f8ch32 VAE chases pixel-shift fidelity

This feels less like a brand-new method than a brute-force way to force translation robustness and subpixel consistency into a decoder. The instinct is solid, but the real question is whether it beats smarter loss design or just burns compute on alignment noise.

• –The closest precedent is patch-sampling augmentation in super-resolution, where informed crop selection improves convergence and detail recovery.
• –If you want exact image identity, feature-space losses like LPIPS are always a compromise: they can sharpen perception, but they stop caring about pixel-for-pixel truth.
• –Recent alias-free and shift-equivariant latent-diffusion work suggests a more principled version of the same idea: regularize shift behavior instead of multiplying crops forever.
• –The strongest ablation here is probably PSNR/SSIM plus shift-consistency against plain L1 and edge-L1 baselines before adding any perceptual or adversarial terms.