AI Trends

ByteDance Seed team releases image model Seedream 5.5, supporting 100 rounds of continuous editing

36Kr · AIUpdated Nov 10 11:27

AI summary

The ByteDance Seed team has released a new-generation image creation model, Seedream 5.5, with a focus on enhancing multi-round editing stability, native RGBA transparent asset generation, and performance on complex text and realistic textures. Multi-round editing supports continuous modification of the same work for up to 100 rounds. The model is already live on Doubao and Doubao Work, with experience access gradually opening on Jimeng AI and Volcano Ark, and a Pro version to be released later. The API bills input images and output images separately, with text tokens free of charge; editing a single 2K square image costs about 0.28 yuan, less than half that of Seedream 5.0 Pro.

Why read this

AI score 71

The original text provides specific data on multi-round editing, transparent assets, and API pricing, along with hands-on testing details, which can be used to assess changes in the image creation workflow.

Topics

Tags

Body · AI translation

Zhidongxi, November 9 report: This afternoon, ByteDance's Seed team officially released its new-generation image creation model, Seedream 5.5. Compared with previous versions, Seedream 5.5 focuses on strengthening three core capabilities: multi-round editing stability, native RGBA transparent asset generation, and complex text and realistic texture rendering.

Overall effect of images generated by Seedream 5.5

Among these, multi-round editing supports up to 100 continuous revisions around the same work; transparent assets can directly generate transparent or semi-transparent images with an Alpha channel, reducing post-processing cutout steps; and text rendering further improves the restoration of complex glyphs and dense small text.

This update comes 4 months after the release of the previous-generation model Seedream 5.0 Pro on July 8 this year.

English text is being prepared.

Seedream 5.5 experience entry

From this real test, the overall experience of Seedream 5.5 is relatively smooth, bringing more obvious upgrades in continuous editing, transparent asset generation, and realistic texture rendering. Generation speed is fast, with a single image taking about 5-10 seconds. In particular, multi-round modifications do not require repeatedly generating from scratch, and transparent assets can also directly enter subsequent design workflows.

In terms of pricing, the Seedream 5.5 API uses separate billing for input images and output images, with text tokens free. Taking 2K square image editing as an example, the cost per image is about 0.28 yuan, less than half that of Seedream 5.0 Pro.

01. 100 rounds of continuous editing, with people, objects, and spaces all editable

Multi-round editing is one of the core capabilities upgraded in Seedream 5.5.

To address this issue, Seedream 5.5 focuses on strengthening the stability of continuous editing. According to the official description, the model can achieve pixel-level preservation of non-edited areas, reduce color shift and image quality degradation caused by continuous modifications, and support up to 100 rounds of continuous creation.

Zhidongxi tested three scenarios: people, desktop objects, and indoor spaces. Overall, Seedream 5.5 can not only modify specified objects according to continuous instructions, but also maintain the subject and composition of the existing image across multiple rounds of creation.

1. Continuous outfit changes for a person, modifying clothing without regeneration

Initial instruction: First generate a realistic photography-style full-body portrait. A young woman stands in a simple cafe, wearing a white shirt and blue jeans, holding a coffee cup. The person stands facing forward, with a clear face, and the background includes wooden tables and chairs, floor-to-ceiling windows, and an outdoor street scene. Natural lighting, realistic photographic texture.

Initial generated image

Round 1: Change the person's white shirt to a black leather jacket, keeping other content unchanged.

Round 2: Change the black leather jacket to an off-white knit sweater, keeping the person's face, hairstyle, pose, coffee cup, and background unchanged.

Round 3: Change the person's blue jeans to a black long skirt, without modifying other areas.

Round 4: Change the coffee cup in the person's hand to a transparent glass cup, keeping the hand pose and other content unchanged.

Results of continuous outfit changes for the person

In this set of tests, the clothing and furniture generated by Seedream 5.5 have good texture, and details such as clothing folds and fabric textures look relatively natural. After continuously adjusting the clothing and held objects, the person's face, hairstyle, and background environment also basically maintained their original state.

