Configuration

All configuration surfaces: environment variables, data files, Django settings, and Docker services.

Environment Variables

Set in .env at the project root. Variables with defaults shown can be omitted for local development.

Database

Variable	Default	Description
`POSTGRES_DB`	`cw`	Database name
`POSTGRES_USER`	`cw`	Database user
`POSTGRES_PASSWORD`	`cw_dev`	Database password
`POSTGRES_HOST`	`localhost`	Database host
`POSTGRES_PORT`	`5435`	Database port (non-standard to avoid conflicts)

Broker

Variable	Default	Description
`VALKEY_HOST`	`localhost`	Valkey/Redis host
`VALKEY_PORT`	`6379`	Valkey/Redis port

Application

Variable	Default	Description
`DJANGO_SECRET_KEY`	(required)	Django secret key for sessions and CSRF
`DJANGO_DEBUG`	`True`	Debug mode (set `False` in production)
`ANTHROPIC_API_KEY`	(empty)	Anthropic API key for LLM prompt enhancement
`CIVITAI_API_KEY`	(empty)	CivitAI API key for LoRA auto-downloads
`MODEL_BASE_PATH`	`{PROJECT_ROOT}/models`	Base directory for local `.safetensors` model files

Video Upload

Variable	Default	Description
`VIDEO_MAX_UPLOAD_SIZE_BYTES`	`524288000`	Maximum upload size (500 MB)

Presets Configuration

data/presets.json defines diffusion models and LoRAs. Synced to the database via import_presets.

Top-Level Structure

{
  "models": [ ... ],
  "loras": [ ... ]
}

Model Fields

Field	Type	Description
`slug`	string	Unique identifier (e.g., `flux1_dev`)
`label`	string	Display name
`path`	string	HuggingFace repo ID or local `.safetensors` path
`pipeline`	string	Diffusers pipeline class (e.g., `FluxPipeline`)
`base_architecture`	string	Model family: `flux1`, `sdxl`, `sd15`, `zimage`, `qwen`
`settings`	object	Generation parameters and behavior flags (see below)

Model Settings

Setting	Type	Description
`steps`	int	Default inference steps
`guidance_scale`	float	Default classifier-free guidance strength
`default_width`	int	Default image width (pixels)
`default_height`	int	Default image height (pixels)
`max_pixels`	int	Maximum total pixels (width x height)
`dtype`	string	Tensor type: `bfloat16` or `float16`
`supports_negative_prompt`	bool	Whether the model uses negative prompts
`scheduler`	string	Diffusers scheduler class name
`token_window`	int	Maximum prompt tokens (77 for CLIP, 512 for Flux)
`vram_usage`	int	Estimated VRAM in MB

Behavior Flags

Optional flags within settings that enable special model behavior:

Flag	Applicable Models	Description
`force_default_guidance`	Turbo models	Ignore user guidance_scale, use model default
`enable_debug_logging`	Any	Enable debug print statements via `DebugLoggingMixin`
`use_sequential_cpu_offload`	CUDA models	Use sequential CPU offload (more memory-efficient)
`max_sequence_length`	Flux variants	Maximum context length for prompt encoding (512)
`load_in_8bit`	Qwen	8-bit quantization to reduce VRAM usage

LoRA Fields

Field	Type	Description
`label`	string	Display name
`base_architecture`	string	Target model family (must match model’s architecture)
`theme`	string	Category tag (e.g., `anime`, `photorealistic`, `comic`)
`air`	string	CivitAI AIR URN: `urn:air:{arch}:lora:civitai:{model_id}@{version_id}`
`prompt`	string	Trigger words appended when LoRA is active
`negative_prompt`	string	Negative prompt to use with this LoRA
`notes`	string	Metadata or CivitAI stats

LoRA Settings

Setting	Type	Description
`strength`	float	LoRA weight multiplier (typical range 0.5–1.3)
`guidance_scale`	float	Recommended guidance when using this LoRA
`clip_skip`	int	CLIP layer skip (1–2) for SD models

Django Settings

Key settings in src/cw/settings.py.

Celery

Setting	Value
`CELERY_BROKER_URL`	`redis://{VALKEY_HOST}:{VALKEY_PORT}/2`
`CELERY_RESULT_BACKEND`	`django-db` (Django ORM via `django-celery-results`)
`CELERY_RESULT_EXTENDED`	`True`
`CELERY_WORKER_POOL`	`solo` (single-threaded for GPU safety)
`CELERY_TASK_TIME_LIMIT`	`3600` (1 hour hard limit)
`CELERY_TASK_SOFT_TIME_LIMIT`	`3300` (55 minutes soft limit)
`CELERY_TASK_SERIALIZER`	`json`
`CELERY_TIMEZONE`	`UTC`

Logging

JSON-structured logging via pythonjsonlogger.JsonFormatter. Rotating file handlers with 10 MB limit and 5 backups per file.

Logger	Handler	Level	Description
`django`	`django_file`	INFO	Server requests and ORM
`celery`	`celery_file`	INFO	Worker lifecycle and routing
`cw.diffusion.tasks`	`tasks_file`	INFO	Image generation tasks
`cw.tvspots.tasks`	`tasks_file`	INFO	Adaptation and storyboard tasks
`cw.lib.adaptation`	`tasks_file`	DEBUG	Pipeline execution
`cw.lib.models`	`tasks_file`	DEBUG	Diffusion model loading
`cw.lib.prompt_enhancer`	`tasks_file`	DEBUG	Prompt enhancement
`cw.lib.storyboard`	`tasks_file`	DEBUG	Storyboard generation
`cw.lib.civitai`	`tasks_file`	DEBUG	CivitAI LoRA downloads
`cw.lib.security`	`tasks_file`	INFO	File upload validation

Video Upload

Setting	Value
`VIDEO_ALLOWED_EXTENSIONS`	`.mp4`, `.mov`, `.avi`, `.mkv`, `.webm`
`VIDEO_ALLOWED_MIME_TYPES`	`video/mp4`, `video/quicktime`, `video/x-msvideo`, `video/x-matroska`, `video/webm`
`VIDEO_VERIFY_MIME_CONTENT`	`True`
`VIDEO_VALIDATE_HEADERS`	`True`
`VIDEO_SANITIZE_FILENAMES`	`True`

Docker Compose Services

Service	Image	Port	Purpose
`postgres`	`postgres:17`	5435	PostgreSQL database
`valkey`	`valkey/valkey:latest`	6379	Celery broker (Redis-compatible)
`loki`	`grafana/loki:latest`	3100	Log aggregation backend
`alloy`	`grafana/alloy:latest`	—	Log collector (ships `logs/*.log` to Loki)
`grafana`	`grafana/grafana:latest`	3000	Log search UI (anonymous login enabled)

Named volumes: postgres_data, valkey_data, loki_data, grafana_data.

Health checks are configured for postgres (pg_isready) and valkey (valkey-cli ping). Use ./start.sh to wait for healthy containers before starting Django.