Latent diffusion: measure the representation bottleneck before denoising it

For an image with height H and width W, a spatial downsampling factor f and latent channel count C_z produce C_z H W/f² latent elements when dimensions divide evenly. Compare element widths and actual storage before quoting a compression factor. A sixteen-bit latent tensor has a calculable raw bit count, but that is not an achieved entropy-code rate or proof about retained information. Decoder weights, metadata, conditioning, and encoder computation are outside this simple tensor-storage count.

A smaller latent grid can reduce work in a convolutional or transformer denoiser, but exact savings depend on channels, patch size, attention, and architecture. Increasing downsampling can remove fine detail or make reconstruction more difficult. The latent diffusion paper studies a particular perceptual-compression design and tradeoff; it does not make every autoencoder equally suitable. A narrow latent bottleneck can generate plausible texture while failing to preserve the exact original detail needed by a reconstruction task.

Evaluate held-out inputs passed through the encoder and decoder without diffusion. Inspect pixel or perceptual distortion using a declared normalization and metric, including small text, thin structures, and rare content. A reconstruction metric can reveal a representation failure before denoiser training starts. It is not a universal lower bound on every generated-image quality metric: a generator can produce perceptually plausible alternatives without reconstructing a particular input. Keep fidelity to an original and distribution-level sample quality separate.

A stochastic encoder may supply a posterior sample or another declared latent representation. Replacing samples with posterior means changes that distribution. Freezing encoder weights does not automatically define its train/evaluation mode or random state. Store the autoencoder artifact, latent scaling and shift, precision, and preprocessing configuration. If latents are cached, preserve the intended augmentation and sampling policy; one cached crop or posterior draw can remove variation that an online pipeline would otherwise present during training.

Measure latent coordinate statistics under the actual encoder path and preprocessing. A configured scaling factor can make the training representation closer to a desired scale, but verify how it is applied and reversed by the decoder. Missing or doubled scaling changes the effective signal-to-noise ratio while leaving tensor shapes valid. Denoiser targets and sampling scheduler must use compatible latent conventions. Test a round trip with fixed inputs and compare a small latent-noising calculation before interpreting a failing sample grid as a backbone problem.

The function lab isolates spatial-area growth with side length. Reset, set a=1 and x=1, predict y=1, then Play or Step. Separately Reset, set a=1 and x=2, predict y=4, then Play or Step. With fixed channels and downsampling, doubling both spatial sides multiplies the latent element count by four. The instrument contains no encoder, decoder, image, or distortion metric. It supports this geometric budget calculation only, not a claim that every denoiser's time or memory scales exactly with area.