The pixel shift technique that lostinspaz is exploring deserves more attention than it's getting. It addresses a real, practical bottleneck in VAE training, reconstruction fidelity, without relying on the usual crutches that introduce their own artifacts. For anyone who has wrestled with LPIPS smoothing or GAN hallucination, this approach offers a refreshingly direct alternative: brute-force accuracy through training data diversity rather than loss function cleverness.
What makes this method compelling is its simplicity. By taking a high-resolution image, resizing it to a slightly offset dimension, and then extracting every possible stride-1 crop, you effectively multiply your training set while teaching the model to handle sub-pixel variations. The result is a VAE that must learn to reconstruct with precision across a continuous range of shifts, not just at fixed grid boundaries. That's a fundamentally different challenge than what standard random cropping provides, and it directly targets the fidelity gap that plagues compressed latent spaces.
The initial results, better than SDXL f8ch4 but not yet matching AuraFlow f8ch16, are honest and useful. They tell us the method has promise, but also that the tuning game is real. The real work ahead lies in weighting loss functions like L1 and edge-aware L1 to balance sharpness against overfitting. That's not a sexy problem, but it's the kind of hands-on optimization that separates a working technique from a robust one. Lostinspaz is right to ask for prior work; this area is underexplored in public literature, and the community would benefit from shared experiments.
We hope others take up this thread. Pixel shift is not a magic bullet, but it is a concrete, reproducible technique that prioritizes fidelity over visual polish. That matters when your downstream task, whether image generation, editing, or analysis, depends on the latent space preserving what was actually there. The next step is not a new loss function or a bigger GAN. It's running the experiments, sharing the weights, and letting the data speak.