Danbooru Corpus Maintenance

If Danbooru Prompt Forge is the front of Eclipse's Danbooru workflow, this is the maintenance bay behind it. Corpus Maintenance does two jobs: it pulls new posts from the live Danbooru API to keep the rating corpora fresh, and it prepares the two-pass LLM categorization batches that assign general tags to categories - or exports a provider-neutral manual work package instead. It's the node you queue periodically, not the one you wire into every generation.

Understand the context first. Prompt Forge reads offline taglist files; those files ship pre-built (four rating corpora, 100k posts each, plus a ~132k-tag categorized index, extracted at first startup). Maintenance is how you grow and curate that corpus yourself. That part needs a Danbooru account and API key, set under Eclipse → General → Danbooru Maintenance. No credentials are needed if you're happy with the shipped corpus or want to drop in your own lists from rainlizard/ComfyUI-Raffle - the node only requires them for online scraping.

How it works

The core is a resumable scraper with checkpoints. It walks /posts.json pages in score bands - the default automatic mode samples the 1024–5000 band first, then halves downward (512–1024, 256–512, …) until it finds unseen posts. Every response is committed before the next request, deduplicated by post ID, and appended to the matching rating file, up to target_per_rating (capped at 100,000). Then it enriches from the /tags.json catalog, publishes authoritative artist/character/copyright/meta tags directly, and funnels only uncategorized general tags into an ordered backlog for LLM categorization. A global 3,000-request safety cap per run keeps you from hammering the API.

The actions chip bar selects the phases: refresh_ratings, optional catalog-only refresh_catalog, mutually exclusive prepare_ai or manual_categorization, and resume. On an imported corpus you can run catalog enrichment alone by turning off refresh_ratings.

Inputs worth knowing

• actions / ratings - the combo-chip bars at the top. Everything on this node runs from those.
• target_per_rating - how many unique posts to retain per rating. Existing larger corpora are never truncated.
• score_range_mode - automatic (adaptive bands) or custom (one fixed custom_score_min/custom_score_max range).
• excluded_post_tags - deny any newly returned post containing an exact listed tag. Whole-tag matching only: excluding rape does not exclude grape.
• ai_batch_size / maximum_ai_batches - tags per SmartLLM request (default 100) and how many batches to prepare per queue.

The AI categorization path

When prepare_ai is selected, the node emits the four outputs that drive a two-pass LLM pipeline: categorization_system_prompt and categorization_prompt feed Smart LM Loader #1, whose output feeds Loader #2's user_prompt; review_system_prompt feeds Loader #2's system_prompt; and batch_token goes to Danbooru Category Apply. That requires the separate ComfyUI SmartLLM pack. The author's tested recipe: a Qwen 3.x instruct model in the 8B/9B class or larger (Qwen 3.8 27B is the strongest tested; Qwen 3.5 9B is the smallest that's exact-preserving), greedy decoding (do_sample=false, num_beams=1), text-only, no Multi-Task. The pack docs name specific models that failed this job - Ministral 3, abliterated Gemma 3 4B, Qwen 2.5 VL 3B - so trial anything else at ai_batch_size 50 or lower first.

No local LLM, no SmartLLM? Pick manual_categorization instead and the node writes a complete, provider-neutral export under prompts/tag_lists/manual_categorization/export-<hash>/ - prompts, rules, numbered inputs - that a remote chat model or agent can process. Eclipse only prepares that package; consuming the results is on you.

Install and gotchas

Same pack as Prompt Forge: ComfyUI Manager → search "Eclipse", or git clone https://github.com/r-vage/ComfyUI_Eclipse custom_nodes/ComfyUI_Eclipse and restart. The maintenance half additionally wants ComfyUI SmartLLM installed.

Real-world traps:

• Resume is your friend. Checkpoints are per-rating, per-score-band, saved after every response. Turning resume off rewinds to page 1, and rewinding only works when the complete checkpoint prefix exists. A network failure that kills a run never loses earlier committed pages.
• Credentials go in Eclipse's private config.json, never into workflow JSON or manifests. Clearing the login or API key removes the stored value.
• Batch manifests expire after seven days. Prepared AI batches older than that get rejected by Category Apply; just re-queue to prepare fresh ones.
• Batch files get two-generation backups under prompts/tag_lists/.backups/ - if a commit hiccups, your corpus isn't silently corrupted.
• For a true fresh start, delete the whole prompts/tag_lists/ directory; Maintenance restores the contracts and empty rating files, and collection restarts from zero.

It's a fiddly node with a lot of knobs, but once it's running it's genuinely set-and-forget: it commits as it goes, respects the API, and picks up where it left off.