Pre-trained flow-based models excel at synthesizing complex scenes yet lack a direct mechanism for disentangling and customizing their underlying concepts from one-shot real-world sources. To demystify this process, we first introduce a novel differential probing technique to isolate and analyze the influence of individual concept tokens on the velocity field over time. This investigation yields a critical insight: the generative process is not monolithic but unfolds in three distinct stages. An initial Blueprint Stage establishes low-frequency structure, followed by a pivotal Instantiation Stage where content concepts emerge with peak intensity and become naturally disentangled, creating an optimal window for manipulation. A final concept-insensitive refinement stage then synthesizes fine-grained details. Guided by this discovery, we propose ConceptWeaver, a framework for one-shot concept disentanglement. ConceptWeaver learns concept-specific semantic offsets from a single reference image using a stage-aware optimization strategy that aligns with the three-stage framework. These learned offsets are then deployed during inference via our novel ConceptWeaver Guidance (CWG) mechanism, which strategically injects them at the appropriate generative stage. Extensive experiments validate that ConceptWeaver enables high-fidelity, compositional synthesis and editing, demonstrating that understanding and leveraging the intrinsic, staged nature of flow models is key to unlocking precise, multi-granularity content manipulation.
Figures & tables
Figure 2: Probing Concept Formation Dynamics. While standard prompt guidance (a) provides a dense, entangled signal, our differential probing technique (b) isolates the dynamic influence of individual concepts. This analysis uncovers a consistent three-stage framework (c): an early Blueprint Stage dominated by structural concepts (e.g.,“riding a bike”), a pivotal mid-stage Instantiation Stage where content concepts (e.g., “panda”) are decoupled and peak in intensity, and a final Refinement Stage where signals fade.
Figure 3: Flowchart of proposed ConceptWeaver. Our framework consists of two main phases: training and inference. (a) During training , a lightweight Semantic Offset Module learns to represent a visual concept by adding a learnable offset to the key/value pairs of its corresponding concept token. This is optimized with a stage-aware loss. (b) During inference , our ConceptWeaver Guidance (CWG) applies these offsets in a stage-aware manner: structural offsets are injected during the Blueprint Stage, and content offsets during the Instantiation Stage, enabling precise compositional control.
Figure 4: Mechanism of ConceptWeaver Guidance (CWG). Our guidance modifies the standard generative path (to x0 ) by injecting learned concept shifts (red vectors) at stage-aware intervals. This timed intervention, visualized by the heatmaps, steers the trajectory from noise ( x1 ) to a customized target ( x0′ ), enabling precise compositional control.
Method
SCLIPT↑
SCLIPI↑
SDINO↑
SGPT↑
TokenVerse∗
35.20
61.75
26.31
3.2
MS-Diffusion
35.02
75.58
70.69
3.6
DreamBooth
34.50
73.14
68.96
3.7
Ours
35.22
76.13
69.57
3.8
Table 1: Quantitative comparison on single-concept generation. The metrics evaluate textual alignment ( SCLIPT ) and concept alignment ( SCLIPI , SDINO , SGPT−4o ). ↑ indicates that higher is better. The best results are highlighted in bold .
Figure 5: Comparison on multi-concept composition generation. The figure illustrates different methods’ performance in combining multiple concepts, focusing on overfitting to original reference concepts and the ability to disentangle target concepts.
Figure 6: Comparison of R2V task results. Contrasts VACE method with our approach in terms of concept disentanglement, structural preservation, and “copy-paste phenomena”.
Figure 7: Qualitative Ablation of Stage-Aware Components. We compare our full model with variants lacking key components. (a) w/o Stage-aware Opt.: Naive training fails to learn a high-fidelity concept. (b) w/o Stage-aware Guid.: Applying guidance uniformly across all stages causes structural artifacts and concept bleeding. (c) Ours: The synergy of both strategies yields the best compositional quality.
Figure 8: Analysis of the Concept Guidance Scale wc . As wc increases, the visual attributes of the generated concepts align more closely with the reference image. This demonstrates that wc provides fine-grained control over the strength of the personalization.
Metric
ConceptWeaver
MS Diffusion
Dreambooth
Concept Consistency
4.06
3.75
3.02
Text Adherence
3.78
3.14
2.88
Aesthetic Quality
4.12
3.75
3.05
Table 2: User Study Results for Concept Generation (Values represent average scores on a 1-5 scale).
Figure 9: Motion Concept Transfer with ConceptWeaver. We learn a motion concept from a single reference video (up) and apply it to new subjects defined by text prompts (down). The generated videos demonstrate high-fidelity motion replication while accurately synthesizing the new subject’s appearance, showcasing effective disentanglement of motion and content.
Figure 10: Image Editing with ConceptWeaver and FlowEdit. By integrating with FlowEdit, we apply a visual concept learned from a reference image to a source image. The resulting edited image preserves the source’s structure while faithfully incorporating the specific appearance of the learned concept.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 11: Qualitative results on single-concept generation. Each row showcases results for a specific learned concept. The first column displays the reference image used to learn the concept. Subsequent columns display generated images from different methods based on a new text prompt.
Figure 12: Qualitative results on motion transfer. Each row demonstrates the ability of ConceptWeaver to compose a subject’s identity, learned from a static Id Concept image, with complex dynamics learned from a Motion Concepts video. The generated videos successfully apply the reference motion (e.g., "surfing", "playing golf") to a completely different subject (e.g., a "dog", a "monster toy") while maintaining high subject fidelity and generating a plausible new scene. This highlights our method’s strong capability for disentangling and recombining structural (motion) and content (identity) concepts.
