Multi-Survey Photo-z

Model

A ViT-Base vision transformer with 87 million parameters. It learns from images first, without labels, and is then tuned for redshift.

1 · Any filters in

Each filter image is cut into 4 × 4 pixel patches. The weights that turn a patch into a token are generated from the filter's wavelength (0.3–6 µm) and pixel scale, so the same model takes 4 Euclid bands, 5 SDSS bands or 42 COSMOS bands. Any band can be missing; in training 20% are dropped at random.

2 · A distance out

Tokens pass through 12 transformer blocks, with a few extra register tokens that give the model scratch space. The head gives a probability for each small redshift bin, so every galaxy gets a full distribution, not one number. At test time the galaxy is rotated and flipped 8 ways and the answers are averaged. Catalogue fluxes can be added as extra tokens.

3 · Learning without labels

Pretraining sees only images. Two altered views of one galaxy go to a student and a teacher, and the student learns to give the same answer (DINO). It also fills in hidden pixels, and a Gram anchor keeps patch features from drifting. The teacher is a slow moving average of the student. Distances are never shown here.

4 · Speed

Time to get one redshift, log scale. Filled: this model, measured on the 20,449 COSMOS-Web test galaxies. Open: other methods, approximate figures from their papers on the hardware stated there (hover for it), so compare orders of magnitude, not exact ratios. Training is a one-off cost and is not included.
MethodPer galaxy595k COSMOS-Web galaxiesHardware
This work, images, 8 rotations2.3 ms23 minone H100 GPU
This work, images, 1 pass0.3 ms3 minone H100 GPU
NNPZ (Euclid DR1 PHZ)~6 ms~1 h24-core Xeon workstation
EAZY~40 ms~6 h2008 laptop CPU
LePhare~1 ms – 1 s10 min – 1 weekone CPU core, depends on the template grid
Phosphoros (Euclid Q1 PHZ)~0.8 s~5.5 days24-core Xeon workstation
Bayesian SED fit (Prospector-type)10–25 CPU-h~10 million CPU-hCPU

Sources: NNPZ and Phosphoros from Euclid Collaboration (2024, arXiv:2407.07940), on a 24-core Xeon Silver 4116. EAZY from Brammer et al. (2008, arXiv:0807.1533): 6,300 galaxies in about 4 minutes. LePhare has no standard benchmark: the Rubin LePhare plan targets about 1 ms per object on a coarse grid, and comparisons often assume about 1 s for a full fit. Bayesian SED fits from Leja et al. and Tacchella et al. (2022), as summarised in arXiv:2506.19919. We found no published runtimes for DEmP or Mizuki. Our 595k times scale the measured 20,449-galaxy runs linearly.