Tile & Tag SEGS

The pain this solves is the annoying one: you're detailing a big image with a SEGS-based workflow, and every region needs its own prompt. If you've ever hand-written prompts for six tiles that turned out to contain things you couldn't see from the thumbnail, you know the feeling. Tile & Tag SEGS is the automation: it cuts your image into a deterministic grid of tile regions, runs a WD14 tagger over each crop, prefixes your shared positive text, and hands back the SEGS together with a matching conditioning batch. Each tile gets sampled with a prompt about what's actually in it.

It's a detailer input generator, not a sampler. You still do the actual refining with a SEGS detailer (SimpleSyrup's own scale-factor detailers, or Impact Pack's, since SEGS is the shared vocabulary here - that's the whole point of the format). This node just removes the "figure out what each region contains" step.

How it works

The node does four things in one pass. First, a tiler builds a deterministic scan-order grid of overlapping bounding boxes across your image. Second, it crops each region and feeds the crops through the connected WD14 tagger, which emits Danbooru-style tags at a confidence threshold. Third, it prefixes universal_positive to every generated tag string and CLIP-encodes each one. Fourth, it returns the tile SEGS and the conditioning batch in the same order, so the Nth SEG's prompt is the Nth conditioning entry. The plumbing that keeps them aligned is the pack's CONDITIONING_BATCH structure - same mechanism its [SEP] prompt batching and regional samplers use.

The inputs that matter

• bbox_size - the target tile size in pixels (default 872). This is your resolution knob; smaller tiles mean more, smaller detail passes.
• threshold vs character_threshold - the confidence cutoffs for general tags (default 0.35) and character tags (default 1.0). A character threshold of 1.0 is effectively "only keep character tags you're certain about," which makes sense for content tiles.
• universal_positive - text prepended to every tile prompt, so your style anchor travels with the auto-tags.
• exclude_tags - comma-separated tags stripped from the generated prompts. The default (1girl, solo, long_hair, short_hair, silhouette) is worth keeping an eye on: those are exactly the tags you don't want in a per-tile prompt.

The tile-geometry controls (crop_factor, min_overlap, filter_segs_dilation, and the mask_irregularity/irregular_mask_mode pair for organic, non-rectangular masks) are the fiddly ones you'll touch when tiles start showing seams or when rectangular masks cause visible region boundaries in the final composite.

Outputs and what they wire into

• segs - the tile regions, usable by any Impact-compatible SEGS detailer.
• positive - the aligned conditioning batch, matching the SEGS order. This is what you feed to a regional sampler or detailer instead of writing prompts by hand.

Installing it

SimpleSyrup installs as a whole pack. ComfyUI Manager: search SimpleSyrup under Node Pack and hit Install, then restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Artificial-Sweetener/SimpleSyrup.git
cd SimpleSyrup
# then install requirements with the Python that runs ComfyUI:
../venv/bin/python -m pip install -r requirements.txt

You also need a Load WD14 Tagger node in the graph feeding this one, and the tagger downloads its ONNX model from Hugging Face on first use - that's a few hundred MB the first time.

Common issues

The tags are Danbooru vocabulary, full stop. If your checkpoint is an Illustrious or NoobAI finetune this is a gift; if you're running a natural-language model like Flux, WD14's tag soup is a mismatch - the generated conditioning won't know what to do with it. And because this node leans on the tagger, it inherits the usual WD14 load: the model download is a real network dependency, and a first run that looks hung is often just the tagger still downloading. Keep character_threshold where you want content tags included; at the default it's basically off, so tiles get only confident general tags plus your universal positive.