Find all node IDs in a workflow matching a given class_type.

Snap a value to the nearest multiple of step.

Find a node ID by its _meta title.

Verify ComfyUI is reachable.

Open WebSocket to ComfyUI for execution tracking.

Load the API-format workflow template JSON.

Submit a workflow to ComfyUI. Returns the prompt_id.

Block on WebSocket until execution completes or errors. Falls back to polling /history if the socket drops. Returns the history dict

Call ollama.chat with automatic token scaling on truncation. Detects truncated JSON output (unbalanced brackets) and retries with doubled tok

Parse JSON from Gemma output, retrying on failure. Handles both formats: - Plain array: [scene1, scene2, ...] - Dict with characters: {"c

Clone the template and inject per-scene values. Only touches prompt text, frame count, and seeds. All other settings (sampler, sigmas, CFG, m

Calculate frame count obeying LTX rule: frames = (8n + 1).

Calculate scene duration from dialogue word count + action time. Returns duration in seconds, clamped to SCENE_MIN/MAX.

Extract frames at sample rate, return list of base64 JPEG strings.

Extract the video file path from a completed history dict.

Verify everything is ready before starting. Auto-launches ComfyUI if needed.

Guess speaking rate from the brief content.

Run a round of keyframe generation attempts. Returns True if one passed.

Multi-page first-run wizard with setup instructions.

Clone the i2v template and inject per-scene values + keyframe image. image_filename should be the name returned by upload_image() (already in Com

Get the character descriptions from the most recent breakdown.

Get the style anchor from the most recent breakdown.

Get the voice descriptions from the most recent breakdown.

Load the i2v API-format workflow template JSON.

Full autonomous pipeline.

Turn a brief into a filesystem-safe project name.

Upload an image to ComfyUI's input directory. Returns the filename.

Auto-detect i2v workflow node IDs from the template.

Auto-detect keyframe workflow node IDs from the template.

Auto-detect video workflow node IDs from the template. Returns a dict with keys: prompt, neg_prompt, frames, seed1, seed2, resolution. Falls

Programmatically enforce FAIL conditions.

Programmatically enforce FAIL conditions even if the model said PASS.

Extract target duration in seconds from user brief, if specified.

Fix common LLM JSON issues. Handles all known failure patterns from Gemma.

Detect if the input looks like a screenplay/script rather than a brief.

Parse eval JSON, handling markdown fences and stray text.

Open video in system default player.

Check if prompt already finished (fallback when WS is unreliable).

Kill and restart Ollama to ensure clean state.

Ask the LLM to rewrite the image prompt based on what kept failing.

Coerce scene fields to correct types. LLMs sometimes output numbers as strings.

Select a take for a scene.

Try to parse JSON with progressively more aggressive fixes.

Unload the heavy creative model from VRAM to free it for ComfyUI.

Build a keyframe image workflow with the given prompt and seed. Auto-detects node IDs if the configured defaults don't match the template.

Concatenate scene clips using FFmpeg concat demuxer (no re-encode).

Thoroughly evaluate a keyframe image against the scene + character descriptions. Returns dict with scores and verdict.

Generate keyframe candidates for a scene with prompt rewriting on failure. Round 1: Try KF_CANDIDATES attempts with the original prompt. If a

Extract image path from ComfyUI history.

Extract a single frame from a video as a PIL Image.

Read an image file and return base64 string.

Launch ComfyUI in the background and wait until it's reachable.

Parse a screenplay/script into scenes using Gemma.

Generate and evaluate a single test scene.

Write an LTX 2.3 prompt for a single scene.

Assemble selected takes into final film.

Crop a square image to the video's widescreen aspect ratio (center crop).

Generate new takes for a scene in a background thread.