TBG ETUR Download PiD Model

PiD - Pixel Difference Transformer - is NVIDIA's super-resolution model, and in ETUR it's the reason the pack can promise 1K→4K tiled upscales that actually look good. But PiD models are multi-gigabyte downloads from HuggingFace, and nobody wants to hand-fetch them. TBG ETUR Download PiD Model is the loader that makes that painless: it refreshes the model list, downloads the selected PiD diffusion model (and its text encoder, if you want it), and hands you a ready-to-wire MODEL and CLIP.

What it does

• pid_model - a dropdown of PiD models, refreshed from HuggingFace (Comfy-Org/PixelDiT). The default is SuperResolution/PID Flux1 1K to 4K BF16, and the menu includes the other families: Flux2 (BF16 and MXFP8), SDXL, SD3, and Qwen Image, each in 1K→4K. Pick the one matching the model family you're refining - this matters, because a Flux1 PiD on a Qwen pipeline is the wrong tool entirely.
• load_clip (default on) - also load and output the PiD text encoder as a CLIP. If your workflow doesn't need prompting on the PiD pass, you can leave it on anyway; it's cheap once downloaded.

Three outputs: PID Model (MODEL), PID CLIP (CLIP), and PID Model Info (a PID_MODEL_INFO metadata handle some other TBG nodes read). Wire PID Model - and PID CLIP where prompted - into the tiled PiD upscale node or the Labs nodes that accept a PID_Model input.

The first run is the whole game

The gotcha is obvious once you know it: the first time you run this node, it downloads gigabytes. PiD models are not small, and a slow connection will make the node look frozen while it streams from HuggingFace. That's normal. Subsequent runs hit the local cache and are near-instant - this node is a loader, not a downloader you run every time.

If you'd rather not use the node at all, the README's model list has the direct links, and the models land in the usual diffusion_models folder. But then you're on your own for matching the exact variant this pack expects, and the node's refresh logic exists precisely to keep that list current as Comfy Org ships new PiD builds.

Installing it

Part of the TBG ETUR pack:

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

Or ComfyUI Manager → search "TBG" → TBG_Enhanced Tiled Upscaler & Refiner FLUX PRO, restart. You need network access to HuggingFace on first run, and the pack's own requirements (transformers, diffusers, accelerate among others) must have installed cleanly - a broken diffusers install is the classic reason the model loads but the node errors out.

Notes

Two things to keep straight. First, PiD is an NVIDIA super-resolution path, and it's worth remembering that NVIDIA's hardware upscaling is usually described as more "Lanczos" than "generative" - but PiD is the generative end of NVIDIA's work, the model that actually synthesizes 4K detail rather than interpolating it. It's a different beast from the RTX video super resolution nodes. Second, don't confuse this node's load_clip with your main model's CLIP. PiD has its own small text encoder; the "prompt" it takes is a loose style hint, not a full scene description. If you connect the wrong CLIP, the upscale will still run but the conditioning will be noise. Keep PID CLIP on the PiD nodes and your main clip on the refiner.