改到第一百轮,这张图还认得出自己吗?Seedream 5.5 读后笔记

2026年11月10日 · AOAI AI

Will the Picture Still Know Itself at Round 100? Notes on Seedream 5.5

你一定遇到过这种事:让模型把背景换成米白,背景确实漂亮了,可模特少了一只耳环,衬衫的褶皱也换了走向。第三轮出这种事还能忍,到第八轮,你已经分不清哪一版才是对的。字节跳动 Seed 团队新发布的 Seedream 5.5,瞄准的就是这种”不敢再改”的心理负担。下面是我读完官方介绍和媒体报道后的整理:先看三项升级,再看演示和评测各自能证明什么,最后算一笔官方没有算的账。

You have surely been here: you ask a model to swap the background for off-white, and the background does look lovely, but the model has lost an earring and the shirt’s folds now run the other way. You can live with that at round three; by round eight you can no longer tell which version was the right one. Seedream 5.5, newly released by ByteDance’s Seed team, aims at exactly this reluctance to edit again. Below are my notes after reading the official announcement and press coverage: first the three upgrades, then what the demos and benchmarks actually prove, and finally a cost calculation the vendor did not make.


官方把升级归为三条。第一是多轮编辑稳定性:未编辑区域做到像素级保持,消除色彩偏移,减少连续修改中的细节丢失与画质衰减,上限是多达 100 轮。第二是原生 RGBA 透明素材:直接输出带 Alpha 通道的透明或半透明图像,处理好轮廓边缘,省掉后期抠图修边。第三是质感与文字:改善光照、材质和皮肤纹理,减少过度油亮与过曝,同时提高复杂字形和密集小字的还原准确度。我的看法是,前两条改变的是工作流程,第三条只是画面变好看。这个区别后面会反复出现。

The official summary lists three upgrades. First, multi-round editing stability: untouched regions are preserved at pixel level, color shifts are eliminated, and detail loss and quality decay across successive edits are reduced, up to 100 rounds. Second, native RGBA output: images come out with an alpha channel, with edges and transparent regions handled, so no cut-out cleanup afterwards. Third, texture and text: better lighting, materials and skin texture, less over-shine and overexposure, plus more accurate rendering of complex glyphs and dense small print. My reading is that the first two change the workflow, while the third merely looks nicer. That distinction will keep coming back.


多轮编辑的证据是演示视频,每一轮都拿上一轮结果当输入。第一段是同一人物的 100 轮连续编辑,时长 39.5 秒。据原文作者抽帧观察,某一帧左边是当前结果,一个粉色折纸字母 R,褶皱、反光和毛边都在,右边层叠着历史帧,最上面压着一个大大的 095。戒指、镜框这类小物件被改动时,与皮肤的接触、遮挡和光影依然对得上。另外两组也带轮次计数:室内改造跑到 45 轮,镜头位置没有漂移;花卉组把同一束花分别做成国画、静物、写实和手账拼贴四种材质。计数器的价值在于说明演示真的跑到了那一轮,而不是改几次就剪出来。

The evidence for multi-round editing is demo video, each round fed with the previous result. The first clip follows one character through 100 consecutive edits, 39.5 seconds long. According to the original author’s frame-by-frame look, one frame shows the current result on the left, a pink origami letter R with creases, reflections and rough paper edges intact, and layered history frames on the right, topped by a huge 095. When small items like rings and glasses frames are modified, their contact, occlusion and lighting with the skin still line up. Two more clips carry round counters: an interior makeover reaching 45 with no camera drift, and a flower set rendering one bouquet as ink painting, still life, photorealism and scrapbook collage. The counter matters because it shows the demo really ran that many rounds rather than being cut together from a few.


透明底这一条看似低调,实际改变了产出物的性质。以前模型给你的是”一张图”,现在给的是”一个素材”,能直接放进排版、合成和动画流程。官方给了三段演示:一只展翅的雪鸮,翅尖羽毛、羽毛之间的缝隙和爪部轮廓被单独处理,放在白底上看不出毛边;一排材质各异的字母,有毛绒、气球、巧克力和金属丝,背景透明;还有同一组蓝色毛绒角色,先贴在黄色列车车身上,再立在红色广告牌上,特征和质感延续。第三段最像真实业务,因为接单时最常见的要求就是同一个角色要在不同场景里保持同一个。

The transparent background sounds modest but changes the nature of the output. Before, the model handed you “a picture”; now it hands you “an asset” that drops straight into layout, compositing and animation pipelines. Three demos were shown: a spread-winged snowy owl whose wingtip feathers, gaps between feathers and talon outlines are handled individually, with no fringe against white; a row of letters in different materials, such as plush, balloon, chocolate and wire, on a transparent background; and one set of blue plush characters used first on the body of a yellow train, then on a red billboard, keeping features and texture. The third looks most like real client work, because the commonest demand is that the same character stays the same across scenes.