Figure 13: Generalization of the Three-Stage Framework Across Different Flow Models. Our differential probing of the "glasses" concept reveals a universal three-stage generative process across diverse flow models ( Wan2.2-5B , Flux , SD3 , and Ours ). All models consistently show a weak signal during the early Blueprint Stage , a sharp peak of influence during the mid- Instantiation Stage , and signal decay during the final Refinement Stage . This confirms that our discovered framework is an intrinsic property of flow-based generation, not a model-specific artifact.
Figure 14: Limitations. The first row illustrates Concept Semantic Confusion , where similar concepts (e.g., dog and doll) lead to undesirable infiltration of one concept into another during generation. The second row demonstrates scenarios where the desired generation task or concept complexity Exceeds Underlying Model Capabilities , exemplified by a rabbit attempting human-like riding motions.
Figure 15: Qualitative results on multi-concept composition generation. Our method successfully disentangles and composes multiple concepts from different source images into novel scenes. Each panel shows concepts (e.g., a specific doll , a hat , glasses , a hairstyle ) learned from one or more reference images (Concept Images) and then synthesized according to a new prompt (Generated Images).
Figure 16: Qualitative results on Stage-Aware Components. This comparison highlights how our stage-aware mechanisms are crucial for both fidelity and pose generalization. Both ablated versions, w/o Stage-aware Opt. and w/o Stage-aware Guid. , exhibit strong pose overfitting, rigidly replicating the posture of the subject from the reference image (e.g., the seated dog in the second row, the lying-down dog in the last row). In contrast, our full model (Ours) successfully disentangles the concept from its original pose, allowing it to be rendered in new, contextually appropriate poses, such as running or lying down. This demonstrates that our method achieves a higher level of compositional control.
Figure 17: Qualitative results on ConceptWeaver Guidance . We fix the CWG guidance for the first stage at g1=0,g2=0 and for the third stage at g1=3,g2=3 , to qualitatively analyze our CWG. From the results, it can be observed that the optimal effect is achieved when the CWG value in the second stage is 8, and there is a clear phenomenon of concept gradual refinement as CWG increases.
Figure 18: Image Editing with ConceptWeaver and FlowEdit. By integrating with FlowEdit, ConceptWeaver enables precise appearance swapping. Our method first disentangles a target concept’s appearance from a Reference Image—like the texture of a straw hat or the look of a VW van.
Recent advances in flow matching models have significantly improved text-to-image generation quality, but also introduce growing safety risks due to the generation of harmful or undesirable content. Existing concept erasure methods are either inference-time interventions with limited effectiveness or rely on supervised fine-tuning (SFT), which requires precisely aligned data and struggles with scalability and multi-concept settings. In this paper, we propose FlowErase-RL, the first GRPO-based framework for concept erasure in flow matching models. We reformulate concept erasure as a reward optimization problem and introduce a dynamic dual-path reward mechanism that jointly optimizes (i) a Concept Erasure (CE) reward to suppress target concepts and (ii) a Non-target Space (NS) reward to preserve generative fidelity. The two reward paths are adaptively balanced during training via a performance-driven switching strategy, enabling stable optimization without explicit supervision. Extensive experiments on nudity, object, and artistic style erasure demonstrate that our method achieves state-of-the-art erasure performance while maintaining strong image quality and semantic alignment. Moreover, it exhibits robust resistance to adversarial attacks and scales effectively to multi-concept scenarios. Our results establish a new paradigm for safe and controllable generation in flow matching models.
Yi Sun, Zhiqi Zhang, Xinhao Zhong +5
Harbin Institute of Technology, Shenzhen · Jilin University · Tsinghua Shenzhen International Graduate School, Tsinghua University +2
Flow Matching enables high-quality visual generation via continuous-time dynamics, but inference remains costly due to multiple sequential function evaluations. Existing acceleration methods reduce the number of function evaluations but often introduce additional training overhead, degrade quality, or fail to account for input-dependent variability. We propose COFLOW, an inference-time method that adaptively selects the step counts each generation based on the prompt features. Our context-aware COFLOW is trained online with an unsupervised reward that balances inference efficiency and generation fidelity. Our method is plug-and-play, requiring no retraining of the underlying generative model. It generalizes to image and video generation, achieving over 2.5x speedup while preserving perceptual and semantic quality. We further provide a theoretical analysis establishing an O(1/K) forward-Euler discretization error bound under standard regularity conditions.
We propose Perceptual Flow Matching (PFM), a simple yet effective framework for few-step generation in flow-matching models. Rather than performing velocity regression in the conventional VAE latent space, PFM supervises flow matching in a perceptual feature space using pretrained perceptual models. This simple change substantially improves the few-step generation capability of flow-matching models, reducing the number of sampling steps from 35-50 to 4-8 while preserving generation quality. Unlike existing acceleration and distillation approaches, PFM requires neither teacher models nor auxiliary score networks and can be integrated into standard flow-matching training pipelines with minimal modifications. Extensive experiments on image generation, video generation, and image editing tasks demonstrate that PFM consistently produces high-quality results while producing fewer artifacts than existing distillation-based methods. We further show that perceptual supervision shifts the regression minimizer from mean-seeking to mode-seeking, biasing predictions toward on-manifold modes that remain accurate under coarse few-step integration. Our results reveal that standard flow-matching training can naturally yield high-quality few-step generators when supervised in an appropriate representation space. We hope this insight inspires future research into representation-aware objectives for efficient generative modeling.
Chuyang Zhao, Yifei Song, Hongfa Wang +7
1Joy Future Academy · 2Fudan University · 3Tsinghua University +1