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

Forum topic · 小凯 · 2026-06-25

Summary

IV-CoT (Implicit Visual Chain-of-Thought) is a latent visual reasoning framework for query-conditioned text-to-image generation, presented in arXiv paper 2506.14664 by Zixuan Li, Haokun Lin, and Yicheng Xiao. Unified multimodal large language models (MLLMs) produce high-quality images but often fail at structure-aware prompt following, where object counts, spatial relations, attribute bindings, and coarse layouts must be preserved. The authors attribute this to entanglement of structural planning and appearance rendering in a single conditioning stream. IV-CoT decomposes visual conditioning queries into a structural-to-semantic cascade: structural queries first form a latent visual plan, then semantic queries render appearance conditioned on that plan. Training-only sketch supervision guides the structural queries, so no sketch extraction or intermediate decoding is needed at inference. IV-CoT performs implicit chain-of-thought reasoning in a single forward pass, achieving strong results on GenEval and T2I-CompBench, with analysis showing structural and semantic queries play complementary roles.

Paper Overview

Field: Computer Vision Authors: Zixuan Li, Haokun Lin, Yicheng Xiao Published: 2026-06-24 arXiv: 2506.14664

Abstract

Unified multi-modal large language models (MLLMs) have achieved strong text-to-image generation quality, but still struggle with structure-aware prompt following, where object counts, spatial relations, attribute bindings, and coarse layouts must be preserved. The authors attribute this limitation in part to the entanglement of structural planning and appearance rendering within a single conditioning stream.

To address this issue, they propose Implicit Visual Chain-of-Thought (IV-CoT), a latent visual reasoning framework for query-conditioned image generation. IV-CoT decomposes the visual conditioning queries into a structural-to-semantic cascade, where structural queries first form a latent visual plan and semantic queries then render appearance conditioned on this plan.

To guide the structural queries, the method introduces training-only sketch supervision, encouraging them to capture structure from sketches without requiring sketch extraction or intermediate decoding at inference time.

Key Contributions

Implicit visual chain-of-thought: performs CoT-style latent reasoning in a single forward pass, with no intermediate image decoding during inference
Structural-to-semantic cascade: separates structural planning from appearance rendering across query groups
Training-only sketch supervision: guides structural queries to capture layout and structure without inference-time overhead

Results

IV-CoT achieves strong results on GenEval and T2I-CompBench. Visualizations and analysis indicate that the learned structural and semantic queries play complementary roles in structure-aware generation.

--- *Auto-collected on 2026-06-25.*