ModelBrief · Seedream 5.5 · Topic Page
Last updated Nov 20, 2026

OmniArt Hands-On: Edits, Transparency and Tiny Text

Published Nov 15, 2026 · Last updated Nov 15, 2026

When ByteDance shipped Seedream 5.5 on November 9, it made three hard claims: multi-round editing that preserves untouched areas "at the pixel level", native RGBA output with real semi-transparency, and sharply lower small-text corruption. OmniArt put all three to a hands-on test on November 11 — the day before the model went live on its own platform — running the same four production tasks on Seedream 5.5 and Seedream 5.0 Pro with identical prompts, reference images, aspect ratios and 1440p resolution. The report was published November 12. Here is our rewrite of what it found.

10-round editing: all ten edits landed, but the untouched wall darkened ~5%

The first task was a chained edit of a café portrait: starting from the same round-0 image, ten rounds changed exactly one thing each — sweater colour, a ring, the drink, the glasses, a potted fern, a notebook, a necklace, the drink again, the ring's style, then the sweater back. Each round fed the previous output in as the sole reference image. Seedream 5.5 applied all ten changes while the face, pose, table and composition held steady. 5.0 Pro applied every visible change too, but the frame drifted cyan mid-chain and the face and hair changed.

"Pixel-level preservation" is not absolute, though. OmniArt measured a patch of wall nobody touched: its average colour slid from RGB 135/138/114 at round 0 to 129/131/107 at round 10 — about 5% darker overall, with the wall's shadows looking slightly mottled at full size. For comparison, 5.0 Pro's same wall moved from 87/87/74 to 74/93/92 and the window frame from 157/161/154 to 139/173/193, a cyan-blue shift visible at a glance and bad enough to need manual colour correction. 5.5's drift is an order of magnitude smaller, but it is not zero. OmniArt's working rule: save every accepted intermediate round, check untouched areas every few rounds, and when a round goes wrong, branch from the previous round instead of trying to fix it with the next edit.

Transparent PNG: 77.6% fully transparent, but the glass bottle is not see-through

The second task set "background" to transparent for a product cutout of a glass perfume bottle. In the downloaded RGBA PNG, 77.6% of pixels were fully transparent, 22.1% fully opaque, and only 0.3% semi-transparent, all of it along the edges — a very clean cutout. The catch is the bottle itself: the glass is drawn in opaque pixels, so the background does not show through it; on a dark background the bottle stays bright, and the soft grounding shadow the prompt asked for never appeared.

That does not make the semi-transparent channel a decoration. In a separate product cutout test, the shadow under the bottle came out as genuinely semi-transparent pixels. So the finding splits in two: semi-transparency is real for edges and soft shadows, but do not count on it for glass, smoke or liquids — check the alpha channel in the downloaded file before using it, not just the preview.

Dense small text: all four quoted lines right, one paragraph collapsed

The third task was a 3:4 newspaper front page. Seedream 5.5 printed all four quoted lines from the prompt exactly, and even added a small masthead and issue number; 5.0 Pro missed only one line, and only on capitalisation. The real gap was in the body copy: 5.5 set the paragraphs larger and kept almost everything readable English, but left two defects — one four-line paragraph that went blurry and garbled, and one word split across a line break and repeated, "resi- resilient". 5.0 Pro packed in more, smaller type and added a sidebar, but the bottom three columns alone contained at least seven garbled or wrong words, including "rvived", "subsudy" and "harbos".

Two caveats belong with that result: 5.5 set its type larger, so this is not a like-for-like density comparison; and ByteDance's own severe and mild text-corruption rates of 0.4% and 1.7% cannot be verified from one page per model. The direction matches the vendor's claim, but the numbers remain the vendor's own. Proofread every line before publishing a page like this.

Skin and light: a difference of style, not of quality

The fourth task was a backlit portrait of an older man by a window. Neither image was overexposed or shiny, so ByteDance's claim about excessive shine produced no separation here. The difference is in look: 5.5 took "visible skin texture and fine lines, no retouching look" literally — deep wrinkles, visible pores, darker, moodier fill light; 5.0 Pro was the cleaner, evenly lit, posed-smile portrait. Pick 5.5 for documentary realism, 5.0 Pro for a friendly campaign face.

Cost: 4 credits per 1440p image, a third of 5.0 Pro

On OmniArt, Seedream 5.5 costs 2 credits per image at 1080p, 3 at 1.5K and 4 at 1440p; 5.0 Pro costs 6 at 1080p and 12 at 1440p. At the same 1440p ceiling, 5.5 costs exactly a third of its predecessor, with the 1.5K tier added as a middle step for drafts that need cropping headroom. Both models require a Starter plan or above.

Limits and standpoint: single runs, not a benchmark

The boundaries of this test come first. Each model ran each task once; the original states plainly that this is not a benchmark, only a spot check of the three claims. The article mentions no retries or cherry-picking, so the results read as single-run outputs. Four tasks, one image each, is a small sample — every finding applies to this set of prompts and framings and cannot be extrapolated into a verdict on the model as a whole.

The source's standpoint is worth flagging too: OmniArt is not ByteDance. It is a multi-model workspace that bundles image, video, audio and music models, with the model integration running through PixVerse. But it also sells access to those models by the credit, and Seedream 5.5 has been live on its platform since November 12. So this is a third-party hands-on, independent of the vendor — but not a fully neutral academic evaluation, since its business is selling model usage.

Our take: All three claims point the right way, and each needs a discount: the edit chain landed all ten rounds with face and composition intact, but the untouched wall darkened about 5% — "pixel-level preservation" is an approximation, not a guarantee. The transparent background is real and the cutout is clean, but the glass bottle is drawn in opaque pixels, so don't expect genuine see-through glass. The small text nailed all four quoted lines and is far cleaner than the previous generation, yet a long paragraph can still collapse wholesale — proofread every line before publishing. On that scale, the edit chain is safe to use, transparent output fits logos, stickers and UI icons, and dense small type suits short quotes rather than full-page body copy. That matches the direction of our 100-round editing article, which found the official evidence was a single portrait demo and the longest independent test ran 18 rounds — "steadier, not frozen". OmniArt's ten rounds add another data point in the same direction: drift is smaller, but still there. The cost cut to a third of 5.0 Pro at 1440p is the most concrete reason to switch.

Reference source: this article is an original rewrite produced by this site; facts and figures were checked against OmniArt — Seedream 5.5 hands-on test (Nov 12, 2026)

← Back to the Seedream 5.5 topic page