ByteDance's Seedream 5.5 Launches on Doubao and Jimeng, Tested with 100 Rounds of Continuous Editing
ByteDance's Seedream 5.5 has launched on Jimeng, Doubao, and Doubao Work, focusing on precise editing and multi-round consistency. In the author's tests, Seedream 5.5 maintained better fidelity than GPT-Image 2.6 in a 100-round continuous editing task, with the latter beginning to distort around the 10th round. According to official information, the Seedream 5.5 API continues the two-part billing of input images plus output images, with text tokens not billed; 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 65
The author conducted hands-on tests involving 100 rounds of continuous editing, multi-image composition, and layer separation, demonstrating the new model's performance in multi-round consistency and precise editing.
Topics
Body · AI translation
Seriously, even something as powerful as the newly released GPT-Image 2.6, when faced with a high-pressure task like continuously editing an image for 100 rounds, has to shout that famous line:
I'm drawing a prison for myself, Lina!
For example, we first feed Image 2.6 the original image below:
Then we throw in a prompt containing 100 rounds of tasks, where each round asks it to increase the time on the watch by 1 second, and use the image generated in the previous round as the input for the next round:
But what looks like a very simple task produced results from Image 2.6 like this (to make the changes easier to see, we turned the 100 images into a video)...
You can clearly see that, in fact, around round 10, the original image begins to distort, and the further it goes, the more severe the image degradation becomes.
If this were put into a real work scenario, it would be a hu-u-uge problem.
But!!!
A newly released Chinese image-generation AI can actually handle this high-pressure task! Let's look at the results:
Although there is still some distortion in the later rounds, overall it is much more faithful than Image 2.6.
And this new Chinese image-generation AI is Seedream 5.5 from ByteDance.
The task above was executed by calling the new model in Doubao.
However, not collapsing after 100 rounds of editing is only one small part of this Seedream 5.5 update. Overall, its highlights also include:
Not only that, Seedream 5.5 now generates much faster than previous versions; at the same time, according to official information, the Seedream 5.5 API continues the two-part billing structure of "input image + output image," with no charge for text tokens. 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.
Talk is cheap, so next, let's start a round of in-depth hands-on testing~
More complex continuous editing
In the first round of testing, we go back to Doubao for an even more complex continuous editing task.
The 100-round editing just now only focused on the time on the watch in the image; this time, based on a realistic living room photo, we continuously edit various details:
To make the comparison easier to see, we again turned the 10 photos into a video:
This round tests exactly the large model's memory.
Because the input for each round is only the output of the previous round, and the model cannot see the original image, any mistake in a later step will lead to mistakes in every subsequent step, and errors will accumulate.
And this is exactly what Seedream 5.5 is emphasizing this time: precise editing and multi-round consistency.
Compositing a high-end advertisement
In the second round of testing, let's do a task close to a work scenario: multi-image composite advertising.
To do this, we prepared three input images in advance: a female model image, a handbag image, and a background image:
The prompt is also relatively simple:
Please fuse the three uploaded reference images into a key visual for a high-end fashion brand's handbag advertisement. Output a 16:9 horizontal, ultra-realistic photographic work at as high a resolution as possible.
Just 1 minute later, the high-end product advertisement shot was ready:
The shooting conditions of these three input images are completely different: the model is under soft studio lighting, the handbag is under side lighting as a still life, and the background is direct outdoor sunlight.
The difficulty of multi-image fusion lies in making them look as if they were shot at the same time.
Looking at the final image, the model has changed from standing to sitting, while the suit fabric, the handbag's slanted flap and metal ring handle are all preserved, and the handbag's size also matches real proportions.
This is exactly Seedream 5.5's upgrade in multi-image fusion: the size, perspective, and lighting relationships between subjects are more coordinated.
This also includes some semi-transparent PNG assets; Seedream 5.5's compositing ability is now quite stable. For example, this time we directly uploaded 5 PNGs:
Then attached this Prompt:
Use the 5 uploaded transparent PNG assets to combine and re-render them into a complete cartoon jelly-style key visual. Aspect ratio 16:9, resolution 3840×2160. The theme is "Jelly Fruit Floating Paradise," like a promotional image for a dessert world from a high-quality animated film, fresh, light, with a clear composition and spatial hierarchy.
Semi-transparent assets have always been a difficult point for image-generation models.
Because if the edges of transparent PNGs are not handled well, white edges or gray edges will appear; and materials like jelly also require handling light transmission, refraction, and colored shadows.
From the results, the purple ring can show the color of the orange behind it, the ground shadows also carry the tones of their respective assets, and the front-to-back occlusion relationships are also very clear.
I thought it was a movie still
This time the rough prompt is as follows:
At five in the morning, winter in Paris, inside an old café that has not yet opened, a woman around forty sits alone by the street-facing window, wearing a dark gray wool coat, her scarf loosened, her hands naturally wrapped around the white porcelain coffee cup on the table...
Although we asked Seedream 5.5 to generate five 2K high-definition images in one go, the speed can be said to be extremely fast, completed in almost 1 minute:
To judge whether an image looks like a movie still, the first thing to look at is the light.
There are two light sources in this image: the cool daylight from outside the window in the early morning, and the warm yellow wall lamps inside the shop. The warm-to-cool transitions on the character's face and coat are very natural.
In terms of detail, there is a faint reflection of the character on the window glass, and the street and car lights outside the window are rendered with shallow depth of field; the character's wrinkles, loosened scarf, and slightly weary expression also make her look more like a real ordinary person.
