4A Prompt Manager

ComfyUI-4A-Prompt-Manager

中文说明

Companion NovelAI generation plugin: ComfyUI-4A-NovelAI

Chinese video guide (Bilibili): https://www.bilibili.com/video/BV1DWKv6NE6t
Update #1 (Bilibili): https://www.bilibili.com/video/BV1Ccg86nErh

Folder-backed prompt library and scheduler for ComfyUI: browse/edit wildcards and JSON cards (LoRAs + sparse generation settings), assemble multi-track prompts, auto-apply settings on Wildcard batch runs, reuse image metadata, and save generations with portable parameters.

Current release: 1.4.3

• Caches Wildcard leaf sizes so large folder __key__ expands stay fast after the first count

1.4.2

• Image Saver writes prompt, parameter, and hash metadata only from connected or filled inputs; unconnected optional inputs no longer add default or empty structures

1.4.1

• Caches the Wildcard library snapshot so official Queue no longer rebuilds the whole library on every click; plain text skips the snapshot entirely
• Remembers Browser folder, view, and send mode per workflow instead of sharing one global UI state

1.4.0

• Adds NovelAI (NAI) Browser mode, sparse nai card fields, NAI image metadata reading, and sending prompts/settings to external NAI nodes
• Moves generation nodes into standalone ComfyUI-4A-NovelAI; Prompt Manager still works without that plugin
• Converts Comfy / NAI weights only when the active mode has no native text; detail view, copy, send, and Wildcard expansion share the same rules
• Flattens NAI characters into charN: lines when a card is used in Comfy
• Writes edits back to the current mode only, without overwriting the other mode's native text
• Persists Meta Apply / Meta Loader previews via /view after upload so they survive workflow switches and restarts

Highlights

Prompt Manager frontend

A full, easy-to-use library UI. Prompts live as a normal folder tree (JSON cards + TXT wildcards), so you can back up or share by copying folders; import/export freely; generate preview images for cards (uses a built-in API workflow, or replace it with your own api.json). The last folder and sidebar expansion are remembered across reloads. After a library refresh you can align card model names to local files by hash.

JSON card LoRAs / generation settings + wildcard auto-apply

JSON cards can sparsely store LoRAs, models, sampler settings (including seed/size), and double-sample settings. Empty fields are omitted from JSON; presence of a double-sample block means second pass ON (no separate enable flag). When Scheduler tracks resolve those cards via wildcard syntax, enable Auto-embed Wildcard LoRAs and/or Auto-apply models / inference parameters to write matching nodes before queueing, then restore the canvas baseline (seed is not restored). With auto-embed on, Group same models / LoRAs queues jobs that share the same stack together to reduce reloads. Conflict rules: LoRAs stack (skip same name); everything else is first-wins per field. Bypass Switch follows whether the card has double-sample fields—wire the second-pass subgraph/parameters into it (one switch per workflow for now). Dropping an image when creating a card still reads prompts only by default; use Load generation settings from image (with overwrite confirm) for models/params/size/LoRA. Detail view can also push models / LoRA / parameters into the canvas in one click.

Multi-track Prompt Scheduler

Stack positives across tracks with random / sequence / shuffle, Impact-compatible wildcard parsing (__key__, {a|b}, weights, multi-select, folder and global name lookup), and optional STRING inputs wired into each track from the graph.

Nested wildcards (great for automation): JSON content / negative and TXT bodies may contain further __folder__ / __path/file__. Expansion resolves recursively and carries nested LoRAs, negatives, and sparse generation settings. Sequential batching uses a leaf space: nested branches sum, side-by-side __a__, __b__ in one track multiply, multiple tracks take the longest; {a|b} and friends stay random and do not lengthen the cycle. A common pattern is “outer cards = LoRAs/presets to compare; inner __scenes__ / suites = shared prompts”—Count tasks fills the job count from that leaf space, not just the top-level folder size.

Full rules and examples: docs/wildcard-nesting.md (中文).

NovelAI (NAI) prompt and metadata management

NAI generation nodes have moved to the standalone ComfyUI-4A-NovelAI plugin. This project retains prompt-card storage, the Wildcard library, weight-display conversion, image-metadata reading/apply, and the bridge that sends prompts and sparse settings to NAI nodes. PM4A still works when the generation plugin is not installed.

