A visual prompt manager + detailer + sampler suite for ComfyUI (formerly Ribbity — ComfyUI Prompt Library). Browse a thumbnail grid of saved prompts, tag them, attach a per-entry LoRA stack, and drop the whole bundle into your workflow with one click. Filesystem-backed — your library survives browser clears and syncs cleanly across machines.

~20 nodes, one shared library, zero ComfyUI custom-node dependencies — every node imports only PyPI packages (torch, ultralytics, sam2, etc.) and ComfyUI core. Nodes that can leverage IPAdapter Plus (Character Anchor) lazy-import it at call time so the rest of the suite loads cleanly without it.

Highlights:

• Library / Style / Multi — visual prompt picker with per-entry LoRA stacks; selection persists across page refreshes via three-tier mirroring (widget value → node properties → localStorage)
• Smart Detailer — one node replaces the 3-node FaceDetailer chain. 6 detail targets (face / eyes / mouth / hands / feet / skin), SAM mask refinement, per-target threshold/denoise/max/steps/crop_factor overrides, color-coded detection-preview output. No Impact Pack required — uses ultralytics directly.
• SDXL + Anima samplers — SDXL_TUPLE-driven (efficiency-nodes wire-compatible) plus a flow-matching sampler for Qwen/Flux/SD3, both with a HiResFix script pipe (5 rounds of caching/perf optimization)
• Save Image (Civitai) — SaveImage replacement that writes A1111/Civitai-compatible PNG metadata; auto-detects model/LoRAs/seed from the workflow trace
• Comic Page (Regional) — single-gen multi-panel conditioning via a color-coded layout mask

The Library node embeds a full thumbnail gallery — search, tag-filter, multi-select, and drop the joined prompt text right into your workflow:

The Comic Frame node combines a character, scene, and background entry into one prompt — so a single workflow can iterate across hundreds of style/character/outfit combinations:

Plug it into your sampler with a Power Lora Loader for one-click LoRA stacks — or use the new Style node (since v0.29.0) to bake the LoRA stack right into the gallery entry:

Get it from one of these — don't grab the unversioned Code → Download ZIP button or random commits.

• ComfyUI-Manager — easiest. Search GrimmRibbity (or Ribbity — Prompt Library) and install. Manager always pulls the latest tagged release.
• GitHub Releases — latest release. Use the Source code (zip) asset, not random older versions.
• git clone — see below; tracks main, which is the same as the latest tag.

1. Download the Source code (zip) asset from the latest release.
2. Extract into ComfyUI_windows_portable\ComfyUI\custom_nodes\. The zip extracts as ComfyUI-PromptLibrary-<version>\ — rename it to ComfyUI-GrimmRibbity so the path becomes …\custom_nodes\ComfyUI-GrimmRibbity\__init__.py.
3. Restart ComfyUI.

cd ComfyUI/custom_nodes
git clone https://github.com/Deaththegrim/ComfyUI-PromptLibrary ComfyUI-GrimmRibbity
# restart ComfyUI

The ComfyUI-GrimmRibbity argument at the end is the local folder name — keep it as shown.

To update later:

cd ComfyUI/custom_nodes/ComfyUI-GrimmRibbity
git pull
# restart ComfyUI

Your data/ folder is gitignored, so git pull won't touch your library. On first launch after a version bump, the package automatically copies data/ to data-backup-<old_version>-<timestamp>/ next to it as a safety net.

After install, find them under GrimmRibbity/ sub-menus in the node picker:

Library (the core suite)

• GrimmRibbity — Library — visual gallery picker. Outputs prompt (positive) and negative STRINGs from the selected entry/entries. If MODEL + CLIP are wired in (optional), the LoRA stack attached to every selected entry is applied in pick-order and patched MODEL/CLIP are emitted on the matching outputs — single node replaces a Library + Power Lora Loader chain. Multi-tile selection stacks every entry's LoRAs.

  

• GrimmRibbity — Style (LoRA + Conditioning) — like the Library node above but takes the next step: also encodes the joined positive + negative on the patched CLIP and emits CONDITIONING outputs (concatted onto optional inputs). Same gallery, same LoRA stack, but skips the separate CLIPTextEncode step. bypass toggle for A/B comparisons; strength_scale multiplies every LoRA at once.

