Build HTTP headers for Ollama or OpenAI-compatible prompt enhancement. For OpenAI / vLLM mode: - Reads OPENAI_API_KEY from environment (never

Amber pulsing "waiting" state on the Enhance button (Looper pattern).

* Toggle between stacked (vertical) and split (two-column landscape) * layouts. Split puts prompting on the left, slots/advanced/LLM on the

Add generation-limit options to an OpenAI-compatible payload. - OpenAI GPT-5 / o-series: max_completion_tokens, omit temperature. - Older Ope

Resize preserving aspect so long edge <= max_size, snap to stride.

* Collect base64-encoded reference images from connected nodes. * Traces the graph links from this node's reference_image_N inputs * backwar

* Extract frames from the connected source video for LLM vision. * Traces source_video input to find VHS_LoadVideo (or similar), * gets the

True for OpenAI chat models that reject the legacy max_tokens field. GPT-5 and reasoning-family models on /v1/chat/completions expect max_com

Load an image from ComfyUI's input directory as a [1,H,W,3] float tensor. Used for the built-in reference slots (drag-and-drop in the editor UI).

Server-side prompt enhancement via Ollama/vLLM. Called during execute() when auto_enhance is True. Uses urllib (stdlib) to avoid async complic

* Re-sync all DOM elements from their underlying widget values. * Called from onConfigure after ComfyUI restores saved widget values * (whic

* Walk backwards through the graph from a given node, following IMAGE * inputs, until we find a node with an "image" widget (LoadImage, etc).

Immediately unload the selected Ollama model from VRAM.

Extract frames from a video in ComfyUI's input folder and return as base64 JPEG. Used to send source video frames to the LLM for vision-aware

Return the full enhancement template for a given task type.

Return task metadata for the JS editor.

Immediately unload the Ollama model from VRAM via keep_alive=0.