Region-Aware Test-Time Scaling for Compositional Image Generation

Conference: ECCV 2026
Paper: ECCV Official
Full Text Cache: paper_cache/ECCV2026/eccv-4058.txt
Area: Image Generation
Keywords: Test-Time Scaling (TTS), Compositional Image Generation, Region-Aware Generation, Diffusion Models, Chain-of-Thought Planning

TL;DR

Addressing the "scaling plateau" of global test-time exploration on complex multi-object and relational compositions, this paper introduces Region-Aware Scaling (RAS) to treat regional decomposition as a dedicated scaling axis, injecting localized guidance only during early denoising to match a 32-sample noise-scaling baseline with just 2 samples.

Background & Motivation

Recent text-to-image diffusion models and flow-matching architectures have substantially pushed the boundaries of visual fidelity and high-resolution generation. In parallel, inspired by the success of test-time compute scaling in large language models, inference-time scaling paradigms—predominantly Best-of-N candidate sampling paired with an automated verifier—have emerged as a compelling avenue to boost visual semantic alignment without parameter updates. Existing vision test-time scaling frameworks largely rely on global stochastic exploration, such as noise scaling via resampling random seeds or prompt scaling via whole-prompt rewriting and LLM augmentation.

However, these purely global search strategies become severely sample-inefficient when applied to compositional prompts that mandate satisfying multiple coupled constraints simultaneously, such as precise object counts, attribute binding, and spatial relations (e.g., "a cat below a backpack"). In these scenarios, valid structural solutions reside in extremely sparse, low-probability modes of the model's unguided prior distribution: among 100 spatial position prompts, more than 30% yield a success rate below 10% even after 32 independent noise-scaling attempts. Global exploration rapidly hits a "scaling plateau" characterized by sharply diminishing returns. The root bottleneck lies in a granularity mismatch: global text conditioning cannot enforce localized, object-centric constraints, and once early diffusion trajectories commit to incorrect spatial layouts, subsequent global sampling rarely fixes the structural failure.

Drawing inspiration from training-free region-aware generation, this paper reframes test-time scaling through the lens of structured task decomposition. While jointly fulfilling multi-object spatial requirements in a single global step is notoriously brittle, generating an individual entity within a localized spatial bounding box is substantially more reliable and robust. By converting a complex compositional prompt into coordinated regional sub-tasks via chain-of-thought (CoT) scene planning, the search space of valid configurations is dramatically densified. Core idea: treat regional decomposition as a new orthogonal scaling dimension, develop an efficient training-free region-aware generation primitive (RAG) that injects localized guidance solely during early denoising to anchor spatial structure, and search over the joint region-prompt-noise configuration space (RAS) to achieve superior compositional alignment under practical inference budgets.

Method

Overall Architecture

The RAS pipeline operates across two coupled levels: an inner training-free Region-Aware Generation (RAG) primitive and an outer Best-of-N Region-Aware Scaling (RAS) framework. Given a complex user prompt, the system first employs a large language model (e.g., GPT-4o) to conduct chain-of-thought (CoT) scene planning and recaptioning, yielding an enriched global description, spatial bounding-box masks for \(R\) distinct semantic regions, and localized sub-prompts capturing object identities and attributes. During inner diffusion sampling, local latents are cropped from the global latent and evolve under their dedicated region prompts in parallel with the global latent for only the first \(T_{\mathrm{rag}}\) denoising steps (typically the initial 10% of steps). After weighted normalization and blending back into the global latent, the process reverts to standard single-trajectory global denoising to refine textures and eliminate boundary seams. The outer RAS loop samples multiple independent configurations across noise seeds, prompt variants, and regional plans, selecting the best candidate via a dedicated verifier.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Compositional Prompt p"] --> B["CoT Scene Decomposition & Region Planning<br/>Generate global recaption, bounding boxes & sub-prompts"]
    B --> C["Early Dual-Path Denoising (t < T_rag)<br/>Parallel evolution of global and regional latents"]
    C --> D["Normalized Weighted Region Fusion<br/>Localized attribute binding and spatial alignment"]
    D --> E["Late Single-Path Global Denoising (t ≥ T_rag)<br/>Artifact-free texture refinement and global coherence"]
    E --> F["Best-of-N Joint Configuration Scaling<br/>Verifier scoring and final candidate selection"]

