scg-prompt-forge

SCG Prompt Forge

A ComfyUI custom node for bulk creative prompt generation. Give it a seed idea ("guerrilla marketing street shots, think DKNY or American Eagle style advertising") and it pumps out batches of distinct, shaped prompts — built for LoRA testing workflows that need hundreds of quality prompts carrying specific trigger keywords (e.g. hex color codes).

It opens a full-screen forge over the ComfyUI canvas where you can:

• Describe the seed idea you're chasing, plus optional things to avoid.
• Optionally add a reference image (click, drag in, or Ctrl+V from the clipboard). With no seed text, the image's scene becomes the springboard; with seed text, the seed stays in charge and can steer how the image is used (only its palette, only its subject, and so on).
• List required keywords (LoRA triggers, hex codes), either injected into every prompt or rotated one-per-prompt across the batch.
• Load wildcard .txt files (click or drag in — same format as scg-wildcards: one option per line, # comments ignored). Each file gets an editable keyword; every prompt draws its own random line. Reference a file as _name_ in the seed idea or required keywords to control where the value lands (keyword-line tokens resolve to the drawn value, so the verbatim check still applies); unreferenced files are still woven into every prompt.
• Pick an output shape from a dropdown — paragraph description, multi-paragraph photography-focused, booru/CSV tag list, Ideogram bbox JSON, Krea bbox JSON (same structure minus color hex codes, since Krea renders hex codes as literal text), simple JSON fields, YAML, markdown — or Custom, where you describe the shape in plain text and the agent runs with it.
• Generate batches of up to 500 prompts. The agent first expands the seed into a creative brief plus distinct concept angles, then writes prompts in chunks so every prompt gets its own angle instead of 200 near-duplicates.
• Review the batch: edit any prompt in place, delete, reroll individual prompts, or Generate more to append. Prompts missing a required keyword are flagged.
• Save to node and step through the batch in your workflow, Copy all to the clipboard, or download the batch as a file:
  • .txt — UTF-8, prompts separated by === PROMPT n/N === lines, so the file stays splittable (/^=== PROMPT \d+\/\d+ ===$/m) no matter what shape the prompts are.
  • .jsonl — UTF-8, one JSON-encoded prompt string per line, bulletproof for scripts.

Node outputs

| Output | Type | Notes | | --- | --- | --- | | prompt | STRING | prompts[index % count] — the current prompt. | | index | INT | The effective (wrapped) index that was emitted. | | count | INT | Total prompts in the saved batch. |

The index widget defaults its control to increment: queue N runs and the node walks the batch one prompt per run, wrapping around at the end. Set the control to fixed to pin a single prompt.

The node also has a Copy Batch button that copies all saved prompts to the clipboard (blank-line separated).

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/SanDiegoDude/scg-prompt-forge.git scg_prompt_forge

Then restart ComfyUI (the plugin registers two small server routes at startup). The node appears as SCG Prompt Forge under SCG/PromptForge.

Dependencies

For the common case (OpenAI-compatible endpoints like LM Studio, OpenAI, Grok, Gemini's OpenAI-compat URL), there is nothing to install — the agent calls are proxied through the ComfyUI server using its bundled aiohttp.

The only optional dependency is google-auth, which is required only if you configure a Vertex AI provider (see below):

pip install -r requirements.txt

Configuring AI providers (.env)

The agents talk to any OpenAI-compatible chat-completions endpoint (LM Studio, OpenAI, x.ai/Grok, Gemini's OpenAI-compat endpoint, etc.). Providers are declared in a local .env file so that API keys stay on the server and are never sent to the browser — the UI only ever sees each provider's id/label/model and asks the ComfyUI server to make the call on its behalf.

