Superside Smart Detail Sheet

Every e-commerce listing needs a shot that shows the product and its interesting details up close - the hinge, the logo, the texture of the fabric. Building those "detail callout" sheets by hand is a fiddly Photoshop job. Superside Smart Detail Sheet automates it end to end: a vision LLM finds the visually interesting details in your photo, the node crops each one, upscales the crops locally, and composites the original plus enlarged callouts into one finished spec-sheet image. Feed it a photo, get a deliverable.

How it works

The detection is a two-mode story, and the mode is picked by product_category:

• auto (default) - a vision LLM (chosen by the model dropdown: gemini-2.5-flash, gemini-2.5-pro, gpt-4o, claude-sonnet-4.6) looks at the image and returns JSON bounding boxes for the num_details (1–6, default 3) most interesting details. The node retries automatically if the model returns prose instead of JSON, a zero-size box, or overlapping zones - a nice robustness touch.
• eyewear - for glasses, free-form LLM judgement proved unreliable (it kept conflating "nose pad" and "hinge" onto the same spot), so this mode forces exactly three fixed zones using Florence-2's grounding endpoint instead: the nose pad, the hinge screw, and a temple tip. It overrides num_details and model.

From there it's local image work: each crop is a fixed-size square centered on the detected detail's center point (not the raw bounding box edges - that's what keeps crops consistent despite imprecise boxes), upscaled with Lanczos (crop_scale, 1–4x, default 2 - no extra API call for the upscaling), and composited into the final layout: side column for portrait originals, a row below for landscape/square. Crops that land on a flat, blank region are auto-discarded - the info output tells you what was kept and what was dropped.

The inputs that matter

• image + api_key - the required pair.
• product_category - auto or eyewear. This is the decision that changes the whole detection path.
• num_details - how many callouts in auto mode (ignored for eyewear).
• detail_hint - free text to steer the auto model toward what you think is interesting.
• crop_size_percent - each crop's square size as a percent of the original's shorter side, default 35%. Smaller keeps callouts subtle; larger makes them dominant.

Outputs: image - the finished sheet - and info, JSON with the kept details, discard count, and settings used.

Installing it

It's in comfyui-superside-nodes - clone, install requirements, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/Superside/comfyui-superside-nodes.git
cd comfyui-superside-nodes
pip install -r requirements.txt

Paste your fal api_key. No model files; the detection call is metered and your product photo leaves the machine for the vision model.

The honest take

This is the most genuinely clever node in the pack - a real production workflow (spec sheets) compressed into one drop. The trade-offs: the auto detection is only as good as the vision model, so check the info output on the first pass and use detail_hint to steer it; and because the detail selection is probabilistic, a batch of fifty products needs a spot-check pass before it ships. For a catalogue where the details must land in the same spots every time, you're better off with the pack's Manual Detail Sheet (you draw the boxes once per profile) - the Smart version shines when every product is different and you want the interesting parts found for you.