Research map / Image generation / Text-to-image diffusion
DynEval: Holistic Evaluations of T2I Generative Models in the Wild
Shyam Marjit, Dheeraj Baiju, Anuj Shikarkhane, Akhil Sakthieswaran, Sayak Paul, Anirban Chakraborty
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| scale the evaluator from 2B to 4B parameters (ablation) | DynEval · Table 2 benchmarks | SRCC: average relative SRCC gain of 4.62% | Improved |
| scale the evaluator from 2B to 4B parameters (ablation) | DynEval · EvalMi | SRCC: 0.7346 -> 0.7944 (+8.14%) | Improved |
| scale the evaluator from 2B to 4B parameters (ablation) | DynEval · T2I-Eval-Bench | SRCC: 0.4535 -> 0.4857 (+7.10%) | Improved |
| scale the evaluator from 2B to 4B parameters (ablation) | DynEval · EvalMuse | SRCC: 0.7765 -> 0.7932 (+2.15%) | Improved |
| use DynEval-4B instead of the previous best evaluators | T2I evaluator · all 11 benchmarks | SRCC: +4.77% relative on average, +7.61% where SOTA | Improved |
| use DynEval-4B instead of the best baseline | T2I evaluator · UniGenBench++ | SRCC: 0.4102 -> 0.4958 (+20.87%) | Improved |
| use DynEval-4B instead of the best baseline | T2I evaluator · TIIF-Bench | SRCC: 0.5495 -> 0.5980 (+8.83%) | Improved |
| use DynEval-4B instead of the best baseline | T2I evaluator · EvalMi | SRCC: 0.7245 -> 0.7944 (+9.65%) | Improved |
| use DynEval-4B instead of the best baseline | T2I evaluator · GenAI-Bench | SRCC: 0.5637 -> 0.5945 (+5.46%) | Improved |
| use DynEval-4B (4B) instead of the 8B T2I-CoReBench evaluator | T2I evaluator · RichHF | SRCC: 0.6311 -> 0.5457 (trails by 9.41%) | Hurt |
| use DynEval-4B instead of the 8B GenEval2 evaluator | T2I evaluator · GenEval | SRCC: 0.6044 -> 0.5894 (trails by 2.48%) | Hurt |
| increase the number of objects in the image from 1-5 to 16-20 (ablation) | DynEval-4B IQA question generation · IQA question quality | Coverage Ratio: 0.885 -> 0.696 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GenEval 2 (8B) — GenEval: An object-focused framework for evaluating text-to-image alignment (NeurIPS 2023) · better on 4 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DynT2I-Eval: A Dynamic Evaluation Framework for Text-to-Image Models (NeurIPS 2026 · similarity 0.92)
- Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating (ICLR 2025 · similarity 0.92)
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.91)
- Who Evaluates the Evaluations? Objectively Scoring Text-to-Image Prompt Coherence Metrics with T2IScoreScore (TS2) (NeurIPS 2024 · similarity 0.90)
- FineGRAIN: Evaluating Failure Modes of Text-to-Image Models with Vision Language Model Judges (NeurIPS 2025 · similarity 0.90)
- PromptCoT: Align Prompt Distribution via Adapted Chain-of-Thought (CVPR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