真实感和小字方面,官方各给了几张样张。人像有南欧街头回眸的女孩,头巾花纹、耳饰、毛孔和贴在肩上的湿发都清楚;还有父亲抱着熟睡婴儿,窗边自然光落在脸和手臂上,暗部保留了皮肤与织物的细节。场景样张有紫亚麻布垂在柠檬黄墙面上的静物,以及雨中的 F1 赛车,路面反光分两层,水雾遮住后车。文字渲染则是两张信息密度极高的图:一张蓝闪蝶图解,八组微距特写各配英文术语和中文说明,”翅鳞””翅脉”这类笔画多的字在小字号下仍可辨认;另一张是莎士比亚主题报纸版面,七篇英文短文挤在旧纸上。这两张好在能被肉眼直接检验。

On realism and small text, the vendor offered several samples. The portraits include a girl glancing back on a southern European street, with headscarf pattern, earrings, pores and wet hair clinging to her shoulder all clear, and a father holding a sleeping baby, window light falling on faces and arms with detail kept in the shadows. The scene samples include a still life of deep purple linen draped on a lemon-yellow wall, and an F1 car in the rain with two layers of road reflection and spray half-hiding the car behind. Text rendering is shown in two very dense images: a blue morpho butterfly diagram with eight macro close-ups, each carrying an English term and a Chinese note, where complex characters stay legible at tiny sizes, and a Shakespeare-themed newspaper layout with seven English articles crammed onto aged paper. These two are good because the eye can check them directly.


评测由官方组织的图像生成专家在 MagicArena 平台上完成,分三块。第一块是电商、营销、影视、游戏、办公、服装设计六类场景的可用率和主观评分,图上只有两条雷达线,没有逐点数值,所以”影视与办公可用率达上一版两倍以上”只能引用官方文字。第二块是 GSB 偏好对比,正数偏向 5.5。文字结构方面,对 GPT Image 2.5 Sunburst 图生图 +20.72%、文生图 +16.01%,对 Nano Banana 2.1 为 +3.82% 和 +5.23%,对自家 5.0 Pro 为 +18.54% 和 +5.00%。真实感方面,对 GPT Image 2.5 Sunburst 为 +36.05% 和 +37.29%,对 Nano Banana 2.1 为 −6.18% 和 +21.79%,对 5.0 Pro 为 +55.62% 和 +39.44%。第三块是文字崩坏率,越低越好:严重崩坏 5.5 为 0.4%,5.0 Pro 为 2.8%,Nano Banana 2.1 为 1.7%,ChatGPT Image 2.5 Sunburst 为 0.7%;轻微崩坏依次为 1.7%、2.6%、3.0%、7.4%。

The evaluation was run by image-generation experts organized by the vendor on the MagicArena platform, in three parts. The first covers usability rate and subjective score across six scenarios: e-commerce, marketing, film, games, office work and fashion design. The charts show only two radar lines with no per-point values, so the claim that film and office usability reached more than twice the previous version can only be quoted from the vendor’s text. The second part is GSB preference, positive favoring 5.5. On text structure: +20.72% image-to-image and +16.01% text-to-image against GPT Image 2.5 Sunburst, +3.82% and +5.23% against Nano Banana 2.1, and +18.54% and +5.00% against its own 5.0 Pro. On realism: +36.05% and +37.29% against GPT Image 2.5 Sunburst, −6.18% and +21.79% against Nano Banana 2.1, and +55.62% and +39.44% against 5.0 Pro. The third is text collapse rate, lower being better: severe collapse is 0.4% for 5.5, 2.8% for 5.0 Pro, 1.7% for Nano Banana 2.1 and 0.7% for ChatGPT Image 2.5 Sunburst; minor collapse is 1.7%, 2.6%, 3.0% and 7.4% in the same order.