For the film and television industry, an image like this can already be used for early-stage concept art, storyboards, and atmosphere reference.
Posters with tons of text are easily handled
To this end, we had Seedream 5.5 create one around the recently popular "world model":
This poster has extremely high information density. From the main title, module titles, and small body text to the source information in the footer, there are at least four or five levels of font size.
Even when zoomed in to look at the smallest level, the text remains clear, with almost no garbled characters.
English text is being prepared.
This corresponds exactly to Seedream 5.5's small-text and multilingual rendering, as well as its image-text layout capabilities.
It can "imitate" an action at a glance
Another feature of the upgraded Seedream 5.5 is that it can learn complex actions at a glance.
For example, we prepared a recently popular "flash step" photo and an image of a person:
We fed both images to Doubao at once, and then simply said:
Make the person in the reference image perform the action of the person in the original action image, and output a 16:9 image.
Again taking only 1 minute, the action was transferred directly:
The difficulty of this task is that the action has to come from the first image and the person from the second image, and the information from both sides must not get mixed up.
From the result, the wide-stride stance, shift in center of gravity, and arm positions were all transferred over, and even the head-turned-to-the-side expression followed along; while the person's appearance, clothing, and studio background all remained consistent with the original image, and the aspect ratio was also expanded to 16:9.
Moreover, even when inputting images of 9 people at once, Seedream 5.5 can still steadily generate them in the specified poses:
Behind this is Seedream 5.5's understanding of human body poses and its preservation of character consistency.
Split into editable layers
The following test is a reverse operation.
We first prepared a finished image, with the goal of first extracting the key elements in the image, and then editing them into different versions as needed.
The prompt for splitting layers is as follows:
Please split this finished perfume advertisement image into layers that can continue to be edited, keeping the original 16:9 canvas, original composition, proportions, colors, and object positions. Use actual layer splitting and text recognition functions, and do not generate a "layer schematic collage".
As can be seen, the product image has been split by Seedream 5.5 into different layers as required.
Next, we continued to have it make detailed adjustments to turn it into a product image suitable for the English market. The prompt is roughly as follows:
Create this market version based on the original advertising project that has already been split. Do not translate the previous market version sequentially. Output a complete 16:9, 3840×2160 advertisement image, and, where the entry point supports it, also retain an editable project.
Lock the background, perfume bottle, gift box, bergamot, cedar branch, and their shadow layers, keeping their position, size, perspective, color, material, and lighting unchanged. Keep STILL, 07, EAU DE PARFUM, and 50 mL on all product packaging, and do not translate or redraw the packaging labels. The STILL wordmark, AUTUMN EDITION / 07, 2026, and the EAU DE PARFUM at the bottom in the image should also remain unchanged.
Why split layers?
Comparing the two images, it can be seen that the background, perfume bottle, gift box, bergamot, cedar branch, and light and shadow are all consistent with the original image, and not a single character of STILL, 07, or EAU DE PARFUM on the packaging has been changed.
English text is being prepared.
What is used in this step is precisely Seedream 5.5's layer splitting and secondary text editing capabilities.
For brands going overseas, this means that one set of key visuals can quickly produce versions for multiple markets.
GIF? Of course it can do that too
English text is being prepared.
The image still directly uses the earlier "flash step" case; this time we make it move. The prompt is as follows:
Search online for a complete video or GIF of "flash step", and use the person just generated and Seedream 5.5 to make a GIF, 16:9, no larger than 10M, and keep it as high-definition as possible.
Then, a GIF that meets the requirements was born just like that:
So, from the overall test results, what Image 2.6 can do, Seedream 5.5 can also do; what Image 2.6 cannot do, Seedream 5.5 can still do.
Fragrant!
What is lacking is an AI that can remain stable when editing and modifying 50 times, or even 100 times.
This is actually also the point that impressed us most about Seedream 5.5 during this test.
And accurate and stable modifications are exactly what e-commerce scenarios need most.
Product images and e-commerce main images almost need secondary modifications every day—changing a color, changing an outfit, changing a background, or changing a line of copy. Now all of these can be completed continuously on the same main image, and a whole set of SKU versions can be derived from one image; plus transparent image direct output, layer splitting, and multilingual text changes, several intermediate steps can be saved.
Film and television and content production can use it as well.
Because when multiple images are fused, the size, perspective, and lighting between characters, props, and scenes can all match, reference images can be directly composited into concept art and storyboards; and a sense of realism like that of a Paris café is also enough to enter early-stage visual development.
Design and marketing materials are the same.
In posters, infographics, and event pages, the most troublesome part is the small text. In the past, designers had to add and fix text one by one; now dense copy and multilingual text can be laid out directly.
Add to that the price advantage mentioned earlier—capabilities close to the first tier, while the price is moving downward—for small and medium-sized merchants and content teams that need to produce a large number of images every day, this is an easy calculation.
Finally, let's return to the "celebrity quote" at the beginning.
On a task like 100 rounds of editing, which most tests the basics, Image 2.6 "drew a prison for itself", while Seedream 5.5 proved that domestic models have already moved ahead of top international models in multi-round editing stability.
At present, Seedream 5.5 has already launched on Jimeng, Doubao, and Doubao Work. Interested friends can try it out for themselves~
Let's also enjoy some stunning images generated by Seedream 5.5:
AI Trends