Group 1 - The core point of the article is the official launch of the Doubao-Seedream-5.0-Lite image creation model on February 13, which is now available at the Volcano Ark Experience Center [1] - The model's API service is set to be launched in mid to late February [1] - Compared to version 4.5, Seedream 5.0 Lite has significantly improved capabilities in three areas: cross-modal understanding and reasoning, precise instruction adherence, and real-time retrieval through networking [1] Group 2 - The model introduces a new real-time retrieval enhancement capability, allowing it to access the latest knowledge and information via the internet [1] - This enhancement enables the model to respond more accurately to time-sensitive creative demands [1]

嘻疯 发自 凹非寺 量子位 | 公众号 QbitAI 豆包升级上新，火山引擎带着图像创作模型Doubao-Seedream-4.5来了。 新模型有三个主打点。 一是强化了原图保持能力，最大化保持原图的人脸、光影与色调、画面细节，可以用来P图。 例如"只保留绿线中的人物，将其他角色都删掉"： 再复杂一些，将白天变为黑夜： 把图片中的英文转成手写体中文： 二是重点强化了多图组合生成能力。 在官方展示中，输入8张参考图，并指定画面布局后，让它生成图画故事书封面： 童话故事书封面：小女孩与小狐狸站在发光森林小屋前，月亮巨大而梦幻，星尘在他们周围飘浮；萤火虫的光点点亮草地；小白花细致点缀； 雾气营造柔和深度；古铜色童话边框华丽包围整个场景；色调是蓝紫与暖金对撞；角色面部特征保持原图一致；整体梦幻、温柔、魔法感强 烈，适合作为儿童绘本封面。 Seedream-4.5能精准执行复杂指令，将多种元素精准识别提取出来，并自然融合： 同样地，让多个角色"拍"一张大合照： 模型也能生成无违和感的群像画面： 反过来，根据一张参考图，一次性生成6张海报，比例分别改成1:1、2:3、4:3、16:9、1:2、9:16： 它能保持风格和元 ...

Core Insights - The article introduces the new image creation model Doubao-Seedream-4.5 from Volcano Engine, highlighting its advanced capabilities in image editing and generation [1][3]. Group 1: Key Features of Doubao-Seedream-4.5 - The model enhances the ability to maintain original image details, including facial features, lighting, and color tones, making it suitable for photo editing [4]. - It significantly improves multi-image combination generation, allowing users to input multiple reference images and specify layouts to create cohesive storybook covers [11]. - Doubao-Seedream-4.5 can execute complex instructions accurately, extracting and blending various elements naturally [13]. Group 2: Creative Applications - The model can generate seamless group portraits and create multiple versions of a reference image in different aspect ratios [19][20]. - It demonstrates strong capabilities in style transfer and material reconstruction, allowing for the generation of creative images for different sports [22][25]. - The model also optimizes poster layout and logo design, showcasing its versatility in various design applications [27]. Group 3: Performance Improvements - In internal benchmark tests, Doubao-Seedream-4.5 shows comprehensive improvements over its predecessor, Seedream 4.0, in instruction adherence, consistency, and aesthetic performance [33]. - The model is now fully open for API use by enterprises and has entered public testing, allowing users to generate up to 200 images for free after registration [35][36]. Group 4: User Experience and Feedback - Users have reported excellent performance in light and environment adjustments, with the model effectively integrating additional elements into scenes [66][71]. - The model's understanding of prompts is strong but relies on clear and specific descriptions for optimal results [72]. - Official guidelines suggest using concise and coherent natural language to describe the subject, action, and environment for better image generation outcomes [75].