⚙️ Configuration¶

Talk2Scene uses Hydra for hierarchical configuration.

📁 Config Groups¶

Group	File	Description
🧠 model	`conf/model/default.yaml`	Whisper and LLM settings
📡 stream	`conf/stream/default.yaml`	Redis stream settings
🖼️ render	`conf/render/default.yaml`	Canvas, render, and video settings
🎨 assets	`conf/assets/default.yaml`	Asset paths and z-order
👤 character	`conf/character/default.yaml`	Character defaults and transitions
📂 io	`conf/io/default.yaml`	Input/output paths and formats

Default model is gpt-4o with JSON mode enabled (response_format: json_object). This guarantees valid JSON output from the scene generator.

Setting	Default	Description
`model.llm.model`	`gpt-4o`	OpenAI model (must support JSON mode)
`model.llm.temperature`	`0.3`	Lower = more deterministic scene codes
`model.llm.max_tokens`	`4096`	Max tokens for scene generation response

Override the model via CLI:

uv run talk2scene mode=text io.input.text_file=input/transcript.jsonl model.llm.model=gpt-4o-mini

Setting	Default	Description
`stream.redis.stream_key`	`stream:mic`	Raw audio stream key
`stream.redis.stt_stream_key`	`stream:stt`	Pre-transcribed text stream key (higher priority)
`stream.redis.consumer_group`	`talk2scene`	Redis consumer group name
`stream.redis.consumer_name`	`worker-1`	Consumer name within the group
`stream.redis.block_ms`	`1000`	Block timeout for XREADGROUP
`stream.redis.batch_size`	`10`	Max messages per read
`stream.redis.backpressure_max`	`100`	Max pending messages before pausing

Setting	Default	Description
`render.canvas.width`	`1024`	Canvas width in pixels
`render.canvas.height`	`1024`	Canvas height in pixels
`render.scene_on_event`	`false`	Render `front_page.png` on each scene event batch (stream mode)
`render.video.fps`	`30`	Video output frame rate
`render.video.crf`	`18`	Constant rate factor (lower = higher quality)
`render.video.format`	`webm`	Video format: `webm`, `mp4`, or `avi`
`render.video.subtitle`	`true`	Burn subtitles into video
`render.video.subtitle_font_size`	`32`	Subtitle font size in pixels
`render.video.preview`	`true`	Open video after rendering

Hydra supports dot-notation overrides:

uv run talk2scene model.whisper.model_size=medium stream.redis.host=myhost