Research map / Image generation / Autoregressive and latent generation

FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Learning-based image quality assessment · 0 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.

Proposes FoMo, an annotation-free pipeline that labels reference-variant image pairs with the diffusion forking timestep as a pointwise perceptual distance and trains reference-based IQA metrics with a RankNet-style ranking loss.

Problem

Reference-based IQA metrics rely on expensive human annotations (MOS or 2AFC), and pairwise labels do not directly support global ranking across image pairs.

Approach

Use the forking moment of a diffusion (FLUX) denoising trajectory, where a reference is noised to a sampled timestep and re-denoised, as a pointwise distance label (earlier fork = larger distance), and supervise with a batch-wise RankNet-style BCE loss over all pairs.

Key results

• FoMo with RankBCE improves mean PIPAL SROCC to 0.683 vs 0.484 for the best human-annotated baseline (PieAPP) across seven backbones.
• On LIVE, FoMo reaches mean SROCC 0.926 vs 0.870 for KADID-10K (DMOS regression).
• Human study: label agrees with 90.2% of individual 2AFC responses and 92.8% of majority-vote items; single-reference Spearman 0.970 vs inter-rater 0.960.

Surprising findings

• Training on only 10k generated pairs from FLUX already beats the best human-annotated dataset for the CNN backbone on PIPAL.
• On TID2013 and CSIQ, the CNN-based FoMo models perform comparably to or below human-annotated baselines (e.g., LPIPS-VGG 0.663 on TID2013 vs 0.757 KADID-10K).

Limitations

• CNN-based FoMo models do not consistently beat human-annotated baselines on TID2013 and CSIQ.
• Performance of Transformer backbones depends on batch size and the choice of generator (FLUX strongest).

Experimental setup

Datasets: ImageNet PIPAL TID2013 CSIQ LIVE KADID-10K

Benchmarks: PIPAL TID2013 CSIQ LIVE

Metrics: SROCC KROCC PLCC

Baselines: BAPPS PieAPP NIGHTS KADID-10K

Builds on: RankNet LPIPS DISTS DreamSim FLUX.1-dev

Closest papers by meaning

• Multidimensional Observer Model and Perceptual Dimensions of Human Image Quality Assessment (NeurIPS 2026 · similarity 0.87)
• LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations (ICML 2025 · similarity 0.87)
• Evaluation of Visual Processing Should Be Human-Centered, Not Metric-Centered (NeurIPS 2026 · similarity 0.87)
• The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric (NeurIPS 2026 · similarity 0.86)
• Rethinking FID: Towards a Better Evaluation Metric for Image Generation (CVPR 2024 · similarity 0.85)
• PerceptionRubrics: Calibrating Multimodal Evaluation to Human Perception (ICML 2026 · similarity 0.85)