Resolution bucketing and tiling

Grouping nearby resolutions can reduce the number of compiled shapes and simplify batch assembly. Padding to a bucket increases processed spatial positions, while resizing or cropping changes image content and conditioning. These are different policies. Preserve the requested aspect ratio and explain any masks, positional coordinates, or conditioning added for padded regions. A pipeline that was trained for particular size and crop metadata may require those inputs to remain consistent with the actual generation and postprocessing steps.

At fixed channel structure, many local operations scale roughly with spatial token count. Global attention arithmetic can scale quadratically in that count, even when a memory-efficient implementation avoids materializing the complete attention matrix. Real model cost mixes several terms and kernel thresholds, so area inflation is only one diagnostic. Profile complete candidate buckets at the actual guidance and precision settings. A bucket with slightly more padding can still be faster if it enables better kernels; the measured result does not erase the additional logical work.

For a chain of odd-sized, unit-stride local convolutions with appropriate matching boundary padding, an interior tile can reproduce the full-image interior when it includes the required halo and discards affected margins. The receptive-field radius grows with the kernels. This statement assumes no operation that depends on the entire spatial domain. Global attention, spatially reduced normalization, resampling, position-dependent context, or other nonlocal stages can invalidate a simple finite-halo argument. Check the actual graph instead of declaring every convolutional decoder exactly tileable.

Overlap and blending can reduce visible seams without making tiled computation identical to full-image computation. Diffusers v0.35.1 documents VAE tiling as a memory option with possible tile-to-tile tone differences. Tiling the VAE decode stage is also different from denoising each spatial tile independently: the latter can lose global composition or use a different noise and conditioning trajectory. Keep the denoiser and decoder choices separate, and compare output artifacts, boundaries, and colors under controlled latents before accepting the memory savings.

Account for overlapping pixels and the edge policy: tiles can duplicate halo computation, require padding beyond the image, and create temporary blending outputs. A memory reduction can increase total arithmetic and transfer costs. Include the final assembly and output buffers in the lifetime budget, and avoid keeping all decoded tiles alive unnecessarily when a bounded streaming assembly is possible. Test extreme aspect ratios, small dimensions, odd divisibility, and high-contrast structures crossing tile boundaries rather than only smooth centered examples.

The function instrument shows how a nonlinear cost term can amplify an input-scale change. Reset, set both a=1 and x=1, predict y=1, then Play or Step. Separately Reset, set both a=1 and x=1.5, predict y=2.25, then Play or Step. Interpreting x as a normalized size ratio is an illustrative analogy to a quadratic term only. The fixed y=ax² model contains no image grid, local convolution, attention kernel, bucket padding, or tiling algorithm. It cannot decide whether a real model has that cost scaling or preserves its output under tiling.