• JSON cards keep generic content / negative separate from native nai.content / nai.negative. Browser prefers the active mode's native text and only converts the other representation as a temporary fallback; edits are saved back to the active mode so {}, [], and N::text:: are not misinterpreted. The detail labels mark whether the visible text is native or a converted preview.
• nai.parameters stores only sparse steps, cfg, sampler, scheduler, cfg_rescale, width, and height. Ordered nai.characters store each character's positive/negative text plus optional per-character x/y; model and generation seed remain on the standalone generation nodes.
• PM4A continues to read NAI positive/negative prompts, character prompts and coordinates, plus the sampler and size fields from images and can apply them to the standalone Scheduler / Sampler.
• NAI-only Wildcard expansion, multi-character scheduling, free-form positioning, serial batch, model/account usage, references, and txt2img/img2img/inpaint switching through a dedicated image-input node belong to the generation plugin.

Metadata reuse (Meta Loader & Meta Apply)

Read prompts and params from common image formats. Drop images from the Browser UI, outside the app, or ComfyUI assets; one click to reuse embedded prompts, sampler settings, and LoRA stacks (LoRA apply needs Lora Manager). Meta Apply has separate toggles for model, LoRA, inference parameters, and prompt. Applying parameters also syncs Bypass Switch from whether double-sample fields are present.

Input Parameters & Image Saver

Simple wiring for sampler / resolution outputs; save PNG/JPEG/WebP with A1111-style metadata so your prompts and params stay portable and reusable (JPEG/WebP need piexif).

Install

ComfyUI-Manager (recommended)

Search for 4A Prompt Manager / ComfyUI-4A-Prompt-Manager and install. Dependencies from requirements.txt / install.py are handled by Manager.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/tsukino4a/ComfyUI-4A-Prompt-Manager.git ComfyUI-4A-Prompt-Manager
cd ComfyUI-4A-Prompt-Manager
python install.py
# or: pip install -r requirements.txt

Restart ComfyUI after install.

Quick start

1. Open Workflow → Browse Templates → ComfyUI-4A-Prompt-Manager and try:
   • 01_single_sampler_workflow — Input Parameters → KSampler → Image Saver
   • 02_double_sampler_workflow — Double Sample Parameters + second pass
2. Open a Prompt Manager Browser node (or the top-menu browser). You should see shipped samples under wildcards/examples/ (JSON cards + TXT wildcards).
3. To learn batch import, use examples/pm4a_examples_bundle.json — the same pm4a-prompt-bundle shape produced by library export (mixed json + txt entries). Import it from the Browser import dialog; do not put this file inside wildcards/ (it is not a single prompt card).

Nodes

| Node | Role | |------|------| | Prompt Manager Browser | Full library UI in a node | | Prompt Scheduler | Build positive/negative strings from tracks | | Meta Loader (Prompt Display) | Inspect image metadata cards; apply LoRA text when Lora Manager is present | | Meta Apply | Auto-apply image metadata to linked targets (including LoRA via Lora Manager) | | Input Parameters | Seed / steps / cfg / sampler / size + JSON | | Double Sample Parameters | Second-pass sampler JSON | | Bypass Switch | Wire-controlled Bypass/Always for connected nodes; ON when the card has double-sample fields | | Image Saver | Save images with hashes & A1111 params |

Example workflows & library samples

| Path | Purpose | |------|---------| | example_workflows/ | UI templates for Browse Templates (+ .png previews) | | wildcards/examples/ | Shipped prompt samples (loaded by the library) | | examples/pm4a_examples_bundle.json | Export/import demo bundle | | workflows/default_api.json | Built-in UNet API graph for in-browser preview generation (replaceable) | | docs/wildcard-nesting.md | Nested wildcards / sequential leaf-space guide (中文) |

Select models in template workflows before running if loaders are empty or point to names you do not have locally.

Dependencies

• Required extra: piexif (>=1.1.3) for JPEG/WebP EXIF when using Image Saver / preview EXIF
• Pillow, NumPy, aiohttp are provided by ComfyUI
• Optional NAI generation plugin: ComfyUI-4A-NovelAI; PM4A no longer registers NAI generation nodes

Without piexif, PNG save still works; JPEG/WebP metadata write will error (preview WebP may skip EXIF silently).

License

This project is released under the MIT License.

Image Saver adaptations from ComfyUI-Image-Saver are documented in THIRD_PARTY_NOTICES.md.