Seedream 5.5's 100-Round Editing: Claim vs Evidence

Seedream 5.5's headline feature is a number: 100. One hundred successive edits on the same image, with untouched areas staying pixel-identical and no color drift. It's the first time an image-editing model has put a round count in its headline. But the evidence chain for that number is short: the official demo covers one portrait, and the longest independent test ran 18 rounds. This article lays out all the evidence there is.

The official evidence: one portrait

The official blog's 100-round demo is a single portrait edited in sequence. The blog has two more multi-turn demo sequences — four floral designs with multi-object, multi-material continuous editing, and an old study gradually becoming a colorful fantasy space (perspective, occlusion and shadows staying coherent as furniture moves) — but neither carries a round count. So the number 100 attaches to the portrait demo only. The blog also publishes no quality/fidelity curve over rounds: how much the image degrades after 100 rounds has no official number.

XenoSpectrum flagged this on launch night, and added context: multi-round editing isn't unique to 5.5 — Black Forest Labs' FLUX 3 Image, released in October, also claims multi-step preservation of non-target areas, just without a round count. What's new about 5.5 is quantifying the rounds as 100, not the multi-round capability itself.

The longest independent test: 18 rounds

ifanr/APPSO, in its Nov 9 evening test, took a five-person poster through 18 rounds of editing from modern styling to a period mystery look, with faces and composition holding up well. That's the longest independent multi-round test we found. The same test covered Mona Lisa outfit and profession changes, and a bare room gradually furnished (the view and sunlight angle outside the window unchanged). 18 rounds does not verify 100 rounds — but it does show that at the dozen-round scale, 5.5 really does hold faces and composition.

Quantslant's observations are finer-grained: across 7 rounds of portrait editing, the face was unchanged in rounds 1–4 and 6, with local edits changing 1.2–3.8% of pixels; across 8 rounds of product editing the bottle didn't move at all, with 0–0.2% pixel change when lighting and product were untouched. But there's a counterexample: a "smile" request was ignored — the model doesn't obey every round.

Six-round drift data: steadier, not frozen

Seedance.io ran the most informative measurement on Nov 12: same prompt, three models, 6 rounds of editing, drift measured round by round (2K output). Wall-color drift: 3.5 for 5.5 vs 17.3 for 5.0 Pro (0–441 scale); label-paper color drift 9.0 vs 16.2; untouched bottle-cap area changed 0.6–1.7 vs 3.6–3.7 per pixel per round (0–255 scale). The conclusion is clear: 5.5 is much steadier than 5.0 Pro, but not frozen — the label paper still darkened slightly after 6 rounds.

That "steadier, not frozen" conclusion matters. It suggests the real meaning of 100 rounds may be "the image is still usable after 100 rounds," not "untouched areas are pixel-identical after 100 rounds." The official "pixel-level preservation" needs to be read against this measured background.

Why 100 rounds is hard to verify

There are two structural reasons. First, standard automatic metrics may miss multi-round degradation: the Banana100 study cited by XenoSpectrum (28,000 images generated with Nano Banana Pro) found that repeated editing accumulates artifacts, and none of 21 image-quality metrics reliably identifies degraded images. Second, the AA editing board doesn't test repeated multi-round editing — OrcaRouter notes the 100-round claim and the board scores are orthogonal questions: a high board score doesn't prove 100 rounds, and the 100-round demo can't be derived from board scores.

There's also a practical reason: each round is an independent generation, billed separately. 5.5 has no session memory — multi-round editing means sending the previous result back as a reference each round. 100 rounds means 100 paid generations — the reproduction cost is itself a barrier.

Conclusion

The current evidence supports a tempered conclusion: 5.5's multi-round editing is genuinely much steadier than 5.0 Pro (6-round drift data), with independent support at the dozen-round scale (18 rounds); but "100 rounds" itself remains a vendor claim — one portrait demo, no quantified curve, no independent reproduction. When someone actually runs the full 100 rounds and publishes round-by-round data, this article will be updated.