However, the model is still not accurate enough in grasping details of human body structure. In this set of images, the person's thumb is too long, and the shape does not conform to real human anatomy; after putting on the skirt, the connection between the legs and the hem is also not natural enough, and the legs appear to be floating.

Overall, Seedream 5.5 performs relatively naturally in character styling and clothing details, and can better preserve the main subject during continuous modifications, but hand structure and human posture still have room for further optimization.

2. Continuous modification of multiple objects, with desktop layout also adjustable

Initial instruction: Generate a desktop still-life photograph. The image contains an open book, a white ceramic coffee cup, a pair of black glasses, a fountain pen, and a small green potted plant. The objects are arranged naturally, the desktop is made of light-colored wood, and the image uses a top-down view with soft natural light.

Round 3: Move the fountain pen to the right side of the book, keeping the positions of other objects unchanged.

Round 4: Replace the potted plant with a bouquet of yellow tulips.

Round 5: Change the desktop material to dark walnut, keeping the shape, position, and lighting relationships of all objects unchanged.

Results of continuous modifications to desktop objects

In the desktop still-life test, Seedream 5.5 still has shortcomings in recognizing multiple objects and restoring details. In the initial generation, the model did not correctly generate the glasses required by the prompt, but instead drew them as a black rack; after subsequently modifying them to wire-rimmed glasses, the frame appeared with the error of three temples, and the object structure was not accurate enough.

In this set of tests, the model can adjust the objects and materials on the desktop according to instructions, but it is still prone to errors in the initial generation and detail restoration of small objects. Especially for objects with relatively fine structures such as glasses, although the model can generate the general appearance, it fails to accurately restore their structure.

3. Continuous renovation of indoor spaces, with perspective maintained after furniture movement

Initial instruction: Generate a modern minimalist living room interior photo. The scene includes a gray sofa, a wooden coffee table, a floor lamp, a TV cabinet, green plants, and floor-to-ceiling windows, using a wide-angle interior photography perspective, with clear furniture positions and natural lighting.

Initial generated image

Round 1: Replace the gray sofa with a dark green fabric sofa, keeping everything else unchanged.

Round 2: Move the coffee table to the left side of the sofa, maintaining reasonable perspective and floor shadows.

Round 3: Move the floor lamp next to the TV cabinet.

Round 4: Add a yellow armchair on the right side of the sofa, ensuring there is no unreasonable overlap between furniture.

Round 5: Change the scene outside the window to a rainy day, keeping the indoor furniture, walls, and floor unchanged.

Results of continuous indoor space renovation

Continuous editing of indoor spaces is more complex than replacing a single object. After furniture is moved, the model not only needs to adjust object positions but also must re-handle perspective, occlusion, and shadow relationships.

In this set of tests, Seedream 5.5 understood the editing instructions fairly accurately. Apart from the coffee table, it was basically able to adjust the color, position, and quantity of furniture as required, and could also modify the environment outside the window. After multiple rounds of modification, the image showed no obvious spatial misalignment, and the perspective and overall layout remained relatively natural.

Judging from this set of results, Seedream 5.5 performs relatively stably in continuous editing of indoor spaces. It can not only identify the content the user wants to adjust, but also handle the spatial relationship between furniture changes and the surrounding environment fairly well.

02. Native RGBA transparent assets, AI image generation can also save the cutout step

In addition to continuous editing, Seedream 5.5 this time also added native RGBA transparent asset generation capability.

In the past, after users generated products, people, or decorative elements through AI tools, if they needed to use the subject separately, they usually had to rely on other tools to remove the background. For objects with complex details or transparent effects, processing was prone to issues such as edge residue, loss of detail, or distorted transparency.

With this upgrade, Seedream 5.5 can directly generate transparent or semi-transparent images with an Alpha channel, making the generated content a design asset that can be used independently.

Zhidongxi tested complex subjects, transparent artistic text, and secondary composition of assets.

