ComfyUI Node
Latent-Tiled PiD Decode
Decodes the latent as overlapping tiles through PiD — every tile stays inside the trained 1024->4096 envelope, avoiding the color collapse of single-shot decoding at large sizes — then feather-blends (seam-free, weights sum to 1).
Latent-Tiled PiD Decode
- model
- positive
- latent
- IMAGE
◄latent_formatqwenimage►
◄max_tile1024►
◄overlap64►
◄seed0►
◄degrade_sigma0.00►
Categorylatent/pid
Inputs (8)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | The PiD checkpoint (UNETLoader). | |
| positive | CONDITIONING | PiD text conditioning (Gemma CLIPTextEncode; empty prompt is fine). | |
| latent | LATENT | The generation latent straight from your KSampler (flux/flux2-klein/sd3/sdxl/qwen-family). | |
| latent_format | COMBO | qwenimage | Family of the STAGE-1 latent. 'flux2' = FLUX.2 dev + Klein 4B/9B (128-ch, 16x); Flux2 is also auto-detected under 'flux' from the channel count, matching core PiDConditioning. |
| max_tile | INT | 1024256–4096 | Max tile size per axis in stage-1 px. 1024 = PiD's trained envelope. |
| overlap | INT | 6416–512 | Tile overlap in stage-1 px (x4 in output; feather band). |
| seed | INT | 00–18446744073709550000 | — |
| degrade_sigma | FLOAT | 0.000–1 | PiD noisy-latent conditioning: raise for partially-denoised latents. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| IMAGE | IMAGE | — |