Restore, Rescale, Reserve & Reshade — A Modular Image Post-Processing Workflow
A modular ComfyUI workflow designed to restore, upscale, and enhance images while keeping unwanted AI alterations to a minimum.
Rather than relying on a single model to do everything, this workflow combines specialized models with distinct roles — from repairing compression artifacts to diffusion-based reconstruction, faithful upscaling, and final visual enhancement.
✦ Four Stages, Four Purposes
01 · Restore — BSRGAN
Repairs common compression artifacts such as color banding, blockiness, and false-color patterns at the original resolution. Designed to clean up degraded images before upscaling, rather than aggressively inventing new details.
02 · Rescale — SeedVR2
Uses diffusion-based reconstruction to upscale clean or restored images. Its reconstruction tends to feel more natural and restrained than aggressive detail-enhancement approaches, helping preserve the overall appearance of patterns, textures, and text — though generative reconstruction can still alter fine details.
03 · Reserve — SwinIR
A lightweight, detail-conscious upscaling stage for already clean, high-quality images. It prioritizes fidelity over aggressive reconstruction and can further increase resolution when SeedVR2's VRAM requirements become a bottleneck.
04 · Reshade — DLSS 5 *
An optional visual enhancement stage focused on the final look of the image. It can be placed after upscaling for full-resolution processing, or before the final SwinIR stage to reduce computational cost.
✦ Flexible by Design
Restore damaged images: Repair compression artifacts before reconstruction.
Upscale clean images: Skip restoration when it is unnecessary.
Go beyond VRAM limits: Use SwinIR for additional upscaling after SeedVR2.
Choose your own finishing stage: Enable, disable, or reposition Reshade according to your needs.
The stages are modular, so you can adapt the workflow to different source images, hardware limitations, and desired output resolutions.
✦ Philosophy
Repair what is broken. Reconstruct what is missing. Preserve what is already good. Enhance what you want to see.
The goal is not simply to make an image larger or sharper, but to give each stage a clearly defined purpose — with a balance between image quality, visual fidelity, and computational cost.
*Note: DLSS 5 ReShade needs a custom node `ComfyUI-DLSS5-NR` and `nvngx_dlssnr.dll`.