1. Transparent generation of complex subjects, even feather edges can be preserved

Instruction 1: Generate a white snowy owl flying with wings spread, front view, both wings fully extended, rich feather layers, with the tiny gaps between feathers clearly visible, talons naturally drawn in, and bright eyes. The subject is complete, with wings and talons not cropped, the background is a truly transparent background, output in RGBA format, preserving natural feather edges and semi-transparent details.

Transparent snowy owl generation result

As can be seen, in the complex subject generation test, Seedream 5.5 was able to generate a complete snowy owl image as required, with the spread wings, layered feathers, and body outline all well presented. Compared with ordinary solid-background generation, transparent assets make it more convenient for users to place the subject separately into other design compositions.

2. Transparent artistic text, text and visual effects can be generated together

Instruction 2: Generate a set of Chinese artistic text assets with a transparent background. The text content is "Stay curious, explore the unknown." The overall design uses a three-dimensional glass material, with clear and complete glyphs, delicate blue-purple gradients and natural refraction effects inside, and soft highlights on the edges. The text must be accurate, with no added or missing characters and no altered strokes. The background is transparent, output in RGBA format.

Transparent artistic text generation result

Transparent asset generation is not limited to people and objects; it can also be used for text design.

In this set of tests, Seedream 5.5 accurately recognized the Chinese content in the prompt and generated the corresponding artistic font as required. The generated text styling was basically consistent with the preset style, the glass material effect was also well presented, the text outline was clear, and there were no obvious typos or glyph confusion.

Judging from the actual results, Seedream 5.5 can not only generate artistic text on a transparent background, but also fairly well reproduce the specified font style and material texture.

3. Secondary composition of assets, generated content can continue to be used for design

Instruction 3: Use the transparent snowy owl and transparent artistic text assets just generated to design a 16:9 horizontal nature-themed promotional poster. The background is a deep blue night sky and snow mountains, the snowy owl is on the right side of the image, and the title "Stay curious, explore the unknown" is on the left side. The overall style uses high-end nature photography and magazine visual design. Keep the snowy owl subject complete, the text clear, and the composition layered.

Poster after combining transparent assets

In the asset reuse test, Seedream 5.5 was able to combine the previously generated subject and text into a new design composition, and the snowy owl was also placed in the snow mountain background, with the overall image connecting fairly naturally. However, compared with the originally generated asset, the snowy owl's appearance and posture changed somewhat, with the body more extended and closer to a flying state, but the subject details remained basically consistent.

03. Upgrades in realistic photography and text rendering, image texture more natural

Image realism and text rendering are also key upgrade areas for Seedream 5.5 this time. According to the official description, Seedream 5.5 improved lighting, materials, and skin texture representation, and further enhanced the ability to reproduce complex glyphs and dense small text.

Zhidongxi conducted tests in four aspects: portrait photography, product photography, complex Chinese infographics, and commercial posters, focusing on observing the model's reproduction of character details, product materials, Chinese text, and image layout.

1. Portrait photography: skin texture and natural light and shadow are more important

Prompt: Generate a portrait photo in a realistic photography style. A young woman sits by the window in a café, slightly turned to look at the camera, with sunlight slanting through the window onto her face and shoulders. Preserve natural skin texture, fine pores, realistic hair strands, and slight facial imperfections; do not use skin smoothing or excessive beauty retouching. The background is a blurred café environment, shot with an 85mm portrait lens, shallow depth of field, natural skin tones.

Portrait photography generation result

In terms of character generation, Seedream 5.5 presents skin, hair, and lighting details fairly well. Faces are no longer just simple smooth-skin effects; the light-and-shadow variations under illumination, hair strand details, and skin texture also make the image closer to real photography.

Compared with simply pursuing refined facial features, this type of test focuses more on whether the image looks natural. For users who need to generate lifestyle images, character promotional photos, and commercial portraits, natural skin texture and light-and-shadow performance can enhance the usability of the images.

2. Product photography: metal and glass materials have more depth

Prompt: Generate an advertising photography image for a high-end consumer electronics product. The main subject is a pair of silver metal smart glasses; the frame uses brushed metal material, the lenses have slight transparency and natural reflections, and the temple structure is clear. The product is placed on a black marble countertop, with a dark gray studio background, using side soft light to highlight the metal edges and lens details, and the image has a realistic commercial photography texture.

