Pixel Anchored Remaster

You searched, you found a node called PixelAnchordRemaster, and something felt off - because it is. "Anchord" is missing an e. That's not a different node, a fork, or an impostor pack. The PixelAnchoredRemaster pack registers the same class under two names: PixelAnchoredRemaster (the correctly spelled one the README documents) and PixelAnchordRemaster (a misspelled alias that points at identical code). Both display as "Pixel Anchored Remaster" in the node list, and both behave exactly the same.

So the first question a beginner asks - "which one do I install?" - has a boring but reassuring answer: neither, separately. The single pack NotProniss/PixelAnchoredRemaster provides both. If you load a shared workflow that references PixelAnchordRemaster, ComfyUI will tell you the pack is missing; install the pack and it resolves, because that class name is registered right alongside the correct one.

What the node actually is

The underlying node is a "remasterer," not a conventional upscaler. It's built for an image that has already gone through a 2x HiRes Fix pass. The idea:

1. Downscale the image in pixel space to 75% of its size (with lanczos interpolation and white padding) - this sheds the harsh, noisy texture that aggressive upscaling leaves behind.
2. VAE-encode the result using tiled encoding, then upscale the latent back up to the input size (default bislerp), giving the sampler room to work.
3. Run a standard KSampler pass - dpmpp_3m_sde_gpu / karras, 16 steps, CFG 4, denoise 0.55 - to rebuild detail, then tiled-decode back to pixels.

It's the hi-res-fix idea (second pass at low denoise) applied after an upscale instead of during one. Expect minor drift from the source - the goal is a better final image, not a pixel-perfect clone. If you want exact fidelity, this is the wrong tool.

The inputs you'll actually touch

• image - your already-hires-fixed 2x image. Wrong input is the #1 mistake.
• model, vae, positive, negative - the checkpoint and conditioning for the remaster pass.
• seed - for reproducible output.
• remaster_denoise (0.55) - the main quality knob. Too high drifts the composition, too low does almost nothing.
• tile_size_vae (512) / overlap (64) - tiled VAE settings; shrink the tile size if VRAM is tight.

Outputs: image (the remastered result) and latent (the pre-decode latent, if you want to chain something else onto it).

Install

Identical to the main node - it's the same pack. No model downloads, no Python dependencies; the node is pure ComfyUI built-ins.

• ComfyUI Manager → search "PixelAnchoredRemaster" → install → restart.
• Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/NotProniss/PixelAnchoredRemaster

…then restart ComfyUI.

The gotchas, specific to this spelling

• Dedupe if you see both. Some workflows load the misspelled name, some the correct one. They're interchangeable, but don't leave two copies of the node in one graph out of confusion - pick one spelling and keep it.
• The pack's limits apply to both names. It's built for 2x upscale workflows only (4x "goes off the rails"), and it's tuned for SDXL / Pony / Illustrious latent spaces. Anima performs poorly - with blurry results on portrait inputs, square images only - and other divergent latent spaces like Chroma are warned against too. Those limitations don't care which spelling you used.