TBG ETUR Refiner CE

This is the node the whole pack is named after. TBG ETUR Refiner CE ("Community Edition") is the tiled sampler-and-refiner: it takes a tiled image from the ETUR upscaler (via the TBG_Pipe), regenerates each tile with its own prompt and denoise, fuses the tiles back into a coherent whole, and hands you an image that's been refined at a resolution that would have OOM'd any non-tiled approach. "Neuro-Generative Tile Fusion" is the fancy name for the part that matters: each new tile remembers what its neighbors generated, so there are no seams even at high denoise. That's the whole trick of the pack, and this is where it lives.

The inputs that matter

• model_type (FLUX1) - which family you're refining. The menu covers FLUX1, FLUX2, Ideogram4, FLUX1 Kontext, Qwen Image, Qwen Image Edit, SDXL, SD3, Z-Image, and others. Pick the one that matches your loaded model - it sets native tile sizes and behavior.
• model / clip / vae - your actual pipeline models.
• TBG_Pipe - from the ETUR tiler/upscaler node. This carries the tiles, per-tile prompts, denoise, seeds, and ControlNet. Without it the refiner has nothing to refine.
• denoise (default 0.27) - the global refinement strength. Low (0.2–0.35) is the safe band for upscaling; the per-tile and per-pixel denoise controls layer on top of this.
• Flux_Guidance (3.5) - Flux guidance, not CFG. The tooltip is blunt: all fusion modes benefit from high guidance, so if areas aren't blending, raise it.
• Color_Match (default TBG ETUR Detail-Preserving Tile Stabilizer) - tile-aware color stabilization with feathered seam blending; falls back to the global stabilizer when tile-aware conditions aren't available.
• Scale-Invariant Feature Transform (on) - SIFT drift correction: after each tile is generated, it's aligned back to the reference tile. Off when you want raw speed; on is why tiles don't wander.
• vae_encode (on = "tiled slow") - toggle between ComfyUI's standard tiled VAE path and TBG's faster color-preserving path. The tooltip also describes an Nvidia PiD 4x decode option (native fast path at 1024², tiled PiD elsewhere) for FLUX1/2, Qwen, SDXL, SD3, Z-Image.
• General_Prompt_Positive / General_Prompt_Negative - the default prompts applied to every tile. Fun fact: the default negative is a long Chinese quality prompt (low quality, blurry, deformed hands…), which is worth knowing if you ever wonder where all those negatives come from. Replace them with your own; they're just defaults.
• Redux_strength (0) - if you've connected Redux_Style_Model + Redux_Clip_Vision, this sets how strongly the Flux Redux style reference steers the refinement. 0 = unused.
• Fast_1_Tile_Preview - process just the first selected tiles at full scale as a preview before committing to the whole set. This is the "don't run 200 tiles to find out the settings are wrong" button, and you should use it. Always.
• Selected_Tiles_Only / Selected_Tiles_By_Numbers - refine only tiles 1,2,3,6 (or whatever) instead of everything. The post-sampling tile repair feature from the README; also great for the "one tile came out wrong" case.

Optional inputs: Controlnet_Pipe, Enrichment_Pipe, Redux_Style_Model, Redux_Clip_Vision. Output: Refined (IMAGE). It's an output node.

The honest context

ETUR is a small, fast-moving, Patreon-backed project. The CE refiner is free and fully usable; PRO features (deeper tile overrides, some Neuro-Generative fusion depth, API-backed bits) sit behind a TB_LAAR membership, and the README's first line warns you this is a beta updated daily. Community reception in the upscaling threads is split between people impressed by the "100MP" demos and people who bounce off the PRO gate and the heavy requirements. If you're a beginner: the CE path works, the YouTube tutorials exist, and the pack's own advice is to test with PRO features off first. The dependencies are heavy (triton, transformers, diffusers, bitsandbytes, opencv-contrib…), and the install asks you to lower Manager's security level to weak - do that only because you trust this specific repo.

Installing it

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" → install, restart. You'll also want the Flux model files and VAE from the README's list (flux1-dev fp8 is fine, t5xxl fp8 encoders, and ae.safetensors).

Notes and gotchas

Where people get burned: running the full tile set before trying the single-tile preview (the preview exists for a reason), setting denoise above 0.5 and wondering why the composition changed (it's supposed to), and forgetting that this node is beta - if a knob misbehaves, the fix is usually git pull inside the pack folder, since the author ships fixes daily. And if you see an error about comfy_api.v0_0_2, that's the pack telling you your ComfyUI is too old - update ComfyUI, not the node.