Research map / Image generation / Text-to-image diffusion

Science-T2I: Addressing Scientific Illusions in Image Synthesis

CVPR 2025 · Image generation › Text-to-image diffusion · approach: T2I evaluation and controllable adapters · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Science-T2I, an expert-annotated adversarial dataset of scientific-knowledge image pairs, the SciScore CLIP-based reward model trained on it, and a two-stage SFT plus masked online DPO fine-tuning of FLUX.1[dev] to improve scientific realism of generated images.

Problem

Text-to-image models often generate superficial images that ignore underlying scientific principles implied by prompts (e.g., "an unripe apple" should look green), and existing evaluators like CLIP and VLMs fail to judge such scientific realism.

Approach

Build an expert-curated preference dataset of implicit, explicit, and superficial prompts with matching images, fine-tune CLIP into a reward model (SciScore) with implicit-prompt alignment and image-encoder losses, and use it as the reward for masked DPO-based online fine-tuning of FLUX after SFT.

Key results

• SciScore reaches 93.14% average accuracy on Science-T2I S and 91.19% on Science-T2I C, above human evaluation (87.01% and 86.02%).
• GPT-4o mini with CoT reaches 74.97% on Science-T2I S, while CLIP-H and BLIPScore are near random (about 50-60%).
• SFT+OFT raises FLUX.1[dev] SciScore from 23.56 to 31.18 on Science-T2I S (82% relative improvement over explicit-prompt upper bound) and from 27.26 to 32.15 on Science-T2I C.

Measured effects reported in the paper

Surprising findings

• OFT without a preceding SFT stage does not improve SciScore, suggesting the model cannot learn from two suboptimal samples without domain transfer.
• Chain-of-thought prompting gives little improvement for GPT-4o mini on the scientific two-choice task.

Limitations

• Evaluation set sizes are small (671 and 227 tuples), with accuracy differences of a few points.
• Reliance on GroundingDINO subject masks, which may fail for prompts without clear subjects.
• Dataset covers only 16 scientific phenomena across physics, chemistry, and biology.

Experimental setup

Datasets: Science-T2I

Benchmarks: Science-T2I S Science-T2I C

Metrics: accuracy in two-choice selection SciScore normalized difference

Baselines: CLIP-H BLIPScore GPT-4o mini GPT-4o mini + CoT FLUX.1[dev] Stable Diffusion v1.5 Stable Diffusion XL Stable Diffusion 3 FLUX.1[schnell]

Builds on: CLIP DDPO DPO FLUX.1[dev] GroundingDINO

Closest papers by meaning

• ScImage: How good are multimodal large language models at scientific text-to-image generation? (ICLR 2025 · similarity 0.85)
• Instruct Where the Model Fails: Generative Data Augmentation via Guided Self-contrastive Fine-tuning (AAAI 2025 · similarity 0.84)
• Learning Visual Generative Priors without Text (CVPR 2025 · similarity 0.84)
• IDEA-Bench: How Far are Generative Models from Professional Designing? (CVPR 2025 · similarity 0.83)
• The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.83)
• MMMG: A Massive, Multidisciplinary, Multi-Tier Generation Benchmark for Text-to-Image Reasoning (NeurIPS 2025 · similarity 0.83)