NA - Styling Detail Change

Whole-frame image editing is great until you want to change just the collar. Ask a generic editor to "make it a v-neck" and it'll happily re-draw the sleeves, the fabric texture, and the lighting while it's at it. NA - Styling Detail Change is Neural Atelier's attempt at the surgical version: change one construction detail on a garment - neckline, sleeve width, hem, fit - while holding the texture, color, and silhouette steady. It's aimed at fashion teams running fit variations without a new photo shoot.

How it works: the pack ships a garments.json that defines a cascading menu - 21 garment types (tshirt, polo, shirt, knitwear, blazer, jogger, and more), each with a set of detail categories (neckline, neckline_height, neckline_binding, sleeves, sleeve_width, body_length, hem, fit, shoulder, details), and each category with concrete options like crew_neck vs v_neck. Pick a garment type and the node's dropdowns repopulate to match - the options you see are genuinely context-dependent, not a flat list. If you leave description empty, the pack fills in a sensible default for the category (e.g. the neckline default template with your chosen option); otherwise your description takes over. Then the usual two-stage pipeline runs: Gemini 3 Flash builds a prompt from the template plus invariant rules that tell it to preserve texture, color, and silhouette, and Nano Banana Pro (gemini-3-pro-image-preview) does the edit with your garment image as reference. Hook up the optional detail_reference_image and the model also gets a visual example of the detail you're after.

Inputs that matter:

• garment_image (required) - the garment you're editing.
• garment_type, detail_category, detail_option - the cascading triple that defines the edit.
• description - your own spin on the detail (optional; defaults kick in if blank).
• brief - broader context, like the intended silhouette or fabric.
• The usual suspects: gemini_api_key, image_model, aspect_ratio, resolution, top_p, rerun_nonce.

Outputs: image, plus prompt (the prompt Flash actually generated - read it when a result drifts and you'll see why), log, and output_json with the change type, option, and timings.

Install:

cd ComfyUI/custom_nodes
git clone https://github.com/valsecchi75/comfyui_neural_atelier
cd comfyui_neural_atelier && pip install -r requirements.txt

then restart ComfyUI (or ComfyUI Manager → search "Neural Atelier"). No model downloads - this runs on Google's paid Gemini API, so count on per-image cost and Google's safety filters.

Where people get burned: the cascading dropdowns depend on configs/NA_Styling_Detail_Change/garments.json - if that file's missing, the node falls back to a stub garment list and the menu loses its intelligence, so re-clone rather than hand-editing. Like its Recolor sibling, this node has a disk cache that only serves results when rerun_nonce is 0, and ComfyUI caches node output independently - if you change the description to iterate and nothing moves, bump rerun_nonce to force a fresh API call. And one honest caveat: "surgical" is the goal, not a guarantee. Nano Banana is powerful, but subtle construction details on busy fabric are exactly where an edit model's idea of a "v-neck" can disagree with a patternmaker's.