Product photography generation result

Product photography has high requirements for material representation. Metal needs to show realistic reflections and textures, glass needs to convey transparency and refraction effects, and different materials must also be clearly distinguished.

From the generation results, Seedream 5.5 reproduces different materials relatively naturally. The metal frame shows the corresponding reflective texture, the transparent lenses do not have obvious cloudiness, and a reflection corresponding to the product is also generated on the marble countertop; the reflection effect is clear, and the overall image has good visual texture.

In this set of tests, Seedream 5.5 distinguishes materials such as metal, glass, and marble fairly accurately, especially in the representation of the transparent lenses and the countertop reflection, basically meeting expectations.

3. Complex Chinese infographic: can small-size text be read clearly?

Prompt: Generate a Chinese popular-science infographic about "how AI chips work," using a 16:9 horizontal design.

The image is divided into three areas:

Left: display a GPU chip, labeled "compute unit," "VRAM," and "data transmission."

Middle: display the HBM high-bandwidth memory stacked structure, labeled "DRAM chip," "through-silicon via," and "logic chip."

Right: display the process of data transmission from memory to the GPU, labeled "data read," "parallel computing," and "result output."

At the bottom, add three knowledge points: "high bandwidth," "low latency," and "high capacity."

Require all Chinese text to be accurate, with no typos, garbled characters, or repeated text. The diagrams should correspond to the text, the layout should be clear, and it should be suitable as an illustration for a technology media popular-science article.

Chinese infographic generation result

Infographics are a comprehensive test of the model's text rendering capability. Compared with ordinary posters, this type of image not only needs to generate a title but also must arrange multiple technical terms, explanatory text, and diagrams within limited space.

From the generation results, Seedream 5.5 understands the prompt fairly accurately and can organize the information hierarchy as required; the text also combines well with the corresponding chip structures and data flow. The text layout in the image is clear, with no obvious text overlap, and the overall layout has the visual style of a technology infographic.

In this set of tests, Seedream 5.5 not only completed Chinese text generation but also handled the correspondence between multiple sets of information and diagrams fairly well; overall readability and design completeness are both quite good.

4. Commercial poster: balancing visual design and text layout

Prompt: Generate a new product launch poster for a high-end consumer electronics brand, with the theme "Explore the Next Generation of Smart Living." The main subject is a silver smart device placed in a futuristic indoor space. The main title is "Make Technology Within Reach," and the subtitle is "Redefine Smart Living." Add three product selling points at the bottom: "Slim Design," "Smart Interaction," and "All-Day Battery Life." All Chinese text must be accurate and clear, with an obvious hierarchy between the title and the selling points; the overall style should be high-end commercial advertising design, 16:9 horizontal.

Commercial poster generation result

Commercial posters not only require the model to generate a visual subject that fits the theme but also need to reasonably arrange the relationship among the title, subtitle, and product selling points.

From the generation results, the images generated by Seedream 5.5 have a strong sense of commercial design, the product subject is prominent, and information such as the title, subtitle, and product selling points is basically complete; the arrangement of each part is fairly coordinated, and the overall result meets the preset requirements. In addition, the generation speed of this set of images was also relatively fast.

In this set of tests, Seedream 5.5 completed the integration of product display and text information fairly well, and the generated posters were relatively in line with expectations in both visual effect and information completeness.

04. Conclusion: from generating one image to continuously refining one image

In this upgrade, Seedream 5.5 further strengthened continuous editing, transparent asset generation, and text rendering capabilities while balancing generation quality and usage cost.

From Zhidongxi's experience, although the model still has flaws in some small details, it showed good performance in the three key upgrade directions this time. In particular, multi-round editing and transparent asset generation can already be well applied to actual image creation workflows.

This article comes from the WeChat public account "Zhidongxi" (ID: zhidxcom), author: Huang Fei, editor: Jiang Yu, published by 36Kr with authorization.

Source

36Kr · AI