Organizations: Fudan University, Shanghai 200433, China · Shanghai Academy of AI for Science, Shanghai 200030, China · Shanghai Innovation Institute, Shanghai 200030, China
Recent breakthroughs of transformer-based diffusion models, particularly with Multimodal Diffusion Transformers (MMDiT) driven models like FLUX and Qwen Image, have facilitated thrilling experiences in visual generation. However, these models rely only on the interactions between textual conditions and visual features to produce semantically aligned images. Once the interactions fail to reflect the nuanced compositional structure of the prompt, the generated images might be unsatisfactory. Thus, a comprehensive understanding of how different blocks and their interactions with textual conditions is crucial for better understanding the intrinsic attributes and for enhancing their interactions accordingly to strengthen the prompts adherence. In this paper, we first develop a systematic pipeline to comprehensively investigate each block's functionality by \textit{removing}, \textit{disabling}, and \textit{enhancing} textual hidden-states at corresponding blocks. Our analysis reveals that 1) semantic information appears in earlier blocks and finer details are rendered in later blocks, 2) removing specific blocks is usually less disruptive than disabling text conditions, and 3) enhancing textual conditions in selective blocks improves semantic attributes. Building on these observations, we propose \method, a novel inference-time method for tailoring block-wise textual guidance. Our approach not only improves text-image alignment but also enables a range of downstream applications, including precise editing and inference acceleration. Extensive experiments demonstrated that our method outperforms various baselines and remains flexible across text-to-image generation, image editing, and inference acceleration. Our method improves T2I-Combench from 56.92% to 63.00% and GenEval from 66.42% to 71.63% on SD3.5, without sacrificing synthesis quality.
Text-to-image diffusion models like Stable Diffusion generate high-quality images from text, but lack a way to inject visual guidance (e.g. sketches, styles) at inference without retraining. Existing methods either require computationally expensive fine-tuning or rely on style transfer techniques that risk semantic misalignment with textual prompts. We introduce Visual Concept Fusion (VCF), the first method offering dual conditioning on both an image and text prompt at inference time without any concept-specific training. VCF enables visual concept injection into Stable Diffusion by aligning CLIP image features with the text embedding space. VCF consists of three components: (1) a lightweight aligner that maps image tokens to the text embedding manifold using InfoNCE and cross-attention reconstruction losses, (2) a fusion strategy that preserves both textual and visual semantics, and (3) an optional Prompt-Noise Optimization (PNO) module for test-time refinement. Our experiments demonstrate that VCF successfully transfers visual attributes including style, composition, and color palette from reference images while maintaining prompt adherence. Quantitative results show a trade-off between text alignment (CLIP score) and visual correspondence (LPIPS), with VCF outperforming baselines in reference fidelity.
While text-to-image diffusion models achieve impressive visual quality, they frequently struggle to maintain precise alignment with complex compositional prompts. An effective strategy is to improve the inference process of diffusion models, thereby better leveraging their pretrained priors to address misalignment. Existing training-free methods can be divided into two categories. The first category focuses on improving the randomly sampled initial noise, either performing costly search over noise pools or manipulating sampled noise without ensuring reliable semantic injection. The second category focuses on improving the denoising trajectory, lacking explicit mechanisms to timely diagnose and correct semantic errors. we propose \textbf{AnchorSteer}, a training-free framework that exerts fine-grained control over \textbf{both initialization} and \textbf{the denoising trajectory}. AnchorSteer consists of two synergistic components: \textbf{Semantic Anchoring} replaces uninformative Gaussian noise with text-aligned initializations via CLIP-based prior extraction and a novel Latent-Prior Score Distillation Sampling (LP-SDS) objective. Specifically, LP-SDS distills CLIP visual priors into the knowledge distribution of diffusion models, mitigating the domain gap between CLIP-based priors and diffusion-based priors. \textbf{Reflective Steering} transforms passive denoising with an active Think--Erase--Retouch loop that enables mid-generation self-correction. It leverages VLM-based diagnosis to detect semantic deviations and performs targeted latent refinement to suppress erroneous content and recover missing attributes. Extensive experiments on GenEval and T2I-CompBench++ demonstrate that AnchorSteer consistently outperforms existing baselines in text--image alignment while preserving high visual quality.
Omnimodal generation is central to a wide range of content creation and editing applications. In-context conditioning is essential to this paradigm. It allows diffusion transformers to process text instructions and visual references in a shared attention sequence. However, each reference image introduces thousands of tokens. Computation therefore grows rapidly with the number of references. Existing methods reduce computation through structured sparse attention, which limits interactions between reference and target tokens. This structure also makes the reference K and V independent of the denoising target, allowing them to be computed once and reused across steps. However, it blocks visual references from attending to the text instruction. This substantially degrades instruction following and reference fidelity in multi-reference editing. To resolve this conflict, we jointly redesign the token sequence and attention mask. Our beyond-mask design uses static text anchors to connect the instruction to the reference branch. It preserves exact K and V reuse without adding parameters. However, this direct architectural conversion degrades generation quality. We recover the lost performance through teacher-forced velocity distillation, followed by a short on-policy stage in which the teacher supervises student-visited states. To our knowledge, this is the first use of on-policy distillation for architectural recovery in diffusion models. Across three image-editing benchmarks, our method matches full-attention generation quality. With five reference images, it accelerates the complete 40-step denoising process by 3.92x, while static text anchors introduce negligible runtime overhead; the speedup reaches 5.47x at ten references in our scaling study.