Research map / Image generation / Autoregressive and latent generation
Hallucination Early Detection in Diffusion Models
Federico Betti, Lorenzo Baraldi, Lorenzo Baraldi, Rita Cucchiara, Nicu Sebe
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 |
|---|---|---|---|
| add HEaD+ to SD 1.4 | SD 1.4 · COCO multi-object | MG3: 50.74 -> 65.03 | Improved |
| add HEaD+ to TokenCompose | TokenCompose · COCO multi-object | MG3-MG5: 76.16 -> 81.27 MG3, 28.81 -> 35.33 MG4, 3.28 -> 4.93 MG5 | Improved |
| add HEaD+ to TokenCompose | TokenCompose · COCO multi-object | MG2: 98.08 -> 97.61 | No clear effect |
| add HEaD+ to PixArt-alpha | PixArt-alpha · COCO multi-object | MG2-MG5: 83.19 -> 89.19 MG3, 42.80 -> 51.65 MG4 | Improved |
| use earlier detection timestep T=5 instead of T=25 (ablation) | HEaD+ · InsideGen time saving | time saved: 14.18 -> 37.73 | Improved |
| use later detection timestep T=25 instead of T=5 (ablation) | HEaD+ · InsideGen | TN-Rate: 50.16 -> 76.95 | Improved |
| use HEaD+ architecture instead of HEaD | HEaD · InsideGen time saving (T=25) | time saved: -11.70 -> 14.18 | Improved |
| use HEaD+ instead of HEaD at T=8 | HEaD · InsideGen time saving (T=8) | time saved: 11.55 -> 37.73 | Improved |
| add textual token v_o as input (ablation) | HEaD+ HP network · InsideGen (T=25) | TN-Rate: 52.16 -> 76.95 | Improved |
| add PFI as input (ablation) | HEaD+ HP network · InsideGen (T=25) | TN-Rate: 62.35 -> 76.95 | Improved |
| add cross-attention maps as input (ablation) | HEaD+ HP network · InsideGen (T=25) | TN-Rate / Recall: 74.84 -> 76.95 TN-Rate, 91.21 -> 93.40 Recall | Mixed |
| add localization module to HEaD+ (TokenCompose) | TokenCompose w/ HEaD+ · two-object relation prompts | Relation Consistency: 50.0 -> 67.5 | Improved |
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.
- Token Compose — TokenCompose: Text-to-Image Diffusion with Token-level Supervision (CVPR 2024) · better on 5 of 5 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
- Understanding Hallucinations in Diffusion Models through Mode Interpolation (NeurIPS 2024 · similarity 0.93)
- GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs (ICLR 2026 · similarity 0.90)
- Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.90)
- Towards Understanding Text Hallucination of Diffusion Models via Local Generation Bias (ICLR 2025 · similarity 0.90)
- RODS: Robust Optimization Inspired Diffusion Sampling for Detecting and Reducing Hallucination in Generative Models (NeurIPS 2025 · similarity 0.90)
- Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination (CVPR 2026 · similarity 0.90)
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.