  

• GrimmRibbity — Multi Library (3 panels) — three independent gallery panels in one node, replaces 3× Library + 2× Join Strings spaghetti. Each panel keeps its own search / tag-filter / sort / selection state, persisted per-panel across page refreshes (v0.57.1: fixed a localStorage collision where the three panels stomped each other's saved selection).

  

• GrimmRibbity — Save — write prompts to the library during workflow runs. Inputs: name, text, optional negative, optional IMAGE thumbnail, optional tags, optional prompt_id (override target id), overwrite_by_name (when set, updates the most-recently-edited entry sharing the typed name instead of appending a new one), loras_json (programmatic LoRA stack as JSON). Logs the lookup path on every save (via=prompt_id / via=overwrite_by_name / via=new) so a sticky widget value is visible in the console.

• GrimmRibbity — Random by Tag — pick a random library entry by tag filter (outputs text, id, negative). Built for overnight loops

• GrimmRibbity — Wildcard Expand — expand {a|b|c} alternatives and __name__ library refs in any string

Comic / character workflow helpers

• GrimmRibbity — Comic Frame — combines character + scene + background entries into one prompt, with per-frame seed offset for comic batches

• GrimmRibbity — Scene — per-frame scene knobs (camera_angle, mood, lighting, framing) + free-text extras

• GrimmRibbity — Background (locked) — locked background preset for series consistency

• GrimmRibbity — Character Anchor — wraps IPAdapter Plus's UnifiedLoader + Apply pair into a single MODEL→MODEL transform. Pin a character's face/style across comic panels with one node instead of three. Includes a bypass toggle and an attn_mask input for regional workflows. Requires ComfyUI_IPAdapter_plus.

• GrimmRibbity — Smart Detailer — one node replaces the 3-node FaceDetailer chain. Six detail targets (face / eyes / mouth / hands / feet / skin), each with its own enable toggle, bbox-detector dropdown, and a per-target column in the in-node grid for threshold / denoise / max N / steps / crop_factor overrides. Optional SAM mask refinement so blends follow the actual region outline instead of a rectangular feather. Tiled VAE decode defaults ON for OOM safety, conditional tile policy avoids the overhead on small crops, NaN-scrubbed output. Color-coded detections_preview IMAGE output draws labelled bboxes per target — wire to a SaveImage to debug detection without queueing sample work. No ComfyUI custom-node dependencies — uses ultralytics (and optionally sam2 / segment_anything) directly.

  Drop YOLO detector .pt files into models/ultralytics/bbox/. Recommended set: Anzhc/Anzhcs_YOLOs for face + eyes (YOLO11n), Civitai #329458 for hands (YOLOv9c — catches partial / off-angle hands the v8s misses), Civitai #1306938 for mouth, Civitai #2511165 for feet/shoes.

• GrimmRibbity — Comic Page (Regional) — single-gen multi-panel conditioning. Take a color-coded panel-layout mask + per-panel prompts (up to 6 panels) and emit one CONDITIONING constrained per region. Optional panel_strengths CSV override ("1.0, , 1.5") lets one panel dominate without changing every panel's binding. Pair with Character Anchor upstream for character lock across panels. No third-party node packs required — uses only ComfyUI's core CLIPTextEncode + ConditioningSetMask.

Output

• GrimmRibbity — Save Image (Civitai) — SaveImage replacement that writes A1111/Civitai-compatible PNG metadata. Auto-detects model, LoRAs, positive/negative, seed, sampler, scheduler from the workflow trace. Override any field if auto-detect picks the wrong sampler in multi-KSampler workflows
• GrimmRibbity — Thumbnail Saver — writes library thumbnails on workflow runs

Sampling (optional, requires torch)

• GrimmRibbity — SDXL Sampler + Pack SDXL Tuple — SDXL_TUPLE-driven sampler. Wires the way efficiency-nodes' SDXL_TUPLE does (drop-in compatible). Exposes denoise for img2img, optional script input for HiResFix.
• GrimmRibbity — Anima Sampler — KSampler-shaped sampler for Qwen / Flux / SD3 / any flow-matching base. Pre-encoded CONDITIONING + MODEL + LATENT directly, no SDXL-specific tuple.
• GrimmRibbity — HiResFix Script (SDXL) — latent / pixel / both upscale, optional checkpoint swap, per-iteration ControlNet anchoring, per-field hires prompt overrides, 1–5 iterations. Pixel upscale model + ControlNet load once per run instead of once per iteration; hires checkpoint cached across workflow re-runs (1-slot LRU). Auto-tiles VAE decode for outputs >1536px.
• GrimmRibbity — Anima HiResFix Script — shape-agnostic latent upscale (works on 5D Qwen latents). Per-iteration ProgressBar, refinement-only mode at scale=1.0, same auto-tile decode.