1. Copy the example file:

   cp .env.example .env
   

2. Add one provider per line:

   AI_PROVIDER_<ID> = Label | model | base_url (blank = official OpenAI) | api_key
   

   Examples:

   AI_PROVIDER_OPENAI = OpenAI (gpt-5.4-mini) | gpt-5.4-mini |  | sk-...
   AI_PROVIDER_GROK   = grok-4.3 | grok-4.3 | https://api.x.ai/v1 | xai-...
   AI_PROVIDER_GEMINI = Gemini 3.5 Flash | gemini-3.5-flash | https://generativelanguage.googleapis.com/v1beta/openai | ...
   AI_PROVIDER_LOCAL  = Local (qwen3) | qwen3-... | http://192.168.0.180:1234 | 123
   

   • Leave the base_url blank to use the official OpenAI endpoint.
   • A bare host (e.g. http://192.168.0.180:1234, LM Studio) automatically gets /v1 appended.
   • Comment a line out with # to hide that provider.

3. In the forge UI (Agent Configuration → Agent settings), pick your provider from the dropdown. After editing .env, click Refresh to reload providers live — no restart needed.

Zero-setup fallback: if this folder has no .env of its own, the plugin automatically reads the one in a sibling scg_json_prompt_agent install, so if you already use that bbox builder there is nothing to configure.

Security: .env is git-ignored. Never commit your keys. The proxy only exposes provider id/label/model to the browser; URLs and keys remain server-side.

Gemini via Vertex AI

Vertex providers reach Gemini through Vertex's OpenAI-compatible endpoint. They use a vertex://PROJECT/LOCATION base URL and authenticate with Google OAuth (no static key in the 4th field):

AI_PROVIDER_<ID> = Label | model | vertex://PROJECT/LOCATION | [path/to/service-account.json]

Replace PROJECT with your Google Cloud project id, and LOCATION with global (the global endpoint) or a region such as us-central1. Leave the 4th field blank to use Application Default Credentials, or set it to a service-account JSON path. The server adds the required google/ model prefix and fetches/refreshes the OAuth token for you, so nothing sensitive reaches the browser.

Setup, step by step:

1. Enable the API. In your Google Cloud project, enable the Vertex AI API and make sure billing is enabled.

2. Install the auth dependency into the same environment ComfyUI runs in:

   pip install -r requirements.txt
   

3. Authenticate — pick one:

   • Application Default Credentials (simplest). Install the gcloud CLI, then run:

     gcloud auth application-default login
     

     Leave the 4th .env field blank.

   • Service account. Create a service account with the Vertex AI User role, download its JSON key, and put the file path in the 4th .env field:

     AI_PROVIDER_GEMINI = Gemini Vertex | gemini-2.0-flash | vertex://my-project-id/global | /home/me/keys/vertex-sa.json
     

4. Add the provider line to .env (replace the placeholders with your own project, region, and model):

   AI_PROVIDER_GEMINI = Gemini Vertex | gemini-2.0-flash | vertex://my-project-id/us-central1 |
   

5. Restart ComfyUI once so the provider registers, then select it from the picker in Agent Configuration → Agent settings. (Later .env edits only need the Refresh button.)

Tip: Gemini "flash/thinking" models spend tokens on internal reasoning, so keep Max tokens generous (8k+) or replies can come back truncated/empty.

A note on OpenAI GPT-5 / o-series models

These reasoning models have a slightly different API surface. The server handles it for you: it sends reasoning_effort: "low", uses max_completion_tokens instead of max_tokens, and omits temperature (which those models reject). Other providers use the standard max_tokens + temperature.

How it works

• nodes.py — the SCG Prompt Forge node; steps through the saved batch.
• providers.py — parses .env and registers the two routes:
  • GET /scg_prompt_forge/providers — provider list (no secrets).
  • POST /scg_prompt_forge/chat — server-side proxy to the selected provider.
• web/prompt_forge.js — the parent extension: hides the batch widget, adds the Open Forge / Copy Batch buttons, and bridges the iframe to the node.
• web/Prompt_Forge.html — the forge UI (iframe over the ComfyUI canvas): batch setup, the two-stage chunked generation pipeline, the reviewable prompt list, and the Agent Configuration panel with editable rulesets.

The generation pipeline runs entirely in the overlay (through the chat proxy); nothing is queued on the ComfyUI graph until you run your own workflow with the saved batch.

License

MIT — see LICENSE.