January 27th.Tencent Hunyuan(26 January) Release mixed image 3.0 Figure 1(HunyuanImage 3.0-Instruct) ModelSupport image editing and multi-graph integrationOn the other hand, the user can experience it directly at the end of the Yuanbao and in the context of the network。
Mixtures image 3.0 The total number of parameters in the map is 8 billion (approximately 1.3 billion active parameters), using a hybrid specialist (MoE) structure based on the original multimodular structure basic model of the mixed image 3.0, introducing multitask data in the map, and achieving a deep understanding and processing of the user input images and editorial commands by fine-tuning the instructions and post-training。
A hybrid image, 3.0Understand input images, thought-based image editing modelsI don't know. After receiving the images and hints entered by the user, the model first understands the content of the image itself and reasons on the basis of the user ' s hint, deduces specific areas requiring editing, detailed editorial steps, and image areas requiring retention, resulting in more detailed editorial instructions to achieve better output。
Mixage Image 3.0 Figs supports diverse photo editing and multi-graph integrationFor example, the ability to edit images by adding, deleting, changing, changing styles, repairing old photographs, character and textual changes, and extract the characters or elements from multiple photographs and produce new photos。
On the dollar treasure, the model-based users can directly produce expression kits, virtual person shoots, social sharing, electrician poster design, game role customization, creative photo production, etc。
AT THE DATA LEVEL, THE HYBRID TEAM CONSTRUCTED TENS OF MILLIONS OF-QUANTITATIVE MAPPING DATA, COVERING 80+ MISSIONS, BY MEANS OF IMAGES, VIDEO-SOURCE DATA MINING AND EXPERT NETWORK SYNTHESIS, AND INJECTED THEM INTO THE ONGOING TRAINING PHASE (CT) TO ENABLE THE MODEL TO ACQUIRE BASIC EDITING FUNCTIONS; AT THE SAME TIME, THE HYBRID TEAM FURTHER CONSTRUCTED THE THINK CHAIN FOR THE GRAPHIC DATA, ALLOWING THE MODEL TO ANALYSE USER IMAGES AND INTENTIONS FIRST AND TO OUTPUT MORE DETAILED EDITORIAL INSTRUCTIONS TO ENHANCE EDITORIAL EFFECTIVENESS。
In the post-training phase, hybrid image 3.0 using self-study MixGRPO algorithms, combined with incentive models, multi-rotational, aligning user preferences, increasing the effect of consistency between command response and non-editing areas。
1AI WITH THE FOLLOWING ENTRY:
https://hunyuan.tencent.com/chat/HunyuanDefault?from=modelSquare&modelId=Hunyuan-Image-3.0-Instruct