Show, Don't Tell: Morphing Latent Reasoning into Image Generation

TL;DR

This work tackles the bottlenecks of explicit reasoning in text-to-image generation by introducing LatentMorph, which conducts reasoning directly in continuous latent space rather than through decoded text. The model interleaves latent thoughts with generation via a short-term and long-term visual memory, a latent translator, and a latent shaper, guided by an RL-trained adaptive invoker. It demonstrates significant gains in fidelity, abstract reasoning, and efficiency across multiple benchmarks, while achieving higher cognitive alignment with human intuition. The approach is model-agnostic, enabling seamless integration with autoregressive generators and offering a scalable path toward cognitively aligned, self-refining visual synthesis.

Abstract

Text-to-image (T2I) generation has achieved remarkable progress, yet existing methods often lack the ability to dynamically reason and refine during generation--a hallmark of human creativity. Current reasoning-augmented paradigms most rely on explicit thought processes, where intermediate reasoning is decoded into discrete text at fixed steps with frequent image decoding and re-encoding, leading to inefficiencies, information loss, and cognitive mismatches. To bridge this gap, we introduce LatentMorph, a novel framework that seamlessly integrates implicit latent reasoning into the T2I generation process. At its core, LatentMorph introduces four lightweight components: (i) a condenser for summarizing intermediate generation states into compact visual memory, (ii) a translator for converting latent thoughts into actionable guidance, (iii) a shaper for dynamically steering next image token predictions, and (iv) an RL-trained invoker for adaptively determining when to invoke reasoning. By performing reasoning entirely in continuous latent spaces, LatentMorph avoids the bottlenecks of explicit reasoning and enables more adaptive self-refinement. Extensive experiments demonstrate that LatentMorph (I) enhances the base model Janus-Pro by $16\%$ on GenEval and $25\%$ on T2I-CompBench; (II) outperforms explicit paradigms (e.g., TwiG) by $15\%$ and $11\%$ on abstract reasoning tasks like WISE and IPV-Txt, (III) while reducing inference time by $44\%$ and token consumption by $51\%$; and (IV) exhibits $71\%$ cognitive alignment with human intuition on reasoning invocation.

Figures (9)

• Figure 1: The comparison of reasoning-augmented image generation paradigms: external-loop, internal-loop, and LatentMorph.
• Figure 2: Overview of LatentMorph. LatentMorph seamlessly integrates implicit reasoning into the autoregressive generation stream. (Middle) Dynamic Monitoring: a short-term condenser compresses recent hidden states $\mathbf{H}_{i-w:i}$ into a local memory, enabling the reason invoker $\mathcal{I}_\text{invoker}$ to adaptively trigger reasoning interventions. (Bottom) Latent Reasoning: upon invocation, a long-term condenser summarizes the global history $\mathbf{H}_{1:i}$ for $\mathrm{UMM}_u$. The resulting latent thoughts $\mathbf{z}$ are transformed by the translator and shaper into control tokens $\mathbf{E}_\text{ctrl}$, which are injected directly into the generator's KV cache to steer subsequent synthesis without explicit text decoding.
• Figure 3: Case study of LatentMorph. The blue stars denote the adaptive reasoning invocations. These interventions align with critical semantic transitions, enabling LatentMorph to correct object omissions or counting errors observed in the baseline without reasoning.
• Figure 4: Qualitative comparison of LatentMorph.
• Figure 5: (Left) Evaluation results on WISE and IPV-Txt. (Middle) Qualitative examples on "impossible prompts" of IPV-Txt. (Right) Differential heatmap between latent and explicit reasoning, highlighting the information loss incurred by discrete text thoughts.