High‑quality Danbooru similarity search and learned style similarity retrieval for ComfyUI. Combines image & tag embeddings with flexible weighting and negative prompt control, plus a separate HNSW index–based style matcher.

Features

Fast FAISS search over Danbooru 2022 embedding index
Hybrid tag embedding: ONNX text encoder + hash fallback
Image + tag blending strategies
Post metadata/tags/url/id retrieval
Auto danbooru image fetching (api/html)
Danbooru Style similarity node

Installation

Place this folder inside ComfyUI/custom_nodes/ then install Python deps

pip install -r requirements.txt

resource files (indexes, style embeddings/checkpoints) are downloaded automatically

Nodes Overview

| Node | Purpose | |------|---------| | DanbooruImageSearch | Multi-image semantic retrieval from Danbooru index using combined image/tag positive & negative logic. Returns images + direct (or fallback) URLs + aggregated IDs. | | GetDanbooru | Fetch a single post by URL (API + HTML fallback) with optional resizing (zoom). | | TagPrompt | Normalize / dedupe / sort / limit a tag string for clean input to search nodes. | | StyleSimilarity | Style-based nearest-neighbour search over large learned embedding set (gustproof). Supports checkpoint versioning and optional similarity overlay. |

Danbooru Lookup Parameters

| Parameter | Type / Values | Default | Purpose | |-----------|---------------|---------|---------| | download_size | preview, thumbnail, large, full | preview | Choose which size variant to request (falls back if unavailable). | | max_images | int [1–50] | 4 | Maximum images to return (search + download budget base). | | positive_image | IMAGE (optional) | – | Image embedding used as positive semantic anchor. | | negative_image | IMAGE (optional) | – | Image embedding subtracted (avoid similar look). | | positive_prompt | STRING (tags/text) | "" | Comma‑separated Danbooru tags or free text for positive semantics. | | negative_prompt | STRING (tags/text) | "" | Tags / text to push away (subtract). | | api_username / api_key | STRING | "" | Optional Danbooru credentials (higher rate limits). | | model_type | clip, siglip | clip | Backbone for text/image tag processing. Requires matching clip_models/{model_type}_text.onnx for ONNX tag embeddings. | | tag_embedding_mode | auto, hash_only, onnx_only | auto | How tag vectors are built (ONNX multi‑hot vs hash fallback). | | combine_strategy | Balanced, Stepwise de-overlap | Balanced | Vector combination mode (see combination notes below). | | tag_new_info_only | bool | False | Removes tag components already present in image vector (orthogonalizes tag vectors). | | reduce_negative_overlap | bool | False | Projects negative components away from positive mixture to avoid over‑cancellation. | | idf_weight_tags | bool | False | Apply IDF weighting from tag frequency file (selected_tags.csv). | | fetch_mode | auto, api, html | auto | Metadata retrieval path (API→HTML fallback or forced). | | extra_api_search | int [0–1000] | 0 | Additional recovery attempts for failed posts (expands candidate fetch). | | debug | bool | False | Verbose internal logging. | | image_weight | float [0–4] | 1.0 | Weight applied to image embeddings before combination. | | tag_weight | float [0–4] | 1.0 | Weight applied to tag embeddings before combination. |

Embedding Combination (Core Logic)

| Strategy | Internal Name | Behavior | |----------|---------------|----------| | Balanced | balanced_diff | L2‑normalize positive (image+tags) and negative sides separately, then subtract and renormalize. | | Stepwise de-overlap | stepwise_deoverlap | Sequentially add positives and subtract negatives, orthogonalizing at each step to reduce component overlap. |

Optional modifiers:

tag_new_info_only: orthogonalize tag vectors against image basis (adds only new semantic info)
reduce_negative_overlap: project negative vectors off the positive mixture (less semantic washout)

Tag Embedding Modes

| Mode | Behavior | When to Use | |------|----------|-------------| | auto | Try ONNX multi‑hot (if model + tag index available), blend hash vectors for unknown tags, fallback to pure hash. | Default robust path. | | hash_only | Deterministic SHA‑256 based 1024‑D vectors per tag (stable, no external model). | Offline or minimal setup. | | onnx_only | Only real tags via ONNX multi‑hot. Unknown tags ignored; if encoder missing returns None. | Highest fidelity (requires clip_text.onnx + selected_tags.csv). |

Style Similarity Parameters

| Parameter | Type / Values | Default | Purpose | |-----------|---------------|---------|---------| | input_image | IMAGE / batch | – | Query image(s) whose style embedding(s) are computed. | | max_images | int [1–100] | 4 | Number of nearest style matches to return. | | model_version | v0.2, v0.3 | v0.3 | Select checkpoint (downloaded on first use). | | download_size | preview, thumbnail, large, full | preview | If API lookup succeeds (not disabled), choose size variant. | | overlay_similarity | bool | False | Renders raw squared L2 distance text onto each returned image. | | disable_api | bool | True | Skip MD5→post metadata/API upgrade (keeps original CDN URLs). |

Style Retrieval Notes

HNSW index persisted after first build (style_hnsw_raw.index).
Distances shown are raw squared L2 (lower = closer style).
Query embeddings coerced to index dimension if mismatch (padding/truncation + renorm).

Attribution

SmilingWolf – Danbooru embeddings
gustproof – Style similarity model
ComfyUI project & community