Upscale

• GrimmRibbity — Upscale SDXL (new in v0.57.1) — single-node tile-diffusion upscaler. Pre-upscales the input (via upscale_model if wired, otherwise Lanczos to upscale_by), optionally applies an SDXL tile ControlNet to the conditioning, then walks the canvas tile-by-tile with a low-denoise sampler and feather-blends each tile back in. Defaults match the Magnific Precise V2 / Clarity recipe (denoise 0.25, CFG 4.0, CN strength 0.5, 512 px tiles). bypass=True skips the diffusion pass and returns the upscaled image directly. Replaces the LoadControlNet + ApplyControlNet + UltimateSDUpscale chain with one node, all knobs exposed.

Both samplers have an opt-in save_prompt_log toggle (default off) — when enabled, every queued sample appends one JSONL line to prompt_log_path (empty = <output>/prompt_logs/prompts.jsonl) with the positive + negative prompts, the LoRA stack walked from the workflow trace, the model, and the sampler params (seed/steps/cfg/sampler_name/scheduler). One growing file per overnight batch — jq-friendly, no per-image directory pollution.

LoRA picker

• GrimmRibbity — LoRA Picker — dropdown picker that emits a <lora:path:weight> token for downstream parsing (rgthree Power LoRA Loader / A1111-style)

Every entry can store an optional negative field alongside its positive text. The Library, Save, and Random nodes all flow it on a second STRING output. Wire that into the negative side of your sampler. Backward-compatible — entries without a negative field load as empty.

The Library node embeds a full gallery in the node body:

• Click a thumbnail to toggle it in the selection. The output is the joined text of every selected entry, separated by the node's separator input (default , ).

• Shift-click range-extends the selection from the last-clicked tile.

• The same selection drives the bulk bar at the top — with one or more selected you get Tag / Export / Delete.

• Right-click a tile for Edit / Duplicate / Export this / Delete.

  

• Toolbar ▦ / ≡ switches between thumbnail grid and single-column list view (the list is good for dense libraries — 56 px thumb on the left, full name on the right).

  

• Drag-and-drop tiles to reorder when sort mode is Manual.

• Keyboard: arrow keys move focus, Enter toggles selection on the focused tile, Delete removes, / focuses search, Esc clears the selection.

Tags use category:value syntax (e.g. style:cyberpunk, character:elf). The chip row groups them by category. Chips OR-filter the grid; combine with the search box for AND.

model:* tags are special — they're lifted out of the chip row into the dedicated dropdown.

Right-click → Edit, or click the + tile to add. The modal is non-blocking — drag it around, the canvas stays interactive. Fields:

• Name — human-readable label

• ID — auto-derived from the name (Cyberpunk Style → cyberpunk_style); override only if you want a specific filesystem name. Collisions auto-append _2, _3, …

• Tags — comma-separated; autocompletes from existing tags via a <datalist>

• Prompt text — the actual prompt string

• Negative prompt — optional, flowed on the second STRING output of the loader nodes

• Rating — 0-5 stars, drives the favorites filter

• Notes — private; not used in generation; searchable

• Reference image — optional thumbnail; PNG/JPG/WebP/GIF/BMP, capped at 16 MB. Paste an image with Ctrl+V while the modal is focused.

• LoRAs — up to 10 per entry. Click + Add LoRA to attach a row (Model dropdown / M+C strength sliders with link toggle / Trigger words / Delete). Each row has independent strength_model and strength_clip controls; the 🔗 / 🔓 button locks them to move together. Soft-delete: the Delete button toggles to Restore until you Save the modal. Missing LoRA files are flagged with a visible warning row above the dropdown. Consumed by the Library node (when MODEL+CLIP are wired) and the Style node (which also encodes prompts on the patched CLIP). Trigger words are appended after the prompt before encoding.