读这些数字,有三处容易被写大。第一,主打能力没有分数:多轮编辑和透明底都没有量化指标,三张评测图考的是场景可用率、文字与真实感、文字崩坏,没有一张回答”改到第几轮会崩”。百轮的证据是一段 39 秒的视频,不是可复现的测试。第二,+55.62% 的对手是上一代 5.0 Pro,不是竞品;对 Nano Banana 2.1 的图生图真实感反而是 −6.18%,这一格输了。官方原话只说”文字结构、文生图真实感方面领先”,范围圈得很准,转述成全面领先就超出了出处。第三,”误差归零”是媒体的说法,官方用的是”像素级保持””消除色彩偏移”,属于定性描述,既没有误差指标,也没有第三方复测。此外,评测专家由官方组织,MagicArena 只是平台,这不等于独立盲测。

Three points here are easy to overstate. First, the headline abilities have no scores: neither multi-round editing nor transparency has a quantified metric, and the three charts test scenario usability, text and realism, and text collapse, none of them answering “at which round does it break.” The 100-round evidence is a 39-second video, not a reproducible test. Second, the +55.62% is against the previous-generation 5.0 Pro, not a competitor; against Nano Banana 2.1 on image-to-image realism it is actually −6.18%, a loss. The vendor’s own wording says only “leading in text structure and text-to-image realism,” a carefully bounded claim, and paraphrasing it as leading across the board goes beyond the source. Third, “zero error” is the media’s phrase; the vendor says “pixel-level preservation” and “eliminating color shift,” which are qualitative, with no error metric and no third-party retest. Also, the evaluators were organized by the vendor and MagicArena is only the platform, which is not the same as an independent blind test.


入口和价格方面,官方只说正陆续上线即梦、豆包、豆包工作和火山方舟,没有给任何价格。以下来自新京报、快科技和 AIBase 三家一致的报道,我按媒体口径标注:方舟体验中心的模型名为 Doubao-Seedream-5.5;计费延续”输入图加输出图”两段制,文本 Token 不收费;2K 方图编辑一张约 0.28 元,官方称比上一代便宜 60%、快近 40%;API 本周在火山方舟上线。有一笔账要自己算:按两段计费,多轮编辑的每一轮都是一次完整计费,100 轮就是 100 次。这是我的推论,官方没有给多轮场景的计价细则。真要做百轮打磨,成本随轮次线性叠加,这个前提比单价更影响预算。

On access and pricing, the vendor says only that the model is rolling out on Jimeng, Doubao, Doubao Work and Volcano Ark, with no price at all. What follows comes from three consistent press reports (Beijing News, Kuai Ke Ji and AIBase), and I label it as press figures: the model name in the Ark experience center is Doubao-Seedream-5.5; billing keeps the two-part “input image plus output image” scheme with no charge for text tokens; one 2K square edit costs about 0.28 yuan, which the vendor says is 60% cheaper and nearly 40% faster than the previous generation; the API goes live on Volcano Ark this week. There is a sum you have to do yourself: under two-part billing, every round of multi-round editing is a full charge, so 100 rounds means 100 charges. That is my inference, since the vendor gave no pricing detail for multi-round use. For a real hundred-round polish, cost stacks linearly with rounds, and that premise affects the budget more than the unit price does.


我的判断是,三项升级里真正改变工作位置的是 RGBA 透明素材。多轮编辑解决的是心理成本,让你敢改;透明底解决的是产出物的形态,让模型从”出图工具”变成”出素材的地方”,也更容易嵌进设计团队现有的流程。百轮编辑真正的瓶颈在计费,不在模型:改到第 95 轮,图更准了,代价是 95 次计费。做一张 5000 元的主图,这点成本可以忽略;做一批 20 元一张的白牌图,就该回到重生成。官方还留了两句话,一句是 5.5 系列后续将推出 Pro 版本,另一句是未来支持更复杂的上下文输入和创作指令,并把能力推向 Agent 工作流。方向清楚,但目前都是计划,不是能力。你自己会怎么选,是续改,还是重新生成?

My verdict is that, of the three upgrades, the one that really changes where the work happens is RGBA transparent assets. Multi-round editing solves a psychological cost and makes you dare to edit; transparency changes the form of the output, turning the model from a picture tool into a place that produces assets, and it slots more easily into a design team’s existing pipeline. The true bottleneck of hundred-round editing is billing, not the model: reaching round 95 makes the picture more accurate, at the price of 95 charges. For a 5,000-yuan hero image the cost is negligible; for a batch of 20-yuan generic product shots you should go back to regenerating. The vendor also left two doors open: a Pro version of the 5.5 series is coming, and more complex context input and creative instructions will be supported, pushing the capability toward agent workflows. The direction is clear, but both are plans, not abilities. Which would you choose yourself: keep editing, or regenerate?