๐ง Smart Tile Detailer
DetailerForEach, but the tiles do the detailing
- image
- segs
- model
- vae
- conditionings
- negative
- bundle
- image
- segs
Impact Pack's DetailerForEach made a name for itself by detecting regions and re-sampling each one with its own prompt. The Smart Tile Detailer is the same idea pointed at a different target: instead of detecting faces or people, it processes your tile grid, giving each tile its own conditioning and compositing the results in one node. It's the one-node-everything workhorse of the Smart Tile pipeline in ComfyUI-ArchAi3d-Qwen (Amir Ferdos's pack) - calculator, SEGS, conditioning in, detailed image out.
How it works
The source is explicit about the design choices. Each SEG (one per tile) gets sampled with the matching conditioning from Smart Tile Conditioning - so tile 3 in your grid uses tile 3's prompt. It uses the SEG masks directly for blending (pre-blur them with SEGS Blur for softer seams), and it processes tiles at their original size rather than upscaling them again - no double-resizing drift. The expensive encoding already happened in the Conditioning node, so the Detailer is just sampling + compositing.
Inputs
- image - the upscaled base (the calculator's
scaled_image). - segs - tile SEGS from Smart Tile SEGS, optionally blurred with SEGS Blur.
- conditionings / negative - the per-tile list and shared negative from Smart Tile Conditioning.
- model / vae / seed / steps / cfg / sampler_name / scheduler / denoise - sampling stack;
denoisedefaults to 0.4 (refine, don't rewrite). - bundle (optional) - calculator's bundle auto-fills
image.
Outputs: image (the composited, detailed result) and segs (the processed SEGS, in case you want to inspect or chain further).
Installing
cd ComfyUI/custom_nodes
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git
cd ComfyUI-ArchAi3d-Qwen
pip install -r requirements.txt
Or ComfyUI Manager โ "ArchAi3d Qwen". No model downloads. Free for personal use; commercial work needs the paid license.
The honest take
This is the node I'd point a beginner at first, because it's the smallest possible "smart tiled detail" workflow: Calculator โ SEGS โ Conditioning โ Detailer, and you're done - the compositing is internal, so there's no Merger to wire and no blend-mode decision to make. The trade is exactly that: the Detailer owns the blending, using the SEG masks, so you get whatever seam quality the masks give you. If you pre-blur with SEGS Blur (and you should), that's plenty. Reach for Sampler + Merger only when you want the Merger's fancier normalized blending or its blend modes. And remember the whole "per-tile prompt" premise only pays off if the Conditioning node got real per-tile prompts - skip that and you've built a slower plain tiled upscale.
Inputs (13)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | Input image (scaled_image from Smart Tile Calculator) | |
| segs | SEGS | SEGS from Smart Tile SEGS (optionally blurred with SEGS Mask Blur) | |
| model | MODEL | Diffusion model for sampling | |
| vae | VAE | VAE for encode/decode | |
| conditionings | CONDITIONING_LIST | List of conditionings from Smart Tile Conditioning | |
| negative | CONDITIONING | Negative conditioning from Smart Tile Conditioning | |
| seed | INT | 00โ18446744073709550000 | Random seed for sampling |
| steps | INT | 201โ100 | Number of sampling steps |
| cfg | FLOAT | 7.01โ30 | CFG scale |
| sampler_name | COMBO | euler | Sampler algorithm |
| scheduler | COMBO | normal | Noise scheduler |
| denoise | FLOAT | 0.400โ1 | Denoise strength (lower = preserve more detail) |
| bundleopt | SMART_TILE_BUNDLE | Connect bundle from Smart Tile Calculator (auto-fills image) |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | โ |
| segs | SEGS | โ |