Master IASD · DataLab 2026

MNIST generation

A2 · FID, precision, recall and generated images

Updated 21 October 2026 · 18:12 CESTDaily run around 05:00 · Paris time

Podium

Ranked by FID · lower is better. Student groups and professor references compete together.

1
Professor reference

BestOf 2025 · 2

7.28 FID

2
Professor reference

BestOf 2025 · 1

8.27 FID

3
Professor reference

VanillaGAN

49.88 FID

Daily progress

FID over time · lower is better.

Drag to zoom horizontally or vertically; scroll to zoom. Use the toolbar to pan or reset. Dots show evaluations, not GitHub activity. Failed runs leave gaps.

Results

FID ↓ · precision ↑ · recall ↑. P and R are fractions between 0 and 1. The generator architecture must stay fixed.

Hover or focus the ! icon to read an error; click to copy it. Open the image link to inspect the generated digits.

RankStatusSamplesSource
1
Professor reference
BestOf 2025 · 2Reference
7.280.8660.701135.69A40Success 10 imagesFrozen referencec64582d3
2
Professor reference
BestOf 2025 · 1Reference
8.270.7940.71165.26A40Success 10 imagesFrozen reference39f46132
3
Professor reference
VanillaGANReference
49.880.4670.17165.41A40Success 10 imagesFrozen reference4b4d8ba1

Generated images

10,000 PNG images, 28 × 28 pixels, are required. Historical FID uses Inception features on all images; historical P/R uses VGG features with k=5 and 9,984 images (complete batches of 256). Scores are recomputed; the leaderboard is not a project grade.

Run 20261021T160156Z · Generation limit: 1200 s. Student generation: 4 CPU threads / 12 GiB; historical references may use one A40. Metrics use the evaluator GPU. Time excludes dependency installation and metric computation.