LamTNguyen/SiT-PCA-FID4K-results

SiT-B/2 single-block PCA experiments

This repository contains quantitative and qualitative results for inserting a PCA projection and inverse reconstruction after each individual transformer block of SiT-B/2. It covers all 48 combinations of:

• block: 1 through 12;
• rank: 64 and 128;
• PCA axis: hidden and token.

Each experiment modifies exactly one block. There are no joint/multi-block interventions in this sweep.

Method

The activation entering the PCA hook has shape [batch, 256 tokens, 768 hidden]. PCA is fitted independently for every sample and every model evaluation.

• Hidden-axis PCA: tokens are observations and hidden channels are features. The centered activation is projected to rank k, inverse-transformed to 768 hidden channels, and passed to the next block.
• Token-axis PCA: the activation is transposed, so hidden channels are observations and tokens are features. It is projected to rank k, inverse-transformed, transposed back to [256, 768], and passed onward.

Sampling protocol:

• model: SiT-B/2, ImageNet 256;
• checkpoint: nyu-visionx/SiT-collections/SiT-B-2-256.pt;
• checkpoint SHA-256: 0c0a3867c24d5f0606de4177f25f60dc4122d5996b07c39dda11247f891ef0f6;
• deterministic balanced sample set: 4 images per ImageNet class, 4,000 images total;
• ODE sampler: Euler, 50 steps;
• CFG: 4.0;
• seed: 0;
• PCA implementation: PyTorch torch.linalg.svd(..., driver="gesvda");
• FID: OpenAI guided-diffusion Inception pool_3 and FIDStatistics against VIRTUAL_imagenet256_labeled.npz.

FID4K results

Lower is better. Bold is the best of the four configurations for that block.

Block	Hidden rank 64	Token rank 64	Hidden rank 128	Token rank 128
1	18.1857	18.1743	18.2044	18.2132
2	18.2055	18.2135	18.1976	18.1887
3	18.1027	18.1174	18.2079	18.2004
4	18.0049	18.0087	18.1519	18.1712
5	17.1081	17.0915	18.2092	18.2294
6	16.5367	16.5131	18.1822	18.1563
7	16.8067	16.8337	18.1639	18.1578
8	16.7375	16.7464	17.8993	17.8906
9	17.4828	17.6022	17.7867	17.7482
10	24.7109	25.5033	17.4716	17.4656
11	30.5780	31.8097	17.5069	17.5158
12	18.5111	18.7334	18.2352	18.2395

The best measured case is token-axis rank 64 after block 6, with FID4K 16.5131. Rank 64 performs well at blocks 5-9 but is highly destructive at blocks 10-11. Rank 128 is much more stable and performs best around blocks 8-11.

There is no matching no-PCA FID4K baseline in this repository, so these values compare PCA variants with each other and should not be presented as improvements over the original checkpoint. FID based on 4,000 samples also has more estimator noise than standard 50K FID; small differences should be interpreted cautiously.

PCA-RGB visualization

The PCA-RGB visualization follows the three-component projection idea in SRA pca-vis. Every case includes:

• PCA-RGB maps for all 12 transformer layers;
• both pre-reconstruction and post-reconstruction maps at the intervention block;
• four noise fractions: 0.7, 0.5, 0.3, and 0.1;
• sign/permutation alignment of the top three components across adjacent maps;
• raw activation reconstruction error, RGB difference metrics, explained variance, and component correlations.

The completed visualization set contains 48 cases and 2,496 individual layer maps.

Repository layout

fid4k/
  RESULTS.md                    # concise results report
  raw/                          # exact evaluator JSON and generation metadata
  qualitative/                  # 48 generated-image grids
pca_rgb/
  axis_<axis>_rank_<rank>_block_<block>/
    pca_rgb_all_layers.png
    target_pre_post.png
    metrics.json
    maps/                       # per-noise, per-layer PCA-RGB maps
code/
  sample_fid4k_pca.py
  fid_only.py
  pca_rgb_visualize.py
  run_rank_axis_sweep.py

The original checkpoint and the ImageNet FID reference statistics are not redistributed here.