TBG ETUR tiled Nvidia PID Image Upscale

PiD - PixelDiT, NVIDIA's Pixel Difference Transformer - is a 1K→4K super-resolution model, and it's the star of ETUR 1.2's standalone nodes. But a 4K-capable diffusion model on a big image is a VRAM wall unless you tile it. This node is the tiling: it takes an input image or latent, chops the job into PiD tiles, runs each one, drift-corrects and color-matches them against the source, and rebuilds the full image. The display name says it all - tiled Nvidia PID Image Upscale - and the two outputs are PID Rebuilt (the finished image) and PID Tiles (a list of the per-tile results, so you can inspect or re-fuse).

It's the standalone PiD path, separate from the refiner. If you've loaded a PiD model with TBG ETUR Download PiD Model, this is what you feed it into.

The inputs that matter

• PID_Model (MODEL) - from the Download PiD node. Required.
• PID_VAE_Compatible_Model (default FLUX1) - tells the pipeline which model family's VAE/latent space the PiD model matches. Get this wrong and the upscale silently degrades.
• upscale_by (default 4, range 0.1–16) - the target factor. 4 is the PiD sweet spot.
• steps (default 4) - PiD is a fast few-step model; 4 is normal, not a typo. Don't crank it to 30 expecting more quality; you'll get more of the same.
• cfg (default 1) - near 1 for PiD. Its conditioning is a light prompt hint, not a full prompt.
• denoise (1.0) / degrade_sigma (0.1) - how much of the input each tile reimagines. degrade_sigma is the knob that says "how noisy/damaged should the model treat this tile as" - the reference upscaler trick for giving the model room to add real detail.
• sampler_name (pid_sde) / scheduler (simple) - the PiD-native sampler; the pack registers it. Leave them unless you have a reason.
• prompt - an optional style hint. PiD's CLIP is small; don't write a novel here.

The two source inputs are exclusive-ish: image alone does a normal PiD upscale with per-tile SIFT drift correction; latent alone makes PiD sample from the latent (VAE-decoding it first for the reference); both connected means PiD samples the latent while the image is used only as the drift/color reference. That last combo is the "clean reference, generative upscale" power move.

Why you'd reach for it

Straight answer: it's the cheapest way to get a genuine 4K generative upscale on consumer VRAM. The upscaling world splits "more pixels" from "more detail" - PiD is squarely in the detail-synthesis camp, and it's NVIDIA-backed, which on an RTX card means the whole pipeline has been tuned with that hardware in mind. The tiling is the part that makes it usable, because an untiled 1K→4K run on a 6–12GB card is asking for the OOM error of shame.

Installing it

Part of the TBG ETUR pack. ComfyUI Manager → search "TBG" → TBG_Enhanced Tiled Upscaler & Refiner FLUX PRO, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Ltamann/ComfyUI-TBG-ETUR
cd ComfyUI-TBG-ETUR
pip install -r requirements.txt

You'll also need the PiD model itself - run the Download PiD node once, and make peace with a multi-gigabyte first fetch from HuggingFace.

Notes and gotchas

Where people get burned: forgetting that the PiD model family must match your pipeline's model (Flux1 PiD on a Flux2 image = wrong), and pushing upscale_by to 8 or 16 on a 4K-trained model (it can, but quality falls off fast past ~4x per pass - do two passes instead). Also: the PID Tiles output is a list, so if you route it into something expecting a single image, you'll get an error that looks like the node is broken. It's not; you just need a list-aware consumer or to skip that output. And the fixed-seed-while-debugging rule applies double here - PiD is deterministic enough that changing the seed and the settings at once will make you chase ghosts.