Latent-Tiled PiD Size Picker

The math that trips people up with the Latent-Tiled PiD Decode node is the relationship between latent size and output size. Output is always 4x the latent, and tile count is derived from that - so to get a specific print size you have to know the matching latent dimensions off the top of your head. Latent-Tiled PiD Size Picker deletes that whole problem. It's one dropdown where every line is a complete, tested configuration:

16:9 | render 1920x1088 -> output 7680x4352 | 33 MP | 4 tiles

That line means: "give me a 1920x1088 latent, it will decode to 7680x4352 (33MP) in 4 tiles." Aspect, stage-1 render size, final output, megapixels, tile count - all of it, encoded in the label. There's nothing else to configure. Wire it where you'd put an EmptySD3LatentImage and pick a line.

Why it exists

The Size Picker is basically the decode planner's dropdown form. It's the "start here" node of the pack: the author render-tested all 56 presets across 9 aspect ratios end-to-end on Krea 2 Turbo, with per-preset verdicts in docs/preset_sweep_report.md. Range is single-tile 4K-class up to 244MP (15616x15616). Nothing in the list is theoretical - every entry was verified to keep all tiles inside PiD's trained 1024→4096 envelope. There's deliberately no sub-1024 single-tile entry: below their trained res, stage-1 models go mushy and PiD reinterprets instead of decoding, so the ~1024 single-tile line is the floor on purpose.

How it works

Under the hood it mirrors core EmptySD3LatentImage - it emits a 16-channel zeros latent at h//8 × w//8 with the 8x downscale_ratio_spacial flag, on the intermediate device. The decode node's planner then derives the tile grid from whatever latent size it gets, so the dropdown's "tiles" column and the decode's plan always agree. Three earlier ~210MP 15-tile rungs were culled in v1.1.1, then restored in v1.2.0 after PiD v1.5's color-fidelity fix cut their chroma roughly in half on retest - a good example of why the QA node exists.

Inputs and output

Two inputs, both boring: size (the dropdown) and batch_size (1–64, default 1). One output: LATENT, wired into whatever you'd feed the sampler - then the Decode node → SaveImage.

Which lane? This picker is for the qwen-family / Krea 2 Turbo 8x-latent world, which is the fully-swept lane. The FLUX.2/Klein family (128-channel, 16x-downscale latents) uses its own node - the FLUX.2 / Klein Size Picker - because its latent space sits on a 16-px grid. Other qwen-family stage-1 models should behave like Krea 2; the Decode node has flux/sd3/sdxl paths but they're untested, so if you run one of those, run the QA node and trust its verdicts over hope.

Install

ComfyUI Manager - search "ComfyUI-Latent-Tiled-PiD" - or clone:

cd ComfyUI/custom_nodes
git clone https://github.com/BennyDaBall930/ComfyUI-Latent-Tiled-PiD

Restart ComfyUI; nodes appear under latent/pid. Requires ComfyUI ≥ 0.28 (PiD in core) and no python dependencies, but you still need the PiD checkpoints and text encoder on disk (NVIDIA weights, non-commercial license).

Gotchas

• Don't use this node for Klein. Grab the Flux2 picker instead - the grids don't line up.
• Dropdown values can change between versions; if one entry throws "Unknown size preset", re-select from the list rather than fighting it.
• The picker only creates the empty latent. If you want a 33MP print you still have to run the actual generation through the Decode node - this node is upstream, not a shortcut around the decode.