AI summaryⓘ
The authors studied how different parts of text-to-image diffusion models respond to specific changes, instead of relying on a single overall quality score. They tested 23 models with various architectures and distillation methods by applying controlled perturbations to three layers and measuring responses with a new selective metric. They found that each layer captures distinct information: the prompt layer detects the average prompt effect, the latent layer identifies the type of prediction used, and the score layer reflects the distillation method. A key discovery is that only models using rectified-flow techniques show a strong signal in the latent layer, distinguishing them from others. These detailed layer-wise insights help understand where differences in model behavior come from during distillation and training.
diffusion modelstext-to-image generationmodel distillationFID scoreCLIPUNetrectified-flowlatent layerdenoiserInception features
Abstract
Few-step distilled diffusion students cut text-to-image inference from ~50 to 1-8 network evaluations, but the quality gap is usually summarised by a single FID/CLIP scalar that cannot say which axis of the conditioning response changed, nor whether a behaviour comes from the architecture, the distillation objective, or simply from being a diffusion model. We replace the scalar with a decomposable probe that injects controlled perturbations along three layers (prompt encoder, denoiser input, denoiser output) under three modes (mean, variance, scale) and six strengths, reporting a bootstrap-median Bures W2^2 selectivity ratio on Inception features. Under a single matched estimator across 23 models -- five teachers and 18 distilled students spanning five backbone families (SDXL, SD1.5, SD3.5, PixArt-alpha, FLUX), three architecture classes (UNet, DiT, MMDiT), and five distillation paradigms -- the three layers read three empirically separable factors: the prompt layer is a universal prompt-mean response (a sanity channel, not a discriminator), the latent layer reads the prediction type, and the score layer reads the distillation objective. Our main result: within this sweep, the latent layer is a near-binary detector of rectified-flow backbones. Its ratio exceeds 1 across a sustained low-to-mid band only for rectified-flow models (SD3.5, FLUX); no epsilon-prediction model qualifies. A matched epsilon-prediction control (PixArt-alpha) rules out wide-T5 conditioning, and the fingerprint survives adversarial (ADD) distillation as both teacher and student. Two secondary score-layer findings hold under narrower scopes: a canonical 4-step ADD-vs-rest contrast on the UNet families with a non-ADD baseline, and a CI-separated trajectory-rollout early-strength score spike on both UNet and DiT. All ratios are CI-citable under one estimator; we release the per-cell tables and the estimator.