TBG ETUR Refiner PRO

The TBG ETUR Refiner PRO is the heart of the whole TBG pack. The Tiler node splits your image and builds the TBG_Pipe; this node is where the actual tiled sampling happens - tile by tile, with each tile getting its own prompt, denoise, and seed, then getting stitched back together so the seams don't look like a quilt. If you've only got room to learn one node in this pack, this is it.

The pack itself is by Tobias Laarmann (TBG / TB_LAAR, TBG______ on Reddit), and it's an unusually ambitious tiled upscaler-and-refiner. It's still labeled beta and gets updated nearly daily, which cuts both ways: you get new features fast, and you also get the occasional regression. PRO nodes like this one need a free Patreon membership plus an API key (TBG_ETUR_API_KEY as an environment variable is safer than pasting it into the node, because pasting it embeds the key in your workflow metadata).

How it works

The refiner takes the TBG_Pipe from the Tiler, which carries your upscaled image, the tiles, masks, and any per-tile prompts. It then samples each tile with your model and fuses them back together. The headline fusion mode is Neuro-Generative Tile Fusion (NGTF): instead of just blurring tile borders, it remembers the surrounding generated content and adapts later sampling steps to it, so you can run higher denoise without the color shifts and seams you'd normally get. There's also a Soft Merge mode for low-denoise work.

The other thing that makes this pack distinctive is how it treats the scheduler. denoise_method defaults to "normalized advanced," which splits sigmas by noise values rather than step percentages, so you can swap samplers without losing the creative behavior you dialed in. And vae_encode defaults to "tiled slow" but offers "tbg Color-preserving fast" and "Nvidia PiD 4x" - the latter hands decode to NVIDIA's pixel-diffusion decoder, which is a fast 4x upscaler in its own right. The KB's take on PiD is worth remembering: it's an upscaler that invents detail, not a lossless VAE swap, and it fixed a mixed CPU/CUDA crash in v1.2.8.

Inputs that actually matter

You must connect model, clip, vae, and TBG_Pipe. Set model_type to match your model (FLUX1, FLUX2, Qwen Image, SDXL, SD3, Z-Image, and more) - it changes the assumed native tile size. Beyond that:

• steps (30) and denoise (0.27) - denoise is the single biggest lever. Low and slow for faithful upscales, higher for creative re-imagining.
• Flux_Guidance (3.5) - the tooltip says it plainly: if areas aren't blending well across tiles, raise it.
• VRAM_Profile - defaults to "Low VRAM Cache (Unload Models)." If you're on 8–12 GB, leave it there; "Fast Cache" precomputes everything and will eat your card.
• Image Stabilizer, Smoother-Sharper, Detail_Enhancer - all default 0 (off). Turn them on one at a time, not all at once.
• Scale-Invariant Feature Transform (on by default) realigns each generated tile back to its reference, which is what kills the creeping drift that plagues big tiled jobs.
• Fast_1_Tile_Preview - render one tile first and check your settings before burning an hour on the full run. Use this. Always.

Optional Controlnet_Pipe, Redux_Style_Model/Redux_Clip_Vision, denoise_mask, and RF_UntwistingRoPE (from the pack's RF pipe node) plug into the same spot. Outputs are Refined (the money shot), plus Refined without Segs, Refined without ColorCorrection, Original Upscaled, Original, and Tiles - the ablated outputs are genuinely useful for diagnosing whether it's the fusion or the color match doing something you don't like.

Install

ComfyUI Manager → search "TBG" and install "TBG Enhanced Upscaler" node set, or:

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

Restart ComfyUI. This pack needs a recent ComfyUI - it registers through the v3 API (comfy_api.v0_0_2) - and its requirements list is heavy: opencv-contrib-python (SIFT), kornia, transformers/diffusers/accelerate (PiD), timm, triton, and optional Qwen-VL bits. The README also lists a long roster of models you'll want (flux fp8, t5xxl encoder, Flux VAE, Redux, sigclip vision, ControlNet models). "A ton of extra models needed to dl before it works" is the community's #1 complaint about this pack, and it's fair.

Common issues

• VRAM OOM on big images - use the Low VRAM or Ultra Low Memory profiles, and don't run "Fast Cache" on a small card.
• Seams or color bands at tile borders - raise Flux_Guidance, lower Fusion Strength on the Tiler, or switch Color_Match.
• Daily-update breakage - the README tells you to update as often as they ship. When a fresh update acts weird, check the changelog before assuming you broke something.