ByteDance's Seedream 5.5 Holds Up Through 100 Edit Rounds in QbitAI Test; 2K Edit Costs About 0.28 Yuan

ByteDance's Seedream 5.5 kept an image more faithful than GPT-Image 2.6 across 100 consecutive editing rounds in a test run by Chinese outlet QbitAI (量子位), though QbitAI notes some distortion still appeared in later rounds. According to official information cited by QbitAI, an API edit of a 2K square image costs about 0.28 yuan, less than half the cost of Seedream 5.0 Pro. The report is in Chinese.

The 100-round test

QbitAI fed both models the same watch photo and asked each to add one second to the time shown on the watch per round, using each round's output as the next round's input. According to QbitAI, GPT-Image 2.6 (referred to in the article as Image 2.6) began distorting the original image around round 10, with distortion worsening in later rounds. Seedream 5.5, run through Doubao (豆包), still showed some distortion in later rounds but was noticeably more faithful overall, per QbitAI.

A second 10-round test on a realistic living room photo had the same structure: each round's input was only the previous output, so the model never saw the original and errors could accumulate. QbitAI presents this as a test of the multi-round consistency and precise editing that Seedream 5.5 emphasizes.

Listed features

QbitAI lists these improvements:

• Realism: people, materials and lighting look more like real photography, and composites of several images are less likely to show seams.
• Small text and multilingual rendering: text in posters, labels and infographics is rarely wrong, across multiple languages.
• Transparent images: direct output of transparent-background PNGs, plus decomposition of a finished image into editable layers, so copy can be changed or localized without regenerating the whole image.
• Speed: generation is faster than earlier versions, according to QbitAI. No figures are given.

Other tests reported

• Multi-image ad composite: three inputs (a model photo, a handbag, a background) shot under different lighting produced a 16:9 fashion ad in about one minute. The suit fabric, the bag's slanted flap and metal ring handle were preserved.
• Translucent PNGs: five transparent PNG assets were combined into a 3840×2160, 16:9 jelly-style scene. QbitAI says colour showed through the translucent elements and shadows carried each asset's tint.
• Cinematic realism: five 2K images were generated in a single request in roughly one minute.
• Dense-text poster: a poster on "world models" with four to five font-size tiers, Chinese, English and Japanese text, and formulas with subscripts and superscripts. QbitAI reports near-zero garbled text.
• Pose transfer: a pose from one image was applied to a person from another, output at 16:9. QbitAI also reports a nine-person input in which each person was generated in the specified pose.
• Layer decomposition and localization: a perfume ad was split into layers, then re-issued as an English-market version at 3840×2160 with the background, bottle, packaging and lighting unchanged and only the copy changed.
• GIF output: QbitAI had the model produce a 16:9 GIF under 10 MB from the pose-transfer result.

Pricing and availability

According to official information, the Seedream 5.5 API keeps the two-part billing structure of input image plus output image, with no charge for text tokens. Per-token pricing and context window are not given. Seedream 5.5 is available in Jimeng (即梦), Doubao and Doubao Work (豆包工作).

What this means

The 100-round result addresses a specific workflow problem: repeated edits to one image, where each pass can degrade the base. QbitAI, which ran the tests itself and compared only against GPT-Image 2.6, argues this matters most for e-commerce, where product images are revised often, and says one master image can yield a full set of SKU variants. These are one outlet's tests on a limited set of prompts, not a standardized benchmark.

Layer decomposition changes the cost of localization: if only the text layer changes, the rest of the image stays fixed, as in the perfume ad example. Combined with the lower per-edit API price ByteDance cites, QbitAI concludes the economics favor small merchants and content teams that generate many images daily. It also states that Chinese models have moved ahead of leading international models in multi-round editing stability, a conclusion based on its own comparison with one rival model.