IV-CoT: Implicit Visual Chain-of-Thought for Structure-Aware Text-to-Image Generation

Unified multi-modal large language models (MLLMs) often struggle with complex structural requirements in text-to-image generation, such as precise object counts, spatial relationships, and attribute binding. While these models can generate visually plausible images, they frequently swap attributes or misplace objects because structural planning and appearance rendering are entangled in a single conditioning stream. This paper asks: can we improve structure-aware generation by internalizing a chain-of-thought (CoT) reasoning process within the latent query space, without the latency costs of explicit intermediate decoding?

The authors propose Implicit Visual Chain-of-Thought (IV-CoT), a framework that decomposes visual conditioning into a structural-to-semantic cascade. Instead of generating explicit intermediate text or images, IV-CoT uses a causal self-attention mechanism in the MLLM to force structural queries to be computed before semantic queries. This creates a latent visual plan that anchors object boundaries and layouts. To ensure these structural queries actually capture spatial information, the authors introduce training-only sketch supervision, using edge-detected sketches to regularize the structural query branch. At inference time, the model performs a single forward pass, producing both structural and semantic queries that guide the diffusion generator.

IV-CoT significantly outperforms baseline unified models on structure-sensitive benchmarks like GenEval and T2I-CompBench. The structural-to-semantic cascade allows the model to better preserve object counts, spatial arrangements, and attribute bindings compared to standard flat-conditioning approaches. Crucially, because the reasoning is implicit and latent, IV-CoT maintains high inference efficiency, achieving 9–15x lower latency than explicit CoT methods that require multi-stage decoding or test-time search. Ablation studies confirm that both the ordered query cascade and the sketch-based structural supervision are necessary to achieve these gains.