Nodes/ComfyUI_StarNodes/⭐ Star Tiled PiD Upscaler
ComfyUI Node

⭐ Star Tiled PiD Upscaler

Upscales an image with a PiD / PixelDiT model by processing it in overlapping tiles to keep VRAM usage low. Each tile is re-rendered in pixel space (4 LCM steps) and blended back seamlessly, with per-tile color matching.

By Starnodes2024·Created 2 years ago·Updated 7 days ago· 103
⭐ Star Tiled PiD Upscaler
  • image
  • IMAGE
model_name
clip_name
vae_name
latent_formatqwenimage
scale4.00
rows2
cols2
tile_overlap0.25
Category⭐StarNodes/Image And Latent

Inputs (9)

NameTypeDefaultDescription
imageIMAGEThe image to upscale.
model_nameCOMBOThe PiD / PixelDiT diffusion model, e.g. pid_qwenimage_1024_to_4096_4step_bf16.safetensors.
clip_nameCOMBOGemma 2 2B text encoder for PixelDiT (gemma_2_2b_it_elm). PiD uses an empty prompt, but the model still needs the text encoder loaded.
vae_nameCOMBOVAE matching the PiD backbone, used to encode the input image (e.g. qwen_image_vae for qwenimage, ae.safetensors for flux).
latent_formatCOMBOqwenimageLatent format of the PiD backbone. Flux1 (16-ch) and Flux2 (128-ch) are auto-detected under 'flux'.
scaleFLOAT4.001–8Upscale factor for the output image. PiD models are trained for 4x.
rowsINT21–16Number of tile rows. More rows = smaller tiles = less VRAM.
colsINT21–16Number of tile columns. More columns = smaller tiles = less VRAM.
tile_overlapFLOAT0.250.05–0.5Overlap ratio between tiles. Higher values reduce seam artifacts but increase VRAM/time.

Outputs (1)

NameTypeDescription
IMAGEIMAGE