Key Designs

1. Chain-of-Thought Scene Decomposition and Recaptioning: Densifying the Solution Space

Text-to-image backbones often fail to bind attributes correctly when presented with monolithic, unstructured prompts containing multiple entities. RAS deploys an LLM planner to parse input prompt \(p\) into \(R\) distinct semantic regions \(\{r_i\}_{i=1}^R\). Each region is allocated a bounding-box mask \(M_i \in [0, 1]^{H \times W}\) along with a specialized sub-prompt \(p_i'\). Concurrently, an enriched global prompt \(\tilde{p}\) is generated to preserve scene-level atmosphere and background harmony. By explicitly transforming a combinatorial relational problem into independent, spatially grounded sub-tasks, the distribution of viable generation trajectories is densified, circumventing the combinatorial failure modes that plague blind global noise sampling.

2. Early-Only Regional Latent Evolution and Weighted Blending: Balancing Local Control and Global Harmony

Existing layout-guided diffusion methods typically enforce regional constraints across the entire sampling trajectory, which incurs high computational overhead that scales linearly with region count and often introduces unnatural seam boundaries. RAS addresses this by restricting regional intervention strictly to the early denoising stage (\(t < T_{\mathrm{rag}}\), e.g., the first 3 of 30 steps on FLUX.1-dev). For each region, a local latent is initialized by element-wise spatial cropping from the global latent \(\mathbf{z}_0^{(i)} = M_i \odot \mathbf{z}_0^{(g)}\) and refined concurrently with the global trajectory:

\[\mathbf{z}_t^{(i)} = f_\theta(\mathbf{z}_{t-1}^{(i)}, \phi(p_i'), t), \quad \mathbf{z}_t^{(g)} = f_\theta(\mathbf{z}_{t-1}^{(g)}, \phi(p), t)\]

To resolve spatial overlaps and ensure numerical stability, masks are normalized via \(\tilde{M}_i = \frac{M_i}{\sum_j M_j + \varepsilon}\). The regional updates are then blended into the global latent using scheduled ratio \(\alpha_t\):

\[\mathbf{z}_t = (1 - \alpha_t) \mathbf{z}_t^{(g)} + \alpha_t \sum_{i=1}^R \tilde{M}_i \odot \mathbf{z}_t^{(i)}\]

Once \(t \ge T_{\mathrm{rag}}\), the regional branches are terminated, and the diffusion model proceeds with conventional global denoising \(\mathbf{z}_t = f_\theta(\mathbf{z}_{t-1}, \phi(p), t)\). This early-only injection establishes the coarse spatial layout and locks object-attribute bindings while allowing subsequent global self-attention to seamlessly smooth boundary artifacts at a negligible ~10% constant computational overhead.

3. Region Axis Scaling Formulation: Expanding Beyond Classical Best-of-N Dimensions

Standard test-time scaling restricts candidate diversity to stochastic noise seeds \(\mathbf{z}_0\) or rewritten global prompts \(\tilde{p}\). RAS generalizes the candidate definition into a holistic configuration tuple \(\xi = (\mathbf{z}_0, \tilde{p}, \{(M_i, p_i')\}_{i=1}^R, T_{\mathrm{rag}})\), explicitly establishing the "Region Axis" (layouts and regional prompts) as an independent search dimension. Given sample budget \(N\), candidates are drawn from a distribution \(q(\xi)\) spanning noise variations, prompt perturbations, and layout adjustments. A verifier \(V(I_k, p)\) evaluates candidates to select the optimal generation:

\[I^\star = \arg\max_{k \in \{1,\dots,N\}} V(I_k, p)\]

Systematic exploration indicates that distributing inference budgets across both stochastic noise and structured regional decomposition significantly outperforms pouring all compute into any single axis.

