Vision-Language Binding in In-Context Image Generation

In-context image generation models such as FLUX.2, in addition to a text prompt, also take an optional reference image that tells the model what visual content to preserve. Internally, text, reference image, and the noise tokens are concatenated and processed through a single attention stream, where all tokens can attend to one another. This leaves open how reference information is actually distributed across these three token sets during the forward pass through the model, and how the relevant reference information is ultimately carried over to the noise tokens to produce the output image. In this work, we show that an implicit cross-modal binding emerges between the text tokens and the reference image: the text tokens absorb a portion of the reference content information during the forward pass, and that absorbed content causally influences the generated output. We surface this binding with three causal interventions on FLUX.2: T2I Lens, which decodes intermediate text-token activations through a text-to-image path; Attention Knockout, which severs specific attention edges; and I2I-to-I2I Patching, which copies text token activations between editing runs. Across 2,875 image editing tasks drawn from SUN397, DreamBench++, and a controlled sweep over color-object and style-subject combinations, we observe a consistent division of labor: language-describable properties of the reference image like color, style, and scene setting are first written into the text tokens, which carry them to the generated image; pixel-exact properties like a specific face or instance identity bypass the text tokens and flow directly from reference to image through image-to-image attention. We further localize the reference-text binding to the padding tokens of the text sequence, exposing the maximum text length as an implicit capacity hyperparameter for cross-modal binding. These results show that text tokens in a multimodal DiT are not just prompt holders, but a structured channel for the part of a reference image that language can describe. More broadly, they suggest that even in unified-attention multimodal generative models, token modality structures how conditioning information is represented and routed across the network.