, which is the practical form of scene consistency.
- Qwen-Image 3.0 Pro: Alibaba pitches dense layouts (newspapers, storyboards, menus, exam papers) in a single pass with up to 4.5k tokens of input.
9. Against the previous version
- Seedream: 5.0 Pro (31 July) added interactive editing and better text and translation in more than 10 languages. ByteDance says 5.5 adds stable multi-turn editing, native transparent and semi-transparent RGBA output, and more natural lighting and skin, and that a 5.5 Pro will follow.
- Qwen-Image: 2.1 (20 September) is the open release; 2.1 Turbo (9 October) cut it to eight steps. 3.0 and 3.0 Pro are API models; we found no 3.0 weights on Hugging Face or in the Qwen-Image-2.1 news list.
10. Text rendering
- Seedream: ByteDance’s own eval, as IT Home reports it, puts 5.5’s severe text-error rate at 0.4% and minor at 1.7% across Chinese, English and dense small text. Vendor numbers.
- Qwen-Image 3.0 Pro: Alibaba claims text as small as 10 px and native rendering in 12 languages. Vendor claim.
11. What users think (labelled)
We did not pull a fresh user sample for this episode. On Artificial Analysis, which is built from blind user votes, Seedream 5.0 Pro’s lead is much larger on editing than on text to image.
12. General quality
On the AA boards, Seedream 5.0 Pro and Qwen-Image-3.0-Pro are within a few Elo points on text to image, well behind the leaders. On editing, Seedream 5.0 Pro is 38 points ahead. We ran no samples.
13. Licence and commercial use
- Seedream: closed model, available through BytePlus ModelArk and ByteDance apps (Dreamina; for 5.5, Jimeng, Doubao and Volcano Ark, per IT Home). ByteDance says outputs carry invisible watermarks, C2PA Content Credentials and visible AI labels in Dreamina. We did not read BytePlus output terms for this draft; check them before commercial use.
- Qwen-Image 3.0: API only. The
watermarkparameter defaults to off. - Qwen-Image-2.1: open weights on Hugging Face and ModelScope under the Qwen Research License, which bars commercial use without a separate licence.
14. Verdict (opinion)
Pick Seedream if your work is editing: layers, region edits, 10 references and the higher AA editing rank. Pick Qwen-Image 3.0 if you want a cheaper, clearly priced API for dense text layouts at 1K or 2K. Pick Qwen-Image-2.1 if you need weights on your own GPU for research, and read the licence first. Revisit Seedream once 5.5 is in the API docs and on the boards.
Summary
| # | Checklist | Seedream 5.0 Pro (5.5 notes) | Qwen-Image 3.0 / 2.1 open |
|---|---|---|---|
| 1 | Max resolution | 2K (about 4.6 MP); 4K on 5.0 Lite / 4.5 | 2048 by 2048 |
| 2 | Aspect ratios | 1:16 to 16:1 | 1:8 to 8:1 |
| 3 | Reference types | edit, multi-ref, 16 layers, arrow/box edits | edit, multi-ref, negative prompt; alpha on 2.1 API |
| 4 | References at once | up to 10 (14 on Lite / 4.5) | 1 to 3 (3.0); 1 to 10 (2.1) |
| 5 | Price (9 Nov) | AA lists $90 / 1k; 5.5 reported under half | $0.04 1K / $0.075 2K (Pro); $0.03 (3.0) |
| 6 | AA rank T2I / edit | #18 / #11 | 3.0 Pro #19 / #23; 2.1 #23 / #26 |
| 7 | Character consistency | 5.5 multi-turn claim (vendor) | multi-ref |
| 8 | Style/scene | layers and region edits | dense layouts (vendor) |
| 9 | vs previous | 5.5: multi-turn, RGBA; Pro to follow | 3.0 API; 2.1 open, Turbo 8 steps |
| 10 | Text rendering | 5.5 error rates (vendor) | 10 px, 12 languages (vendor) |
| 11 | Users | AA edit lead | AA mid-pack |
| 12 | Quality | close on T2I, ahead on edit | close on T2I |
| 13 | Licence | closed; C2PA in Dreamina; terms unread | 3.0 API; 2.1 research licence, no commercial use |
| 14 | Verdict (opinion) | editing pick | cheaper API / run-it-yourself research pick |
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment