EPIC: Efficient Predicate-Guided Inference-Time Control for Compositional Text-to-Image Generation

Sunung Mun, Sunghyun Cho, Jungseul OkarXiv 2026 · arXiv

TL;DR

EPIC (Efficient Predicate-Guided Inference-Time Control) is a training-free framework that drastically improves prompt-following in text-to-image models by treating refinement as a structured, predicate-guided search. By parsing a prompt once into a fixed "visual program" and using cheap, modular perception tools to verify individual conditions, EPIC intelligently routes local failures to targeted image editing and global failures to resampling. This approach more than doubles prompt-level accuracy on complex benchmarks while significantly reducing computational and MLLM token costs compared to prior agentic baselines.

What Problem Does This Paper Address?

State-of-the-art text-to-image (T2I) generators can create stunningly realistic images, but they frequently stumble when given complex, compositional prompts. If you ask for "four yellow candles on top of a spotted raccoon jumping over four stone koalas," a base model might easily miss an object, miscount the items, or swap attributes (e.g., making the koalas yellow instead of the candles).

Retraining these massive models to better follow prompts is incredibly expensive and impossible for closed-source systems. To fix this, researchers have turned to inference-time scaling—using extra computation during generation to refine the output. However, existing methods fall into two flawed camps:

  1. Selection/Search Scaling: Generates many candidates and ranks them using a single scalar alignment score. This score is too coarse; it tells the system that an image failed, but not why or where.
  2. Agentic Refinement: Uses Multimodal Large Language Models (MLLMs) to repeatedly critique and rewrite prompts in a natural-language loop. While flexible, this is computationally expensive and prone to "feedback drift," where the refinement state loses track of the original prompt over multiple rounds.
Comparison of inference-time scaling strategies for compositional text-to-image generation.

Methods

EPIC reformulates inference-time refinement as a deterministic control problem guided by a fixed visual program. The pipeline operates in three distinct phases:

  1. Visual-Program Construction: When a user inputs a prompt, an MLLM parses it exactly once into a fixed visual program $\Phi$ consisting of object variables and typed predicates (e.g., object presence, exact counts, attributes, and spatial relations). This program acts as an immutable semantic contract.
  2. Predicate Verification: Instead of relying on an expensive MLLM to critique the entire image at every round, EPIC uses a suite of lightweight, modular perception tools (like SAM3 for segmentation, SigLIP2 for region-text scoring, and MiDaS for depth estimation). Each predicate is evaluated locally and assigned a discrete state: satisfied, uncertain, or violated.
  3. Predicate-Driven Refinement Policy: If any predicates fail, they form a "blocking set." EPIC's dependency-aware policy uses this set to choose the next action. Localized errors (like an incorrect attribute or a minor counting mistake) are routed to a targeted Image Editor. Structural or layout-level failures (like a completely missing object category) trigger Resampling from a precomputed pool of meaning-preserving prompt rewrites.
Figure 2: Method overview. EPIC compiles the input prompt once into a fixed visual program Φ of typed predicates and prepares meaning-preserving prompt rewrites for candidate generation. Starting from an initial candidate generated from a rewrite, each round verifies the current candidate, forms a blocking set from non-satisfied predicates, and uses this state to choose acceptance, resampling from a rewrite, or targeted editing. If the budget is exhausted before acceptance, EPIC returns the best partial candidate from the verified history as a fallback.

Datasets

The authors evaluated EPIC primarily on GenEval2, a rigorous benchmark specifically designed to measure compositional prompt alignment across varying levels of "atomicity" (the number of primitive visual elements in a prompt). To demonstrate broad semantic improvement, they also tested the framework on DrawBench and the original GenEval benchmark.

Key Insights and Ideas

  • Decomposition as a Stable Control State: A major pitfall of prior agentic methods is that their natural-language checklists evolve and drift across refinement rounds. EPIC's core "aha moment" is that decomposing the prompt once into a fixed program anchors the entire refinement process, ensuring the system never loses sight of the original request.
  • Modular Perception over Holistic Critique: Asking a large MLLM "Is this image correct?" is computationally heavy and often inaccurate for fine-grained details. Breaking the verification down into local, specialized vision tasks (bounding boxes, depth maps, mask-level attributes) yields a much cleaner and cheaper feedback signal.
  • Smart Failure Routing: Not all prompt failures are created equal. By separating errors into local attribute mismatches (best fixed by an editor) and global layout failures (best fixed by regenerating with a rewritten prompt), EPIC allocates its compute budget far more efficiently than brute-force best-of-N sampling.

Key Results

Using FLUX.2-Klein-9B as the base generator and Qwen3-VL-32B as the MLLM backend, EPIC achieved outstanding results on GenEval2:

  • Accuracy Boost: EPIC raised prompt-level accuracy from 34.16% (single-pass base model) to 71.46%.
  • Outperforming Baselines: Under the same maximum execution budget, EPIC outperformed RAISE (the strongest prior refinement baseline) by 19.23 percentage points.
  • Massive Cost Reductions: Despite achieving higher accuracy, EPIC reduced realized costs compared to RAISE by 31% in image-model executions, 72% in MLLM calls, and 81% in MLLM tokens per prompt.
Figure 3: Main GenEval2 results with deterministic MLLM decoding (T = 0). Left: prompt-level accuracy by atomicity. Right: accuracy versus realized image-model executions under budget sweeps. All scores use the GenEval2 Soft-TIFA-GM evaluator.

Limitations and Future Work

While highly effective, EPIC is fundamentally bound by the accuracy of its initial parsed visual program and its underlying perception verifiers. If the parser generates a semantically flawed program, the system will spend its budget optimizing for the wrong target. Furthermore, false negatives from the verification tools can trigger unnecessary edits, while false positives can cause premature acceptance.

Currently, EPIC is optimized for explicit, checkable visual properties (counts, colors, spatial layouts). The authors note that extending the framework to handle abstract reasoning, open-ended prompts, and interactive human-in-the-loop settings remain important avenues for future research.

Summary updated

Abstract

Recent text-to-image (T2I) generators can synthesize realistic images, but still struggle with compositional prompts involving multiple objects, counts, attributes, and relations. We introduce EPIC (Efficient Predicate-Guided Inference-Time Control), a training-free inference-time refinement framework for compositional T2I generation. EPIC casts refinement as predicate-guided search: it parses the original prompt once into a fixed visual program of object variables and typed predicates, covering checkable conditions such as object presence, counts, attributes, and relations. Each generated or edited image is verified against this program using visual evidence extracted from that image. An image is judged to satisfy the prompt only when all predicates are satisfied; otherwise, failed predicates decide the next step, routing local failures to targeted editing and global failures to resampling while the fixed visual program remains unchanged. On GenEval2, EPIC improves prompt-level accuracy from 34.16% for single-pass generation with the base generator to 71.46%. Under the same generator/editor setting and maximum image-model execution budget, EPIC outperforms the strongest prior refinement baseline by 19.23 points while reducing realized cost by 31% in image-model executions, 72% in MLLM calls, and 81% in MLLM tokens per prompt.
BibTeX
@misc{mun2026epic,
  author = {Sunung Mun and Sunghyun Cho and Jungseul Ok},
  title = {{EPIC: Efficient Predicate-Guided Inference-Time Control for Compositional Text-to-Image Generation}},
  year = {2026},
  eprint = {2605.11722},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.11722},
}

Loading related content...

Figure