Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling

Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling Yuran Wang1,2* Bohan Zeng1,2* Chengzhuo Tong1,2 Wenxuan Liu1 Yang Shi1,2, Xiaochen Ma1 Hao Liang1 Yuanxing Zhang2 Wentao Zhang1† 1Peking University 2Kling Team, Kuaishou Technology Abstract Subject-driven image generation has advanced from single- to multi-subject composition, while neglecting dis- tinction, the ability to identify and generate the correct subject when inputs contain multiple candidates. This limitation restricts effectiveness in complex, realistic vi- sual settings. We propose Scone, a unified understanding- generation method that integrates composition and distinc- tion. Scone enables the understanding expert to act as a semantic bridge, conveying semantic information and guid- ing the generation expert to preserve subject identity while minimizing interference. A two-stage training scheme first learns composition, then enhances distinction through se- mantic alignment and attention-based masking. We also in- troduce SconeEval, a benchmark for evaluating both com- position and distinction across diverse scenarios. Exper- iments demonstrate that Scone outperforms existing open- source models in composition and distinction tasks on two benchmarks. Our model, benchmark, and training data are available at: https://github.com/Ryann-Ran/Scone. 1. Introduction Image generation methods [3, 8, 36] have demonstrated ex- ceptional capabilities, enabling the generation of desired images across diverse scenarios [35]. Subject-driven image generation has recently gained significant attention, with the focus evolving from single-subject to multi-subject gen- eration, incorporating more input images. Existing meth- ods [36, 37, 39, 40] can process two or more input im- ages and combine subjects based on instructions. Moreover, methods such as [8, 44] extend this capability by accepting more than four images, showcasing potential for more com- plex composition tasks. However, existing works primarily focus on expanding subject combinations while neglecting the ability to dis- *Equal contribution †Corresponding author: wentao.zhang@pku.edu.cn GPT-4o Gemini-2.5 -Flash-Image USO (a) Problem The third boot from the right in Image 1 lies on a muddy path. Subject redundancy Subject omission The baby lion in image 1 is gazing toward the horizon. Subject redundancy The bird with a purple neck and a blue belly is walking on street. Success Und. Und. + Gen. Subject error (c) Challenge 2 Understanding Generation (b) Challenge 1 Success Subject error : Target Scone (Ours) Figure 1. The distinction problem and challenges. (a) Prob- lem. State-of-the-art methods have limitations in distinguishing target subjects specified by the instruction. (b) Challenge 1: semantic deficiency in generation. Reference image informa- tion from the understanding and generation experts in the unified model is used to compute semantic similarity with instruction. (c) Challenge 2: biased understanding and misaligned generation. “Und.” and “Und.+Gen.” indicate whether texture information from generation expert in the unified model is included to collabo- rate with understanding expert. The unified model is BAGEL [7]. tinguish target subjects in complex contexts. As shown in Fig. 1(a), although current models can combine multiple subjects, they may fail to distinguish and generate the cor- rect target subject when a reference image contains multiple candidates, leading to problems such as subject omissions (none of the candidate subjects appear) or errors (misiden- tification of the target subject). Real-world images often in- volve interference and intricate details [19, 32], further lim- iting practical performance. Thus, we emphasize examining the input subjects themselves, focusing on the model’s abil- ity to distinguish the target subject within complex con- texts and leverage this information for generation. A core challenge is extracting useful information from complex references, which remains difficult for genera- tion models. Subject distinction relies on semantic under- standing of instruction’s expression of references, where understanding models are more proficient [1, 17, 47]. As 1 arXiv:2512.12675v1 [cs.CV] 14 Dec 2025 shown in Fig. 1(b), in a unified understanding-generation model consisting of an understanding expert and a genera- tion expert, the information encoded by the understanding expert is more similar to the instruction, which means more aligned with instruction than that encoded by the genera- tion expert, revealing generation models’ deficiency and un- derstanding model’s advantage in interacting with instruc- tions and semantically understanding reference informa- tion. However, this semantic advantage of understanding models is not entirely reliable: understanding models of- ten exhibit biases [14, 18, 31, 46], which become problem- atic when directly used to assist generation. A

This content is AI-processed based on open access ArXiv data.