COMPOSITIONAL TEXT-TO-IMAGE GENERATION

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

Accepted at NeurIPS 2026

Sunung Mun1Sunghyun Cho1,2Jungseul Ok1,2

1 Graduate School of Artificial Intelligence, POSTECH

2 Department of Computer Science and Engineering, POSTECH

Read the paper Code

Parse once. Verify each candidate. Refine what fails.

COUNTS & MATERIALSSix rabbits on top of a wooden suitcase.
ATTRIBUTE BINDINGFive red monkeys, and five glass donuts to the left of a metal dog.
SPATIAL RELATIONSA yellow car to the left of a purple bicycle in front of five checkered chairs.

Selected EPIC outputs. All images are actual results from our experiments.

THE IDEA

Keep the requirements fixed.
Let the image improve.

Compositional prompts ask an image to satisfy several requirements at once: the right objects, the right number of instances, the right attributes, and the right relationships.

EPIC makes those requirements explicit. It compiles the original prompt into a fixed visual program of typed predicates. Each candidate is checked against that same program using visual evidence. The remaining failures guide either targeted editing or resampling, turning verification into a control signal for refinement.

The framework uses pretrained components and requires no controller or generator training.

01

Parse

Compile the original prompt once into objects and checkable visual predicates.

02

Verify

Recompute predicate states from the visual evidence in every candidate.

03

Refine

Route local failures to editing and broader failures to fresh generation.

EPIC framework: a prompt is parsed into a fixed visual program; a predicate verifier and refinement policy guide generation or editing until acceptance or budget exhaustion.
The EPIC refinement loop. Candidates passing the verification gate are accepted; at budget exhaustion, the best partial candidate is returned.

QUANTITATIVE RESULTS

Stronger composition.
More efficient refinement.

On GenEval2, EPIC improves alignment while using fewer image-model executions and MLLM resources than RAISE.

71.46%

Prompt-level accuracy

34.16 for single-pass generation

31%

Fewer image-model executions

13.6 vs. 19.8 per prompt, compared with RAISE

81%

Fewer MLLM tokens

9.2k vs. 49.4k per prompt, compared with RAISE

Prompt-level accuracy

Prompt-level accuracy (%). Higher is better.

GenEval2 results with FLUX.2-Klein-9B
MethodGM ↑AM ↑Image calls ↓MLLM calls ↓MLLM tokens (k) ↓
FLUX.2-Klein-9B34.1679.061.00——
BoN+NVILA41.6082.3832.0——
T2I-Copilot36.5681.191.244.4815.2
RAISE52.2386.2819.89.1249.4
EPIC Ours71.4690.2513.62.539.2

800 prompts · Maximum budget of 32 image-model executions · Deterministic MLLM decoding (T = 0). Prompt-level and atom-level accuracy are estimated with Soft-TIFA GM and AM, respectively, on a 0–100 scale. Costs are realized averages per prompt; “—” denotes an unreported value.

EPIC targets explicit compositional requirements. Performance and visual-quality trade-offs vary across benchmarks; see the paper for additional evaluations and limitations.

QUALITATIVE COMPARISONS

Better compositional alignment with EPIC.

1 / 10
PROMPT

“five red monkeys, and five glass donuts to the left of a metal dog”

EPIC (Ours)
FLUX.2-Klein-9B
BoN+NVILA
T2I-Copilot
RAISE

Selected comparisons using the same prompt and FLUX.2-Klein-9B backbone.

CITE THIS WORK

BibTeX

If you find this work useful, please consider citing our paper.

@inproceedings{mun2026epic,
  title     = {{EPIC}: Efficient Predicate-Guided Inference-Time Control
               for Compositional Text-to-Image Generation},
  author    = {Mun, Sunung and Cho, Sunghyun and Ok, Jungseul},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  url       = {https://arxiv.org/abs/2605.11722}
}

IMAGE COMPARISON

1 / 5← → Compare methods   ·   Esc to close