Title
Two‑Stage Cascade Rendering Workflow for High‑Fidelity SD Upscaling Using Proteus Models
Abstract
This document describes a two‑stage cascade rendering workflow designed to maximize restoration quality when upscaling SD sources (NTSC, PAL, analog TV, digital SD) to 4K resolution using Topaz Video AI. The workflow relies on a strict 2× intermediate scaling factor, 10‑bit processing, controlled noise injection, and a refinement‑focused second inference pass. The method significantly improves temporal stability, texture fidelity, and perceived sharpness while avoiding over‑enhancement artifacts.
1. Overview of the Cascade Rendering Approach
The workflow uses two sequential inference passes:
-
First Pass — Proteus Standard
Purpose: structural cleaning, denoising, artifact removal -
Second Pass — Proteus Natural
Purpose: detail refinement, texture enhancement, cinematic rendering
The two passes serve different roles and must not be interchanged.
2. Rationale for the 2× Intermediate Scaling Factor
The intermediate upscale is performed strictly at 2×, e.g.:
-
480i/p → 960p
-
576i/p → 1152p
This choice is critical for three reasons:
-
Model Training Alignment
Proteus models behave optimally at integer scaling factors.
Non‑integer factors (e.g., 1.7×, 2.3×) introduce:-
temporal instability
-
micro‑vibrations
-
loss of fine detail
-
inconsistent sharpening
-
-
Optimal Deblurring and Denoising
A 2× upscale maximizes the model’s ability to:-
remove compression artifacts
-
neutralize analog overscan noise
-
stabilize edges and contours
-
-
Stable Base for the Second Inference
The second pass requires a clean, stable, artifact‑free base.
A 2× first pass produces the most consistent signal for refinement.
3. First Inference Pass — Proteus Standard (Cleaning Stage)
Objective
Remove noise, compression artifacts, ringing, mosquito noise, and overscan contamination.
Key Principles
-
No added detail
-
No added grain
-
No sharpening
-
No enhancement of residual imperfections
Reasoning
Adding detail or grain during the first pass amplifies:
-
MPEG‑2 artifacts
-
analog noise
-
motion blur edges
-
background inconsistencies
The first pass must produce a neutral, clean, stable signal.
4. Second Inference Pass — Proteus Natural (Refinement Stage)
Objective
Enhance real detail while preserving optical characteristics such as:
-
depth of field
-
motion blur
-
background softness
-
facial focus hierarchy
Settings
-
Noise: 2
Adds controlled “clean noise” to break 8‑bit banding and smooth gradients. -
Recover Details: 20
Enhances real textures (skin, fabric, iris detail) without creating artificial crispness.
Why Proteus Natural
Proteus Standard enhances everything uniformly.
Proteus Natural enhances only areas detected as sharp, preserving:
-
motion blur
-
out‑of‑focus regions
-
background softness
-
secondary faces
-
optical realism
This produces a cinematic, non‑synthetic final image.
5. Importance of 10‑bit Processing
All processing is performed in 10‑bit, not 8‑bit.
Reason
8‑bit pipelines introduce severe color banding, especially after:
-
denoising
-
sharpening
-
gradient reconstruction
-
compression artifact removal
10‑bit processing ensures:
-
smooth gradients
-
stable skies and walls
-
no posterization
-
better noise modeling
-
higher fidelity for the second inference pass
This is a critical requirement for high‑quality restoration.
6. Results and Benefits
The workflow provides:
-
superior temporal stability
-
preserved optical characteristics
-
natural skin texture
-
stable motion blur
-
clean backgrounds
-
no over‑sharpening
-
no artificial detail
-
dramatically improved SD‑to‑4K fidelity
This method often produces results comparable to professional studio restorations.
Cheers, Vincent.
