Krea Moodboard Search

The name is the whole pitch: type what you want the style to feel like - dark teal gothic, warm pastel product, grainy documentary - and Krea Moodboard Search digs through all 3,549 bundled Krea moodboards and hands you the best match as prompt text. No images, no API calls, no key, nothing to download. It's the text-only workhorse of the pack, and it's the node to reach for when you want style control that's automatable and deterministic rather than eyeballed.

Context first. Krea's hosted product has a moodboard/style-reference module, and Krea explicitly kept it closed when the Krea 2 weights went open in June 2026. The community answer was to copy styles from the public gallery and let an LLM merge them in - people did this by hand and got great results, which is exactly the workflow this pack industrializes. Search is the "find me a style" step, as pure text.

How it works

This is genuinely local, dependency-free search. The pack ships a prompt-only catalog (data/krea_moodboards_slim.json) where every moodboard has been rewritten into a typed style schema - palette, lighting, medium and texture, composition, atmosphere, era. Search scores each board across eight fields: title, keywords, taste profile, prompt guidance, style axes, conditioning notes, source summary, and negative guidance.

The nice touch is the query expansion. Common words get mapped to the way Krea actually words things - dark finds noir, photo finds documentary, bright finds luminous, anime finds manga, retro finds vintage - and matches in a title or keyword field score higher than matches buried in prose. So you can search the way you talk, not the way the catalog was written.

The search always returns something usable: top_k controls how many matches you preview, min_score is a floor, and if you set random_from_top_k above 0 it picks a random board from the top N matches (seeded by seed) instead of always taking the best one.

The inputs that matter

• query - your style idea. Default is dark teal gothic, but anything mood/color/lighting/medium-ish works.
• top_k - how many matches show in preview (1–25, default 5).
• min_score - the acceptance floor (default 1). If you're getting nothing, drop this.
• random_from_top_k + seed - set both if you want a seeded random pick from the top matches rather than the single best.
• strength - concise / normal / strong; how much detail goes into the style text.
• random_mode - balanced / any / non_photo / photo. Relevant only when you use random selection.

The outputs

positive (the style block), negative (filtered negative guidance), title, metadata_json, and preview (a scored list of matches with URLs). The wiring is simple:

• positive and negative → Krea Moodboard Apply → the Krea 2 Turbo subgraph's Text String (User Prompt) input
• metadata_json → a board_1–board_4 slot on Krea Moodboard Mashup

The negative output deserves a note: it's quality-safe. Guidance has been filtered so it can never ban your subject or fight image quality - no "avoid detail" or "avoid photorealism" leaking in, which used to quietly sabotage Krea 2 renders.

Install

Comes with the Krea Moodboards pack. Manager search "Krea Moodboards", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Andro-Meta/ComfyUI-Krea-Moodboards.git

Restart ComfyUI. No external Python dependencies, no weights, no key - this pack is prompt text only. Nodes show up under Andro.Meta / Moodboards.

Common issues

The big one: if nothing matches your query, the node raises an error ("No Krea moodboards matched query") and the workflow stops. Loosen it - drop min_score, use simpler words, or lean on the synonym mapping (dark → noir, not gloomy gothic shadows drenched in despair). Check preview if a pick surprises you: it shows the score and which fields matched, so you can see why "warm pastel product" returned a particular board. And remember, Search is text - if you want to see thumbnails before committing, that's the Visual Browser's job.