Loss & Training

The framework is entirely training-free. No parameters of the diffusion backbone, text encoder, or planning LLM are modified. All compositional enhancements emerge purely at inference time through structured CoT region planning, early localized latent updates, and verifier-guided candidate selection.

Key Experimental Results

Main Results

On the comprehensive GenEval benchmark, RAS is evaluated across the FLUX.1-dev and SANA-1.0-1.6B backbones, comparing against unscaled baselines, prompt rewriting techniques (RePrompt, RePrompt+GRPO), and prominent test-time scaling methods (Noise Scaling, Prompt Scaling, Reflect-DiT, ReflectionFlow).

Table 1: Quantitative comparisons on the GenEval benchmark across sample budgets

Method / Backbone	Samples (N)	Overall ↑	Single ↑	Two ↑	Counting ↑	Colors ↑	Position ↑	Attribution ↑
FLUX.1-dev (Vanilla)	1	0.67	0.99	0.81	0.75	0.80	0.21	0.48
+ Reprompt	1	0.72	0.99	0.79	0.68	0.84	0.51	0.52
+ Reprompt+GRPO	1	0.76	0.99	0.87	0.77	0.85	0.62	0.49
+ RAS (Ours)	1	0.77	0.99	0.84	0.76	0.84	0.67	0.54
+ Noise Scaling	32	0.85	1.00	0.96	0.91	0.91	0.52	0.78
+ RAS (Ours)	2	0.85	0.99	0.94	0.84	0.88	0.77	0.66
+ Prompt Scaling	32	0.87	0.99	0.94	0.85	0.91	0.80	0.71
+ RAS (Ours)	4	0.88	0.99	0.94	0.88	0.88	0.88	0.70
ReflectionFlow	32	0.91	1.00	0.98	0.90	0.96	0.93	0.72
+ RAS (Ours)	8	0.91	1.00	0.98	0.93	0.93	0.88	0.75
+ RAS (Ours)	32	0.92	1.00	0.99	0.92	0.93	0.93	0.76

Ablation Study

Table 2: Scaling configuration ablation under a fixed total budget of 8 on the Position task (Table 3)

#Noise	#Global	#Region	Position Accuracy (%)	Note
8	1	-	82%	Pure noise scaling baseline; weak spatial control
8	1	1	89%	Adding a single region layout yields +7% jump without region resampling
1	8	8	90%	Pure prompt & region scaling without stochastic noise exploration
4	2	2	91%	Balanced split with modest noise exploration
2	4	4	92%	Optimal configuration: balanced noise diversity and region-prompt search

Table 3: Planner robustness and component ablation on GenEval Position (Table 4)

Configuration	Accuracy (Acc. ↑)	Note
FLUX vanilla	0.21	Baseline without guidance; fails on relative positions
+ LLM Rewrite	0.40	Global text enrichment provides modest gains
+ LLM Plan + RAS (N=1)	0.60	Regional planning and early injection boost accuracy by +20%
+ Best-of-2 (RAS N=2)	0.77	Modest 2-sample scaling delivers dramatic reliability gains

Table 4: Fine-grained compositional evaluation on T2I-CompBench (Table 5)

Model	Color ↑	Shape ↑	Texture ↑	Spatial ↑	Non-Spatial ↑	Complex ↑	HPSv3 ↑
Vanilla FLUX	0.8821	0.7304	0.8662	0.5442	0.8823	0.8308	9.22
RPG (ICML 2024)	0.8902	0.7534	0.8952	0.7042	0.9025	0.8358	9.08
RAS (Ours)	0.8921	0.7790	0.9107	0.7310	0.9133	0.8425	9.20

