TBG ETUR Upscaler and Tile Generator PRO

The PRO Tiler is the CE Tiler with the training wheels off. Same job - upscale, tile, generate per-tile VLM prompts, hand a TBG_Pipe to the Refiner - but it adds per-tile control, a real fusion mode, and the settings that decide whether your 100MP upscale looks seamless or looks like a grid of differently-colored squares. This is also the node where the pack's Patreon line gets drawn: PRO features need a free TB_LAAR membership and an API key, which you can pass as PRO_api_token or (better) as the TBG_ETUR_API_KEY environment variable so it doesn't get baked into your workflow JSON.

The pack is by TBG (Tobias Laarmann) and is an honest beta - updated daily, so expect the occasional regression along with the new toys.

How it works

The pipeline matches the CE node: ESRGAN-style upscale, split into tiles, optional VLM description per tile, then package into TBG_Pipe. The difference is the fusion strategy. The CE node uses Soft Merge, which limits how much creative (high-denoise) refinement you can do before seams appear. The PRO tiler defaults to Neuro-Generative Tile Fusion (NGTF): each tile's generated surroundings are remembered and fed into subsequent sampling steps, so later tiles "know" what's already been generated around them. That's what lets the Refiner run high denoise and still produce seamless results - the README's whole pitch is "no need for blended borders."

The controls cluster around that:

• Fusion Margin (64) - the blur width used to blend NGTF tiles.
• Fusion Strength (0.95) - strength of the fusion mask. The tooltip gives you the exact fix for the most common complaint: "If you see light color shifts on tile borders, lower this a little." 0 = off, 1 = full.
• Feather Mask (16) - final-composite gradient, separate from the NGTF fusion itself.
• Fusion Reference Margin (0) - reference border content kept per tile and cropped at stitch time.

Inputs that matter

• presets - drops in curated combinations so you don't have to hand-tune everything. Handy until you outgrow them.
• Fragmentation (1.0, 0.5–4) - how aggressively the image is cut up. Higher = more, smaller tiles = less VRAM per pass but more seams to fuse. Raise it only if you're OOMing.
• tile_size_w / tile_size_h (1024 × 1024) - unlike the CE node you can tile non-square, which matters for wide panoramas.
• upscale_by (2.0) and upscale_model ("NONE") - same as CE: pick a real 4x model from upscale_models if you want actual pixel upscaling first, and don't expect magic from a bare NONE rescale.
• Optimize_Upscale_Factor_For_Tile_Use - keeps your tile grid, trims the effective scale so only the needed overlap is used. Leave it on.
• VLM_Model / VLM_Prompt / VLM_seed - per-tile prompt generation. The PRO version adds VLM_Process_Selected_Tiles_Only + VLM_Selected_Tiles_Index_Numbers (e.g. "1,2,3,6"), so you can regenerate prompts for just the tiles you edited.
• Optional Segment_Mask lets you tile around a segmentation mask (objects get their own handling - useful with the per-tile Refiner workflow).

Outputs: TBG_Pipe and the Upscaled Image. Everything interesting happens downstream in the Refiner.

Install

Same as the whole pack - ComfyUI Manager ("TBG Enhanced Upscaler") 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. Heavy requirements (opencv-contrib-python, kornia, timm, triton, transformers/diffusers), needs the v3 API of a current ComfyUI, and you'll be downloading a lot of models per the README list - the "too many model downloads" complaint is the most common pushback you'll see in the community.

Common issues

• Color bands at tile borders - lower Fusion Strength (the tooltip literally tells you to), check Color_Match in the Refiner, or raise Fusion Reference Margin for more context per tile.
• OOM - raise Fragmentation to make smaller tiles rather than dropping your whole workflow.
• API key pasted into the node - it ends up in your workflow and image metadata. Use the TBG_ETUR_API_KEY env var instead.