Three different meanings of “better images”
An interactive field guide
An image can look plausible and still fail to preserve its input. A representation can reconstruct well and still be poor for understanding. Measure each goal separately.
Before this article: Autoencoders · Probability and tensors
What was already possible
Section titled “What was already possible”Pixel metrics compare pairs. MSE averages squared pixel errors; PSNR is a logarithmic transform of MSE with a stated pixel range. SSIM compares local structure. LPIPS compares learned features. Each metric encodes a particular notion of similarity.
What changes, and why
Section titled “What changes, and why”FID instead compares feature distributions. Reconstruction FID uses reconstructed images; generation FID uses newly generated images. The same statistic receives different inputs. Neither directly checks that every reconstructed object or letter matches its paired original.
Read from left to right. Select a module to connect its role, tensor shape, and calculation. Snippets are teaching pseudocode. Sources provide the paper method or audited implementation.
Paired reconstruction
Keep the pairing for fidelity metrics.
mse = mean((x_hat - x)**2)Distribution comparison
Compare moments of feature distributions with a fixed evaluator.
stats_real = feature_stats(real_images)
stats_model = feature_stats(model_images)Downstream understanding
A separate protocol tests whether the features support a task.
features = frozen_encoder(x)
prediction = trained_probe(features)Work through the mechanism
Section titled “Work through the mechanism”The example contains five scalar samples. Permute them: paired MSE changes, but their mean and variance do not. Then add a brightness offset: the mean changes. This is an exact small counterexample to confusing distribution agreement with pairwise fidelity, not a computation of image FID.
Change a parameter, inspect the computed values, then compare this small example with the model described above.
Analytical teaching model · controls change the example, not a trained research model.
A permutation changes paired error but preserves the distribution. This is not an Inception FID calculation.
| Quantity | Value |
|---|---|
| MSE (paired) | 0.000 |
| Squared mean difference | 0.000 |
| Variance difference | 0 |
Training is not sampling
Section titled “Training is not sampling”A linear probe freezes the encoder and trains a linear classifier. A zero-shot text-image classifier additionally needs a compatible text encoder. VQA needs a language system and its own evaluation. These are different tests, even when they share visual features.
What this does not establish
Section titled “What this does not establish”Record dataset split, resolution, sample count, preprocessing, model size, prior training, and sampling settings. Guidance and rejection sampling can change both quality and cost. Throughput at large batch size is not latency for one image.
Check your understanding
Section titled “Check your understanding”Can lower reconstruction FID establish better VQA?
No. It measures a different target. Use a controlled understanding evaluation with the same training and input conditions.
Sources and next steps
Section titled “Sources and next steps”Diagrams and snippets are teaching reconstructions. Paper reports, implementation observations, and analytical examples are distinguished. No research model runs in this page.