Nodes/Anima PiD/Anima PiD Decode (4x SR)
ComfyUI Node

Anima PiD Decode (4x SR)

A ComfyUI node in Anima/PiD with 10 inputs and 1 output.

By sorryhyun·Created 2 months ago·Updated 30 days ago· 14
Anima PiD Decode (4x SR)
  • pid_model
  • latent
  • IMAGE
◄steps4►
◄sigma0.00►
◄seed0►
◄tile_latent64►
◄tile_overlap16►
◄compilefalse►
◄use_calibfalse►
◄attentionauto►
CategoryAnima/PiD

Inputs (10)

NameTypeDefaultDescription
pid_modelANIMA_PID—
latentLATENT—
stepsINT41–8—
sigmaFLOAT0.000–1Latent degradation level PiD assumes. 0.0 = clean decode; higher lets PiD synthesize/hallucinate more detail.
seedINT00–18446744073709550000—
tile_latentINT640–2560 = decode whole image at once (4K output may OOM on <=16GB). >0 = tile the latent (each tile -> tile*32 px) with feather blending. 64 -> 2048px tiles.
tile_overlapINT160–64Latent-space overlap between tiles (pixels = overlap*32). Larger = fewer seams, slower.
compileBOOLEANfalsePer-block torch.compile of the PiD net: each transformer block is compiled as its own small graph (faster compile, fewer graph breaks than whole-net — mirrors Anima Block Compile). First run per output size is slow (compilation), then fast; with tiling on all tiles share one size so the blocks compile once.
use_calibBOOLEANfalseApply the bundled PiD->native-VAE color-match transform after decode. It was fitted against the pre-v1.5 qwenimage checkpoint, which decoded flat and desaturated vs the native Qwen VAE. v1.5 fixes colour accuracy upstream, so this is OFF by default — applying it on v1.5 would over-correct. Turn on only for a hand-placed v1 checkpoint.
attentionCOMBOautoAttention backend for the PiD net. 'auto' honors ComfyUI's --use-sage-attention / --use-flash-attention launch flags (sage wins), else plain SDPA. 'sdpa' already runs PyTorch's built-in flash kernel here, so 'flash' is ~identical; 'sage' (INT8) is the real speedup. Masked/fp32/CPU paths always fall back to SDPA.

Outputs (1)

NameTypeDescription
IMAGEIMAGE—