  

• History — disclosure showing every prior version of this entry (max 20). Click any row to revert. Tracks LoRA-only changes too.

  

A toolbar ⚠ N badge appears when the validator finds unhealed issues. Click it for a categorised modal:

• Broken LoRA references — entries whose loras list points at a .safetensors no longer in models/loras/. Per-row Edit jumps into the entry.
• Orphan thumbnails — image files in data/images/ with no matching entry. One-click bulk cleanup.
• Invalid entry IDs — would refuse to upsert.
• Empty-text entries — half-completed saves.

The badge auto-refreshes after every gallery refresh; clean libraries get no UI clutter (badge hidden at zero issues).

Inside any prompt text:

• {red|green|blue} — picks one at random per generation
• __character__ — resolves to a random library entry whose id matches character, or whose tags contain character. Recurse-safe (max depth 8); cycles bottom out as literals.

For a fully randomized overnight loop, chain three Random by Tag nodes:

[Random by Tag: character]  ─┐
[Random by Tag: background] ─┼─ concat ─→ CLIP Text Encode ─→ KSampler
[Random by Tag: action]      ─┘

Each node has a seed input — set control_after_generate=randomize to draw a fresh combination every queue. Wildcards in the picked prompts (old {grumpy|wise} wizard) re-roll per generation.

The Wildcard Expand node has independent toggles for expand_choices and expand_named_refs if you want one without the other.

Everything lives in ComfyUI-GrimmRibbity/data/:

data/
├── prompts.json           # entries: id, name, text, tags, history, timestamps
├── images/<id>.<ext>      # thumbnails
└── .last_version          # auto-backup marker

Both prompts.json and images/ are gitignored. Back up the whole data/ folder to keep your library safe across reinstalls.

External edits to prompts.json are picked up automatically — the gallery polls the file every 2 s and refreshes on change. So you can version-control your library in git or edit it in your favourite text editor.

• Export some prompts: filter the gallery (search, tags, model, anything), click Export. You get a zip with prompts.json + thumbnails.
• Send the zip to a friend.
• They click Import in their gallery and pick the zip. Existing entries with the same id get history-bumped before being overwritten; new ones are appended.

Same flow works for backups: export the whole library to a zip and stash it.

For seeding from a spreadsheet, save as CSV with these columns (header row required):

• name is required
• text is the prompt body
• tags use ; as the in-cell separator (because , is the CSV delimiter)
• id is optional; blank means auto-derive from name

Five scripts in tools/ run independently of a live ComfyUI server. Useful for batch maintenance over an SSH session or as cron jobs.

# Diagnose library issues — reports broken LoRA refs / orphan thumbnails /
# invalid ids / empty-text entries. Read-only by default; --fix-orphans
# bulk-deletes the stray thumbnails.
python3 tools/library_validate.py [--fix-orphans] [--quiet]

# Backfill A1111/Civitai-compatible 'Hashes:' metadata on PNGs from before
# the Civitai Save node was wired in. Walks a folder, reads each PNG's
# embedded workflow trace, computes SHA256s for the model + LoRAs, and
# rewrites the parameters chunk. Idempotent — already-tagged PNGs skip.
python3 tools/civitai_backfill.py /path/to/output [--recursive] [--dry-run]

# Repair Smart Detailer widget-value shifts in workflow JSONs saved against
# an older version of the node. Walks every GrimmRibbitySmartDetailer node,
# type-coerces salvageable values into the current widget order, and falls
# back to defaults for unrecoverable slots. Original kept as <file>.bak.
python3 tools/fix_workflow_widgets.py <workflow.json> [-o]

# Live monitor for ComfyUI process — polls /queue, samples RSS / CPU / VRAM
# / GPU temp every 0.5s, prints per-run summary with peak RSS delta and
# VRAM peak when the queue empties. Run alongside ComfyUI for the bug-and-
# monitor matrix on long Smart Detailer batches.
python3 tools/detailer_monitor.py [--interval 0.5] [--port 8188]

# Capture deterministic gallery + modal screenshots via headless Firefox
# (useful for refreshing the docs after a UI change).
python3 tools/screenshots/run.py

The ComfyUI server gains 17 routes under /prompt_library/*. Highlights:

• GET /list / POST /upsert / POST /delete / POST /bulk_delete — CRUD over entries
• POST /export / POST /import_zip / POST /import_csv — bulk transfer
• GET /validate / POST /fix_orphans — library health check
• GET /loras — installed LoRA filenames (drives the modal dropdown)
• GET /history/{id} / POST /revert — per-entry history
• POST /scan_loras — bulk-import LoRAs as library entries
• POST /restore_snapshot / GET /snapshots — recover from a destructive bulk op

Bodies are validated; size caps applied (16 MB image, 50 MB CSV, 500 MB ZIP compressed / 2 GB uncompressed). All POST routes accept JSON; all return JSON. The export route returns a zip body with Content-Disposition: attachment.

