HyperTile (bleh)

HyperTile is one of the older speed tricks still floating around ComfyUI - it dates back to October 2023, well before SageAttention or step-distilled models became the default answer to "how do I make this faster." It's not dead, exactly, just quiet: most of today's speed conversations happen elsewhere, so if you go looking for fresh community chatter about it you won't find much. What you will find is a genuinely real technique with a real quality tradeoff, and this node adds scheduling controls the original never had.

The mechanism: self-attention is expensive, and its cost scales badly with image size. HyperTile tiles the attention computation into smaller local chunks instead of computing it globally, which is much cheaper - the README's original pitch (from the technique's creator, not blepping) claimed 3-4x speedups at 4K resolution. The tradeoff is that local-only attention can introduce structural repetition or artifacts if you push it too hard, since parts of the image lose the "big picture" context that global attention would have given them.

The credits trail on this node is almost a running joke in the README itself: originally from tfernd's HyperTile repo, taken into ComfyUI core, taken from core into the Inspire pack, taken from Inspire into bleh. What bleh actually contributes on top is the scheduling layer - a seed, a timestep range, and a step interval - so HyperTile doesn't have to run for the entire generation. Positive interval values apply it every N steps; negative values (like -3) apply it for the first two of every three steps and skip the third, the same interval convention bleh uses elsewhere in the pack. start_step/end_step gate it by ComfyUI's raw timestep scale, which runs from 999 down to 0 and isn't linear - the README doesn't pretend this is intuitive, noting that figuring out which sampling step a given timestep corresponds to "is left as an exercise for you, dear node user."

Inputs that matter: model, tile_size (256 default - the size of each local attention tile), swap_size (2 default - how tiles get shuffled between passes so seams don't line up predictably), max_depth (0 default - how many U-Net depth levels get the tiling treatment), plus seed, interval, start_step, and end_step for the scheduling bleh adds. Output is a patched MODEL.

Worth taking at face value: the author flags this node himself, in his own words, as "Not well tested, and I just assumed the Inspire version works which may or may not be the case." That's about as candid as node documentation gets, and it's a reasonable warning to actually heed - if you hit weird behavior, you're mostly on your own testing and the README, not a deep well of community troubleshooting threads.

Install is standard: ComfyUI Manager search for "ComfyUI-bleh," or a manual git clone https://github.com/blepping/ComfyUI-bleh into custom_nodes followed by a restart. No extra dependencies for this node. If you see visible tiling seams or texture repetition, that's the known cost of the technique working as designed, not a configuration mistake - try a smaller tile_size, a smaller max_depth, or narrowing the step range it's active for rather than assuming something's broken.