Quantslant Tests 4 Image Models: H3 Cheapest, Weak on Text

Seedream 5.5 arrived November 9, H3-Image on November 10 — two Chinese image models landing a day apart. Quantslant's November 12 article does something rare: it runs the same prompts with the same counting method through both newcomers and the two models it had already tested, Nano Banana 2.1 and GPT-Image-2.6 — six categories, every one backed by numbers. The tests were completed November 10 and the article published at 18:30 Beijing time on November 12. The conclusion is clean: H3-Image is the cheapest of the four, and the weakest on text, faces and transparency.

Six categories: the full table

Prices are what the author read from the cost fields in his November 10 tests (Nano Banana 2.1 at 2K via OpenRouter, GPT-Image-2.6 Sunburst medium at 1024×1024, Seedream 5.5 at WaveSpeedAI's published price); resolutions and reference counts come from provider catalogues as of November 10. A hyphen means he did not check.

Polish text: Nano Banana wins outright, H3 trails

Five prompts — a bakery sign, a poster, a book cover, a menu and a pangram, 389 characters in total. Nano Banana 2.1 and GPT-Image-2.6 were nearly flawless, Seedream 5.5 averaged 3 errors, H3-Image 13. H3's most common mistake was writing "zl" instead of "zł" (the Polish currency sign), in nearly every price. Short slogans tell the same story: H3 got 17 of 30 images error-free, Seedream 27, Nano Banana all 30. Conclusion: for non-English text in the image, don't pick H3.

Faces: GPT-Image-2.6 leads, and H3 has an EXIF trap

Same setup for every model — 10 prompts, the same 3 photos of Waco (who agreed to the use of his likeness in the tutorials), ArcFace scored identically: GPT-Image-2.6 0.729, Seedream 5.5 0.632, H3-Image 0.573, Nano Banana 2.1 0.508. H3 keeps the character and clothing consistent, but the face is "similar", not "the same". One pitfall deserves its own note: H3's first run scored 0.273 — a stranger — because the author sent phone photos without applying EXIF rotation, so the model received faces lying on their side; the other APIs rotate photos on their own, this one does not. After rotating, the score was 0.573. Straighten your photos before using H3 for face references.

Transparency: Seedream 10/10, H3 0/5

H3-Image has no transparent background: the documentation lists it as unsupported, and in the test it managed 0 out of 5; Seedream 5.5 returned 10 files with a real alpha channel (glass excepted). For product cutouts, H3 needs a separate background-removal tool. The author did not test transparency on Nano Banana or GPT-Image-2.6.

Editing: single edits excellent, chained edits degrade

A single edit from the original is very good on H3: one serum bottle moved into six scenes, the label identical all six times. Chained editing — editing the result of a previous edit — is worse: after 2–3 rounds a texture appears on the label paper, and after 5 rounds the image looks oversharpened; in the same test Seedream 5.5 got through 8 rounds cleanly. One-shot edits are fine on H3; for repeated edits of the same image, Seedream holds up better.

Resolution and reference images: H3's home ground

Here H3 has no competition: 16MP straight away with no upscaler, in 22–24 seconds for $0.034; the 16-reference test (an empty room plus 15 objects) produced a living room with all 15 objects in both attempts. At 2K it costs $0.023 and takes 13–17 seconds — the cheapest of the four, and as fast as Nano Banana 2.1. GPT-Image-2.6 at high quality was the slowest, up to 43 seconds per image.

Which one to pick

The author's verdict splits cleanly by use case: graphics with Polish diacritics — Nano Banana 2.1 (GPT-Image-2.6 as an alternative); a specific face from a few photos — GPT-Image-2.6; a product on a transparent background — Seedream 5.5; large images, print, scenes with many objects, one product in many settings, on a budget — H3-Image.