Key Findings

• 16× Reduction in Sampling Budget: On FLUX.1-dev, RAS achieves an overall GenEval score of 0.85 with merely 2 samples, matching the performance of a 32-sample noise-scaling baseline. On an NVIDIA H100 GPU, this translates to a 7.5× end-to-end wall-clock speedup (50.8s vs. 380.8s).
• Overcoming the Spatial Relation Barrier: Baseline FLUX scores a dismal 0.21 on GenEval Position, improving to only 0.52 under 32-sample noise scaling. In stark contrast, RAS attains 0.77 with \(N=2\) and 0.88 with \(N=4\), proving that spatial alignment requires geometric task decomposition rather than brute-force noise perturbation.
• Synergy Across Scaling Dimensions: Ablating the budget allocation under \(N=8\) demonstrates that neither pure noise exploration (82%) nor pure prompt/region exploration (90%) is optimal. A balanced allocation (2 noise × 4 global × 4 region) achieves the highest accuracy (92%), confirming the complementarity of the noise and region axes.
• Uncompromised Visual Aesthetics: While prior region-guided methods like RPG degrade overall aesthetic quality due to persistent bounding-box constraints throughout all steps (HPSv3 drops from 9.22 to 9.08), RAS confines regional guidance to the first 10% of denoising steps. It achieves a top-tier spatial score (0.7310 vs. 0.7042) while preserving native aesthetic quality at 9.20.

Highlights & Insights

• Elevating Spatial Regions to a Core Test-Time Scaling Axis: Rather than viewing test-time scaling merely as random seed resampling borrowed from LLM reasoning, this work recognizes spatial decomposition as an intrinsic geometric dimension for visual generative models.
• Early-Only Regional Injection Primitive: By leveraging the empirical property that diffusion models crystallize structural layout early on and refine high-frequency details later, early injection avoids boundary stitching artifacts while keeping inference overhead at a constant ~10%.
• Seamless Training-Free Plug-and-Play Integration: RAS functions entirely at inference time without requiring retraining or LoRA adapters, offering plug-and-play compatibility across diverse backbones (FLUX, SANA) and verifiers.

Limitations & Future Work

• Vulnerability to LLM Planning Errors: When dealing with cluttered scenes containing numerous densely packed objects or ambiguous relational prepositions, the planning LLM can generate flawed bounding boxes, which directly misguide the regional latents; geometric rule constraints are currently needed as safeguards.
• Peak Compute in Parallel Exploration: While total sample count drops by an order of magnitude, concurrent parallel generation of multiple candidates still demands considerable peak VRAM, motivating future exploration of adaptive sequential reflection loops.

Related Work & Insights

• vs Noise Scaling / Prompt Scaling: Conventional noise and prompt scaling explore stochastic trajectories or lexical rewrites globally, failing to anchor localized objects. RAS introduces explicit spatial bounds, converting rare combinatorial configurations into accessible single-object sub-tasks.
• vs RPG: RPG enforces regional attention guidance across all denoising steps, incurring runtime overhead proportional to region count and risking unnatural boundary seams. RAS injects guidance only during early steps, achieving constant-time scalability and superior global coherence.
• vs Reflection-Tuned Models (Reflect-DiT, ReflectionFlow): Reflection tuning requires expensive post-training alignment and fine-tuning with multimodal critics. In contrast, RAS delivers state-of-the-art compositional alignment in a completely training-free manner.

Rating

• Novelty: ⭐⭐⭐⭐⭐ Formulates the first systematic test-time scaling framework that establishes spatial region decomposition as a dedicated, actionable search axis.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Extensive evaluations across GenEval, T2I-CompBench, and HPSv3 metrics, supported by end-to-end runtime breakdowns and multi-axis budget ablations.
• Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally clear narrative, structured mathematical formulation, and insightful visualizations.
• Value: ⭐⭐⭐⭐⭐ Provides a highly practical, training-free solution for compositional text-to-image synthesis, setting a strong precedent for inference-time compute allocation in diffusion models.

Related Papers

• [ECCV 2026] From Open Loop to Closed Loop: A Test-Time Iterative Optimization Framework for Reference-Consistent Image Generation
• [CVPR 2026] Progress by Pieces: Test-Time Scaling for Autoregressive Image Generation
• [ECCV 2026] Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models
• [CVPR 2026] From Scale to Speed: Adaptive Test-Time Scaling for Image Editing
• [CVPR 2026] Compositional Text-to-Image Generation Via Region-aware Bimodal Direct Preference Optimization