v1.2

Big update for Pixel Anchored Remaster.

What changed in v1.2

The main improvement is that HiRes Fix now feeds directly into the final Pixel Anchored Remaster pass instead of treating them more separately.
This ended up being a huge quality improvement and is basically the core reason v1.2 is better.

Main v1.2 changes

• HiRes Fix output now becomes the input for the main Pixel Anchored Remaster stage

• switched the final latent upscale method from bicubic to bislerp

• increased final remaster denoise to 0.55

• cleaned up and reorganized the workflow

What the workflow is doing

This is not a traditional upscaler and it is not meant to perfectly preserve every detail.

The goal is:

controlled drift for better final output

The workflow works best as a creative remaster pipeline:

• start with a good original image

• improve it with HiRes Fix

• then run it through the final Pixel Anchored Remaster stage

• accept some drift if the overall vibe, composition, and subject stay intact and the final image looks better

Compatibility testing so far

v1.2 seems to work well with refined SDXL / Illustrious / Pony-family checkpoints, with some model-specific tweaking.

Good results with:

• WonderMix v16

• WAI Illustrious

• some Pony-derived checkpoints

• some realistic refined checkpoints

Less successful with:

• base Illustrious

• base Pony v6

• Anima (anything that uses 16 point latents)

So far the pattern seems to be:

If a model can make a strong original image, the HiRes Fix + Pixel Anchored Remaster pipeline will usually improve it, sometimes significantly.

Important note

This workflow is better thought of as a remaster than an upscale.
If you want strict preservation, this may not be the right tool.
If you want the same image idea/vibe but cleaner, more detailed, and more polished, this is where it shines.

v1.1

Based on community feedback regarding background softening, v1.1 introduces key mathematical refinements to the pixel-anchor stage to dramatically improve architectural crispness, background preservation, and edge definition—while keeping the main subject completely perfect.

🛠️ What Changed under the Hood:

• Resize Interpolation: area ➔ lanczos (Restores sharp geometric lines and window leading)

• Upscale Method: bilinear ➔ bicubic (Cleaner pixel-space translation before the latent pass)

• Remaster KSampler CFG: 3 ➔ 4 (Tighter prompt and structural adherence)

• Remaster KSampler Denoise: 0.55 ➔ 0.50 (Slightly lower denoise to anchor original details while preventing hallucinations)

⚠️ Note on NSFW Content

Because this is a hybrid remaster method that relies on a second KSampler re-imagining details over a pixel foundation, it can still occasionally struggle with explicit NSFW anatomy or specific interactions. If things look a bit weird or muffled, try slightly lowering the Remaster KSampler denoise (down to 0.45 or 0.48) to keep closer to the original structural layout!

v1

🔍 Pixel-Anchored Remaster vs. Hires Fix: The Ultimate Comparison Canvas

Ever wondered exactly how your upscaling methods stack up side-by-side? This is a clean, organized comparison tool designed to test three distinct upscaling philosophies simultaneously from a single master seed.

Instead of running separate generations and guessing which one looks better, this workflow generates your base image once and instantly branches out to give you a definitive visual shootout.

🎛️ The 3-Way Shootout Branches:

1. Raw Pixel 2x: A baseline upscale utilizing the standard model loader 2xAoMR_mosr.pth) to show the raw, un-denoised upscaled foundation.

2. Standard Hires Fix: The classic latent-space enhancement method utilizing VAE Encode/Decode passes to introduce fresh detail.

3. Pixel-Anchored Remaster: An advanced approach that anchors downscaling/area-mapping metrics before the final latent pass to preserve prompt fidelity while significantly modernizing details.

🌟 Key Features:

* Fully Bundled & Grouped: Clean bounding boxes ("Settings", "Original Image", "Hires Fix", etc.) make the board incredibly easy to navigate.

* Shared Seed Link: A single master Seed Primitive controls both generation KSamplers so your comparisons remain perfectly synchronized.

* Massive Preview Fields: Pre-arranged, massive preview blocks make it easy to inspect fine details (like eyes, textures, and lineart) right inside ComfyUI without needing an external image viewer.

🛠️ Requirements & Notes:

Built on *ComfyUI Core** nodes for maximum compatibility.

Uses native *SDXL** parameters (1024x1024 base resolution).

* Requires an upscale model loader (configured for 2xAoMR_mosr by default, but swappable with your favorite ESRGAN/DAT model).

Feel free to leave a review or drop your comparison grids below! If you enjoy the template, keep an eye out for the streamlined, single-output production version coming soon.