ByteDance releases image generation AI 'Seedream 4.0,' capable of generating 4K resolution images and featuring image editing capabilities, outperforming Google and OpenAI in some tests
ByteDance, the company behind TikTok, has released the image generation AI ' Seedream 4.0 .' Seedream 4.0 can generate images up to 4K resolution, and supports not only simple 'image generation from text' but also image generation using model knowledge and inference. It also features high-quality image editing functions, and has recorded scores higher than products from Google and OpenAI in various benchmarks.
Seedream 4.0
Seedream 4.0 can generate new images while inheriting the characteristics of the input image. For example, along with the following image, 'Refer to this image, generate three images from different views: 360-degree overhead view, an elevated view, and a back view. Don't change the girl's appearance and outfit.'
The generated result is shown below. The subject's appearance, clothing, hairstyle, and surrounding scenery are all preserved.
You can also input a logo image to generate multiple product images.
The generated result looks like this. We were able to generate images that look like the logo is printed directly onto paper bags, hats, etc.
You can also input multiple images and combine each element into a single image.
Elements contained in the input image, such as a woman in a white dress, a silver bike, and a silver parachute, were retained as they were.
It is also possible to generate images using the knowledge and inference of AI models. For example, by simply providing an equation, the system can infer the intermediate steps and the answer, and output the results as an image.
Just say, 'Write an ice cream recipe in English,' and it will print out detailed instructions.
You can also edit the input image while preserving most of it. In the photo below, we ask you to remove the person in the foreground.
The part where the person was in the photo has been completely erased. The key point is that the other parts remain the same.
You can also rewrite the text.
You can also colorize black and white photos. The input image looks like this.
The coloring results are as follows:
It is also possible to change the overall atmosphere of the image.
You can make it anime-style or art-style.
The following figure compares the image generation benchmark results of 'Seedream 4.0,' 'Gemini 2.5 (Nano Banana),' 'GPT Image 1,' 'Seedream 3.0,' and 'Qwen-Image.' Seedream 4.0 outperforms Nano Banana in all tests and even scores higher than GPT Image 1 in some tests.
Here's a comparison of image editing performance, with Seedream 4.0 achieving the highest scores in most tests.
Seedream 4.0 is already available for multiple platforms, including FAL and ComfyUI , and a prompt guide from ByteDance is available at the link below.
Seedream 4.0 User Guide - Feishu Docs
https://bytedance.larkoffice.com/docx/PBvldM6Xlo5OHKxsRNVcyAq4nFe
Related Posts:
