source string | shards_used int64 | bytes_read int64 |
|---|---|---|
timm/imagenet-1k-wds | 76 | 11,451,115,520 |
Does REPA narrow SiT representations toward the teacher?
Linear-probe study of frozen SiT-XL/2 features: does REPA-style alignment to a frozen vision encoder make the diffusion model's representation "narrow" toward that teacher (better at what the teacher is good at, worse at the rest), improve it uniformly, or move information across layers? No diffusion training; only forward passes of released checkpoints and linear probes.
The full report (in Vietnamese) is REPORT.md. Short version below.
Short answer
No. At 400K steps (and at 4M), REPA does not narrow the representation in the sense of losing what the teacher does not emphasise. It improves every probe, moves information up to the layers around the aligned block, and the teacher decides only a small part of where the gains go.
- Against an unaligned SiT-XL/2 at 400K, all 16 aligned students (8 teachers × REPA/iREPA) are better at their best layer on every probe (each difference ≥ 2 seed-σ): ImageNet top-1 +11.8 to +31.3 points, NYUv2 depth δ1 +0.075 to +0.167, clean-latent reconstruction equal or slightly better (≤ 0.005 R²).
- The teacher's classification strength does not carry over (Spearman over 8 teachers, ρ −0.17 to 0.19). Its depth strength does with iREPA: profile test (depth − classification, controlling for overall teacher strength) ρ = 0.81, p = 0.022, consistent for s = 0.1–0.5; not with REPA (ρ ≤ 0.55, p ≥ 0.17). Moderate evidence.
- Alignment displaces the classification/depth peak from block 20 (unaligned) to blocks 8–12. Right after the peak (blocks 12–16) latent R² is up to 0.019 below the unaligned model — the only loss observed: small, local, teacher-independent, recovered in later blocks.
- Same teacher: iREPA beats REPA at block 8 for all 8 teachers (top-1 +3.4 to +14.5 points, δ1 +0.016 to +0.063); REPA matches or beats iREPA at blocks 16–20.
- Late training: REPA 4M beats REPA 400K on every probe; HASTE (alignment stopped at 250K, 2.5M steps) has lower CKNNA with DINOv2 than REPA 4M but is not better on depth or reconstruction. SiT 7M (unaligned, similar FID to REPA 400K) is far behind REPA 400K on classification (−16.5 points) and depth (−0.059 δ1), equal on reconstruction.
Not tested: semantic segmentation (optional probe, not run), non-linear or higher-resolution probes, intermediate checkpoints of one run, causal links to FID, alignment at blocks other than 8, and model size: only SiT-XL/2 was probed (the iREPA collection and the original REPA release contain only XL). A smaller model could narrow from lack of capacity; a larger or longer-trained one could follow its teacher more closely. Section 7 of the report gives a minimal design.
Setup
- Set 1: 16 SiT-XL/2 400K checkpoints from REPA-E/iREPA-collections
(
exps_irepa_400k: DINOv2-B, DINOv3-B16, DINOv3-H16+, WebSSL-DINO-1B, MoCo v3-B, CLIP-L, PE-G, LangPE-G; each with REPA and iREPA), the 8 teachers themselves (last-layer patch tokens), and a noisy-VAE-latent lower baseline. - Set 2: unaligned SiT-XL/2 400K (snap-research/elit), original SiT-XL/2 7M (willisma/SiT), REPA 4M, HASTE 2.5M, SiT + Dispersive loss 400K.
- Inputs: 256×256 center crops, SD-VAE posterior mean × 0.18215, null class. Noise ratio s ∈ {0.1, 0.3, 0.5, 0.7} with x = (1−s)·x0 + s·ε, mapped to each codebase's time convention; same noise bank for every model. Features at the output of blocks 4, 8, …, 28; EMA weights, strict load.
- Probes (linear, frozen features, identical hyper-parameters for every model): ImageNet classification (mean-pooled tokens, 50 imgs/class train, 10k val), NYUv2 depth (per-token log-depth, RMSE and δ1), clean-latent reconstruction (R²), CKNNA with the teacher / DINOv2-B (5k images), token participation ratio. 3 seeds (each holds out a different 10% fold to select regularisation); differences < 2 seed-σ are treated as inconclusive.
Files
| Path | Content |
|---|---|
REPORT.md |
Full report (Vietnamese): pre-registered scenarios, results, deviations, known weaknesses, rerun commands |
results/results_long.csv |
All measurements, long format: model, method, teacher, steps, layer, s, task, metric, mean, std, n |
results/set1_results_long.csv, results/set2_results_long.csv |
Same, split by set |
analysis/analysis.md |
Teacher scores, Spearman tests, profile tests, pairwise comparisons |
analysis/*.csv |
The same tables as CSV, plus per-model layer profiles |
analysis/heatmaps/ |
Layer × noise heatmaps for every metric and model |
samples/ |
4×4 class-conditional sample grid (CFG 4) per diffusion checkpoint, a sanity check of each load |
probe/ |
Pipeline code; probe/config/{smoke,set1,set2}.yaml are the configs of the three runs |
Reproducing
probe/ runs from the root of a checkout of end2end-diffusion/irepa
(base commit 99ad4ac; it imports that repo's ldm/ for SiT and the teachers). On one 24 GB GPU:
cp probe/config/set1.yaml probe/config/active.yaml # or smoke.yaml / set2.yaml
bash probe/run.sh # installs probe/requirements.txt, then python -m probe.main
python -m probe.analyze --csv results/set1/results_long.csv --csv results/set2/results_long.csv --out analysis
run.sh assumes a vast.ai-style image (/venv/main, data under /workspace/repa_probe); edit the paths at the top
of run.sh and probe/common.py for other machines. ImageNet train images come from timm/imagenet-1k-wds (gated;
accept its terms on Hugging Face). A full SiT-XL/2 checkpoint takes about 25 minutes on an RTX 4090.
Licence
No licence has been chosen for the code and tables yet. The probed checkpoints, teachers and datasets keep their own licences; none of them are redistributed here.
- Downloads last month
- 128