Identity Signal Analysis in Diffusion Model Latent Spaces

Investigating whether facial identity information can be isolated, measured, and manipulated in diffusion model latent spaces through frequency-domain and channel analysis.

Setup

Experiments run on Google Colab (Pro, L4/T4 GPU). The notebook clones this repo, installs deps, and runs each experiment.

1. Open identity_signal_analysis.ipynb in Colab
2. Connect a GPU runtime (L4 recommended)
3. Run setup cells, then the active experiment

Experiments

#	Name	Status	Key Finding
1	Paired Frequency Analysis	Done	Ch 2 shows largest frequency shifts on identity change
2	Within-Identity Invariance	Done	Ch 3 is best identity discriminator (ratio 0.046)
3	Identity Emergence	Done	Identity locks in at steps 7-9 during denoising
5	Channel Identity Transplant	Done	UNet heals single-channel swaps; Ch 3 swap shows most identity shift
6	Channel Importance (Zeroing)	Done	Ch 0 catastrophic, Ch 3 significant, Ch 1/2 minimal
7	PCA on Identity	Next	Is identity a linear subspace in Ch 3?

See RESEARCH.md for full findings and the hierarchical channel model.

Architecture

identity_analysis/       # Core library (pipeline, FFT, scoring, plotting)
experiments/             # Experiment scripts (each has run() entry point)
identity_signal_analysis.ipynb   # Colab notebook
EXP{N}_RESEARCH.md      # Per-experiment reports
RESEARCH.md              # Cross-experiment synthesis

Model

SDXL Base 1.0 (fp16), 1024x1024, 12 steps, guidance 7.5 + negative prompt.