FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance

A distance label from the generator itself

Perceptual metrics normally learn from people, and human labels are expensive, noisy and hard to compare across images. A diffusion model already builds images coarse-to-fine, so we use that order. Re-enter the trajectory at step s and generate again. Fork late and only fine detail changes. Fork early and only coarse structure survives. The forking step is the distance label.

We generated 480,000 labelled pairs with FLUX and no annotators. Because every pair carries its own label, any two pairs in a batch can be compared, so the model learns a global ranking and not only a choice within one reference.

From forks to a trained metric

Left: trajectories fork from noise, and a later fork gives a more similar image. Right: 2AFC training compares two variants of one reference, while RankBCE compares every pair in the batch.

People agree, in two studies

Study 1: same reference

Raters ordered five forks of one image by similarity to it.

Reference (outlined), then forks from least to most similar. Example images.

0.970

rank correlation with the fork order, 1,982 annotations

Study 2: across references

Raters chose which of two pairs, from different references, looked more alike.

90.2%

of 5,713 annotations agree

92.8%

of items, by majority vote

Fork label vs. personPerson vs. person

1-1011-2021-3031-4041-50

Agreement (%) by gap between the two labels, in steps. The label is weakest exactly where people disagree with each other.

Try a real Study 2 item

Two image pairs. Which pair shows two images more similar to each other?

Which pair looks more alike to you?

RankBCE: rank every pair

A 2AFC label only says which of two variants of one reference is closer. Our labels are pointwise, so every pair can be compared with every other pair, even across references.

2AFC

B labels, same reference only

human

human voteno label

comparisons per batch

RankBCE

B² labels, any two pairs

s=45

s=5

s=25

s=45

0.5

s=5

0.5

s=25

0.5

row closercolumn closertie

4,096

comparisons per batch

Same FoMo labels, different loss

PIPAL SROCC. RankBCE is never beaten; on DINOv3 it matches L1 within seed variation.

LPIPS-Alex

RankBCE

0.733

2AFC

0.592

0.440

DINOv3

RankBCE

0.699

2AFC

0.622

0.694

Results

Each row trains the same seven models on a different dataset, so only the supervision changes. FoMo is the only row with no human labels. Scores are mean SROCC over 5 seeds, and higher is better.

7 / 7backbones lead on PIPAL

12 / 12frozen-ViT cells lead

18 / 28of all cells lead

Trained on	Objective	LPIPS-Alex	LPIPS-VGG	DISTS	DINOv3	CLIP	MAE	DreamSim^†	Avg.
BAPPS	2AFC	0.622	0.634	0.489	0.287	0.151	0.338	0.760	0.469
PieAPP	2AFC	0.602	0.635	0.564	0.325	0.258	0.294	0.711	0.484
NIGHTS	2AFC	0.577	0.545	−0.294	0.234	0.277	0.234	0.661	0.319
KADID-10K	L1	0.577	0.590	0.518	0.298	0.312	0.253	0.617	0.452
FoMo (ours)	RankBCE	0.733	0.683	0.615	0.699	0.644	0.632	0.776	0.683

Trained on	Objective	LPIPS-Alex	LPIPS-VGG	DISTS	DINOv3	CLIP	MAE	DreamSim^†	Avg.
BAPPS	2AFC	0.779	0.671	0.632	0.341	0.285	0.287	0.815	0.544
PieAPP	2AFC	0.761	0.722	0.680	0.315	0.302	0.467	0.768	0.574
NIGHTS	2AFC	0.779	0.653	−0.497	0.230	0.294	0.358	0.760	0.368
KADID-10K	L1	0.793	0.757	0.834	0.613	0.519	0.595	0.788	0.700
FoMo (ours)	RankBCE	0.785	0.663	0.691	0.713	0.736	0.643	0.801	0.719

Trained on	Objective	LPIPS-Alex	LPIPS-VGG	DISTS	DINOv3	CLIP	MAE	DreamSim^†	Avg.
BAPPS	2AFC	0.943	0.887	0.855	0.411	0.427	0.455	0.912	0.699
PieAPP	2AFC	0.936	0.896	0.821	0.331	0.514	0.587	0.904	0.713
NIGHTS	2AFC	0.945	0.862	−0.560	0.335	0.364	0.427	0.867	0.463
KADID-10K	L1	0.935	0.895	0.938	0.616	0.604	0.698	0.874	0.794
FoMo (ours)	RankBCE	0.938	0.859	0.918	0.811	0.901	0.795	0.894	0.874

Trained on	Objective	LPIPS-Alex	LPIPS-VGG	DISTS	DINOv3	CLIP	MAE	DreamSim^†	Avg.
BAPPS	2AFC	0.952	0.929	0.843	0.688	0.425	0.298	0.928	0.723
PieAPP	2AFC	0.934	0.921	0.860	0.458	0.415	0.537	0.908	0.719
NIGHTS	2AFC	0.949	0.919	−0.774	0.439	0.560	0.376	0.874	0.478
KADID-10K	L1	0.942	0.912	0.950	0.765	0.827	0.797	0.896	0.870
FoMo (ours)	RankBCE	0.948	0.923	0.954	0.897	0.925	0.907	0.931	0.926

Best in each column is highlighted. ^†Non-frozen backbone.

Use it

pip install fomo-iqa

import fomo

metric = fomo.FoMo('dinov3')
d = metric(img_ref, img_test)        # lower = more similar

# as a perceptual loss
loss = metric(prediction, target, output='loss').mean()
loss.backward()

ModelsLPIPS-Alex, LPIPS-VGG, DISTS, DINOv3, CLIP, MAE, DreamSim
Datasetfomo_480k: 480,000 samples, 512×512
Codegithub.com/JHLew/FoMo
Packagefomo-iqa on PyPI
NoteDINOv3 is gated: accept its license on Hugging Face before first use.

Citation and poster

@inproceedings{lew2026fomo,
  title     = {FoMo: Forking Moment in Generative
               Trajectory as a Perceptual Distance},
  author    = {Lew, Jaihyun and Jung, Mingi and
               Park, Minjun and Song, Wooseok and
               Yoon, Sungroh},
  booktitle = {Advances in Neural Information
               Processing Systems},
  year      = {2026}
}