ComfyUI-PBaSR

ComfyUI-PBaSR

A minimal ComfyUI custom node for PBaSR x4 inference. It vendors only the inference-time architecture code needed for:

• PBaSR_SRFormer_best.pth
• PBaSR_ESRGAN_best.pth

FeMaSR is intentionally not included in this lightweight node because the PBaSR FeMaSR config depends on the additional FeMaSR HQ-stage weight.

About PBaSR

PBaSR is an upscaling model for depth-of-field blur-kept image super resolution, especially for realistic photo.

Install

Copy this folder into:

ComfyUI/custom_nodes/ComfyUI-PBaSR

Download models here and put them in:

ComfyUI/models/pbasr/PBaSR_SRFormer_best.pth
ComfyUI/models/pbasr/PBaSR_ESRGAN_best.pth

Restart ComfyUI. Add node:

image/upscaling -> PBaSR Upscale (SRFormer/ESRGAN)

Settings

Recommended first run for SRFormer:

arch = PBaSR_SRFormer_x4
model_name = PBaSR_SRFormer_best.pth
tile_size = 512
tile_pad = 64
precision = fp16

For lower VRAM, use tile_size=384 or precision=fp32 if fp16 causes artifacts on your GPU.

Notes

• This is an inference-only port and does not include training, metrics, datasets, pyiqa, lmdb, or full BasicSR.
• The node loads the PBaSR wrapper ParallelNet_ncb and uses the averaged combinet branch, matching PBaSR test-time behavior.
• Large images require tiled inference. SRFormer x4 is much heavier than RRDB/ESRGAN.