comfyui-coras-textgen-nodes

Cora's Textgen — a ComfyUI custom-node pack with three nodes: Prompt Enhancer and Recaption send your text (or an image) plus a style-driven system prompt to an OpenAI-compatible /chat/completions endpoint and output the reply with <think> reasoning blocks stripped; Refine Text lets you manually edit a captured prompt before re-running downstream nodes.

• Targets the Nodes 2.0 (V3) API.
• Endpoint and API key are set in the ComfyUI Settings panel (group: Cora's Textgen); system prompts come from per-node style dropdowns backed by user-editable YAML files.
• Works with any OpenAI-compatible server: OpenAI, Ollama, LM Studio, llama.cpp, vLLM, etc.

Install

Clone (or copy) this folder into your ComfyUI custom_nodes directory and restart ComfyUI:

ComfyUI/custom_nodes/comfyui-coras-textgen-nodes

requests is the only runtime dependency and is already bundled with ComfyUI; requirements.txt lists it for standalone use.

Configuration

Open Settings → Cora's Textgen and set:

| Setting | Default | Notes | |---|---|---| | API Base URL | https://api.openai.com/v1 | The node calls <base>/chat/completions. | | API Key | (empty) | Required. Stored locally, never in workflow files. | | Temperature | 0.7 | | | Request Timeout (s) | 60 | | | Filter Vision-Capable Models (Recaption) | true | Substring heuristic on /models for the Recaption dropdown. Disable to see every model. |

The API key has no default. A ComfyUI setting is only written to disk after you change it once, so you must enter the key in Settings (or provide CORAS_TEXTGEN_API_KEY) — otherwise the node fails with a clear message.

Environment-variable alternative (headless / Docker)

Each setting has an env fallback, used when the setting is unset: CORAS_TEXTGEN_API_BASE, CORAS_TEXTGEN_API_KEY, CORAS_TEXTGEN_TEMPERATURE, CORAS_TEXTGEN_TIMEOUT, CORAS_TEXTGEN_RECAPTION_FILTER_VISION.

Precedence (highest first): comfy.settings.json → environment variable → built-in default. The model and system_prompt per chat call come from on-node selections (model dropdown + style YAML) and are not configured in Settings.

The model dropdowns talk to /coras_textgen/models (and /coras_textgen/models/vision for Recaption), which proxy <api_base>/models.

The nodes

Prompt Enhancer

Inputs

• text (required, multiline) — the draft prompt to be rewritten.
• style (dropdown) — populated from YAML files in <user>/default/coras_textgen/prompts/prompt_enhancer/. Each file has name (the display label) and system_prompt (the instruction sent to the model). On first launch the extension seeds stable_diffusion.yml (tag-style rewriter) and flux.yml (natural-language rewriter); existing files are never clobbered, so edits and additions survive restarts.
• model (dropdown) — populated by calling <api_base>/models on the configured endpoint.

Output

• text — the assistant reply with reasoning removed.

Recaption

Inputs

• image (required) — only the first frame of a batched image is captioned.
• style (dropdown) — populated from YAML files in <user>/default/coras_textgen/prompts/recaption/. Each file has name (the display label) and system_prompt (sent to the model). On first launch the extension seeds stable_diffusion.yml and flux.yml with tag-style and natural-language captioners respectively; existing files are never clobbered, so edits and additions survive restarts.
• model (dropdown) — populated from <api_base>/models, then filtered by a name heuristic (vision, vl, llava, gpt-4o, gemma, pixtral, molmo, claude, minicpm, florence, moondream, multimodal). If the heuristic filters out every model the endpoint advertised, the unfiltered list is shown instead so the dropdown is never empty; toggle Filter Vision-Capable Models off in Settings to skip the heuristic entirely.

Output

• text — the assistant reply with reasoning removed.

The image is encoded as a base64 PNG and sent as an OpenAI-compatible vision message (image_url content part). Works against any endpoint that implements the vision content array — OpenAI, LM Studio (with a VL model), LocalAI, vLLM, etc.

Refine Text

A passthrough-with-override node for the workflow where you want to rerun downstream nodes with a manually tweaked version of an LLM-generated prompt without rerunning the LLM.

Inputs

• text (socket, optional) — upstream string. Typically Prompt Enhancer's output, but anything string-typed works.
• lock (checkbox) — when off, the node passes the upstream input through and writes it into the editable widget below after each run; when on, the edited widget value is output and the upstream input is ignored.
• edited_text (multiline textarea) — the editable capture of the upstream text. Typing into it flips lock to on automatically, so editing and rerunning takes a single keystroke past the edit itself.

Output

• text — edited_text while locked, otherwise the upstream input.

Typical workflow

1. Wire Prompt Enhancer → Refine Text → <downstream>. lock starts off.
2. Run — the enhanced prompt appears in edited_text and flows downstream.
3. Edit edited_text. lock flips on. Run again — downstream sees your edit; the Prompt Enhancer call doesn't re-fire (ComfyUI caches it since its inputs are unchanged).
4. Untick lock to resync from the upstream LLM output.

The lock state and edited text are saved into the workflow JSON, so reloading a workflow restores both.

Stale-lock warning. While lock is on, the node remembers the last upstream text it captured. If a subsequent run produces a different upstream value (e.g. you tweaked Prompt Enhancer's style or model and reran without unticking lock), the node title bar turns orange and a warning is logged to the browser console. Untick lock to resync from upstream and clear the warning. The warning color is not persisted into saved workflows.

<think> stripping

The output has reasoning content removed, handling the common variants:

• Balanced <think>…</think> blocks (any number, case-insensitive).
• A dangling </think> with no opening tag (some servers emit reasoning, then </think>, then the answer) — only the text after the last </think> is kept.
• An unclosed <think> (truncated reasoning) — everything from <think> onward is dropped.

Text with no think tags passes through unchanged (whitespace-trimmed).

Caveats

• The backend reads ComfyUI's persisted settings file (user/default/comfy.settings.json) directly — there is no Python settings API. Reads are defensive (missing/empty/bad file → env → defaults).
• Multi-user ComfyUI stores settings under user/<name>/…; only the default user (plus env vars) is supported initially.
• V3 only — requires a ComfyUI build that ships comfy_api.latest.

Development

python3 -m venv .venv
. .venv/bin/activate
pip install requests pytest pyyaml
pytest -q

Tests cover strip_think (all tag variants), resolve_config precedence, the vision-model name heuristic, and the YAML-backed prompt-style loader; none require a running ComfyUI.

The Refine Text node's lock state machine has its own JS unit tests using Node's built-in test runner (no extra deps):

node --test tests/test_refine_text_state.mjs

The reducers in web/js/refine_text_state.js are pure (no browser or ComfyUI references) so the spec is testable in isolation; the DOM glue in web/js/coras_refine_text.js is a thin wrapper around them.

Publishing

Before publishing to the ComfyUI Registry, set PublisherId and the repository URL in pyproject.toml (currently CHANGEME).

License

GPL-3.0-or-later. See LICENSE.