• ComfyUI any reasonably recent version (tested with 0.19.x, frontend 1.42.x)
• Python 3.10+
• No third-party deps beyond what ComfyUI already ships (aiohttp, numpy, Pillow)

git clone https://github.com/Deaththegrim/ComfyUI-PromptLibrary ComfyUI-GrimmRibbity
cd ComfyUI-GrimmRibbity
python3 -m venv .testenv
.testenv/bin/pip install aiohttp pillow
.testenv/bin/python -m unittest discover tests

322 tests, runs in ~3 s. The Comic Page tests + the Smart Detailer + Style node helper tests skip without torch installed (.testenv doesn't ship it). The Set/Get compatibility tests skip when comfy.* isn't on PYTHONPATH.

Both PromptLibrarySave and PromptLibraryThumbnailSaver log their target on every run:

[PromptLibrary] saved id='X' name='X' via=prompt_id ...        ← targeting by stuck id
[PromptLibrary] saved id='X' name='X' via=overwrite_by_name ... ← name-match fired
[PromptLibrary] saved id='X' name='X' via=new ...              ← new entry created
[ThumbnailSaver] wrote thumbnail to id='X' name='Y'

If via=prompt_id fires when you expected overwrite_by_name, the Save node has a prompt_id widget value pinned in the workflow JSON (a one-time test value, an accidental wire, etc.) — clear that field and overwrite_by_name will fire.

When multiple entries share a name, overwrite_by_name targets the most-recently-edited match (since v0.46.1).

The Library node's MODEL + CLIP inputs are both-or-neither for LoRA application. Wiring just one logs:

[PromptLibrary] WARNING: only one of MODEL / CLIP is wired — LoRA application requires BOTH.

Wire the matching socket or unwire both.

Click the ⚠ N badge in the gallery toolbar (or run python3 tools/library_validate.py from the shell). Lists broken LoRA refs / orphan thumbnails / invalid ids / empty-text rows.

ComfyUI loads workflow widget values by position. When new required widgets are added to a node across versions, saved values shift and end up in the wrong widget — symptom: validation errors saying things like sampler_name='karras' (a scheduler value) or max_size=1 (was a boolean).

Two fixes:

• Re-add the node: right-click the Smart Detailer → Remove → re-add fresh. The new node has correct defaults.
• Migrate the saved workflow JSON: python3 tools/fix_workflow_widgets.py <workflow.json> -o. Walks every Smart Detailer node, type-coerces salvageable values into the current widget order, falls back to defaults for unrecoverable slots. Original is preserved as <workflow.json>.bak.

Going forward, new widgets land in optional so they don't shift required-widget positions — saved workflows from v0.53.1+ should remain stable.

The detail pass pipeline used to skip a NaN scrub between VAE decode and the alpha-blend composite, so a sample-time NaN propagated into the final image and downstream clip(0,1).astype(uint8) casts produced black pixels. Fixed in v0.49.0 — nan_to_num runs after every refined-image path. If you still see this on v0.49.0+, share the console log; there's a different bug.

See CREDITS.md for the full list. Short version: the suite is co-created by Deaththegrim and RibbityRabbit, pair-programmed with Claude Code. The bundled tag library is a community contribution — curated by AceVanquish, Drow, and the rest of the contributors.

If you contributed tags or want to, head to the GitHub repo and open a PR or issue.

If you fork, remix, or publish a derivative, a credit line back to Deaththegrim/ComfyUI-PromptLibrary is appreciated.

MIT — see LICENSE. You're free to use, modify, redistribute, and include this in commercial work. The one ask: keep the copyright notice ("Copyright (c) 2026 Deaththegrim and RibbityRabbit") intact in copies and substantial portions.