Environment Variables¶

All environment variables that configure ComfyUI Studio, organized by category.

Authentication & API Tokens¶

These are the variables you set when creating your RunPod pod or Docker container.

Variable	Default	Recommended	Description
`API_KEY`	`changeme`	Change it	The password for the web UI login and for programmatic access via the `X-API-Key` header. Anyone with this password has full access to your instance — model downloads, job execution, file management, settings. Choose a strong, unique password.
`CIVITAI_API_KEY`	(empty)	Yes	Your CivitAI API token. Required to download models hosted on CivitAI and to fetch model metadata (descriptions, preview images, tags). Without this, CivitAI downloads fail silently. Get your token from civitai.com/user/account → API Keys.
`HF_TOKEN`	(empty)	Yes	Your HuggingFace access token. Required to download gated models — Flux, WAN, HunyuanVideo, and many others require accepting a license on HuggingFace before downloading. Without this token, gated model downloads return 401 errors. Get your token from huggingface.co/settings/tokens → New token → Read access is sufficient.

Where tokens are used

These tokens are stored as environment variables on the pod. The backend injects them into download requests automatically -- HF_TOKEN as a Bearer header for HuggingFace URLs, CIVITAI_API_KEY as a query parameter for CivitAI URLs. They are never sent to any other service.

Runtime-editable auth settings

API_KEY, CIVITAI_API_KEY, HF_TOKEN, TELEGRAM_BOT_TOKEN, and TELEGRAM_BOT_NAME can be edited at runtime from the Settings page without restarting the backend. When edited, the new values are persisted to both os.environ and the SQLite database. On subsequent boots, database values override environment variable defaults.

Configurable¶

These variables let you customize the behavior of the system. All have sensible defaults.

Variable	Default	Description
`STUDIO_DIR`	`/workspace/studio`	Root directory for all Studio data on the persistent volume. All other paths derive from this. Change this only if you want Studio data in a different location.
`STUDIO_PORT`	`8000`	The port the Studio backend listens on. Change this if you have a port conflict or want to run multiple instances.
`COMFYUI_DIR`	`/workspace/ComfyUI`	Path to the ComfyUI installation. On first boot, ComfyUI is copied from the Docker image (`/comfyui`) to this location. Downloaded models go into `COMFYUI_DIR/models/`.
`COMFYUI_PORT`	`8188`	The port ComfyUI listens on. The Studio backend connects to ComfyUI on `localhost:COMFYUI_PORT`.
`COMFYUI_FLAGS`	`--highvram`	ComfyUI launch flags that control VRAM management. Choose based on your GPU's VRAM:

COMFYUI_FLAGS Options¶

Flag	VRAM	Behavior
`--highvram`	> 24 GB	Keeps models in VRAM. Fastest inference but uses the most memory. Default for datacenter GPUs (A100, H100, B200).
`--normalvram`	12-24 GB	Balances between VRAM and system RAM. Good for RTX 4090, A5000, L4.
`--lowvram`	< 12 GB	Aggressively offloads to system RAM. Slower but works on 8 GB GPUs.

Variable	Default	Description
`COMFYUI_EXTRA_ARGS`	(empty)	Additional ComfyUI command-line arguments appended after `COMFYUI_FLAGS`. For example, `--dont-upcast-attention` or `--force-fp16`. See ComfyUI CLI arguments for all options.
`LLAMA_SERVER_PATH`	`/opt/llama-server`	Path to the llama.cpp server binary. If ENABLE_LLM was set to true during the Docker build, the binary is at `/opt/llama-server`. If you compiled it manually on the pod, point this to wherever you put the binary.
`LLAMA_SERVER_PORT`	`8080`	The port llama-server listens on when started from the LLM page. The Studio backend connects to it on `localhost:LLAMA_SERVER_PORT`.
`MAX_CONCURRENT_DOWNLOADS`	`3`	Maximum number of model downloads that run in parallel. Range: 1-10. Can also be changed at runtime from the Settings page without restarting. Higher values download faster but use more bandwidth and can cause timeouts on slow connections.

Read-only (Set by Infrastructure)¶

These are set by the Docker image or the RunPod platform. Do not change them manually.

Variable	Default	Set by	Description
`HOSTING`	(auto)	config.py	Hosting platform identifier. Auto-detected as `"runpod"` when `RUNPOD_POD_ID` is present. Affects which environment variable groups are shown in Settings and how disk usage is calculated.
`RUNTIME_VERSION`	`3`	Dockerfile	The Docker image version number. The application checks this against `min_runtime` in `version.json` to ensure compatibility. If your image is older than what the application requires, you'll see a "Runtime Incompatible" message and need to pull a newer image. This number is incremented when the Docker image includes breaking changes (new dependencies, new custom nodes, updated CUDA).
`REPO_URL`	`https://github.com/diego-devita/comfyui-studio.git`	Dockerfile	The Git repository URL used by the bootstrap script and update mechanism. The application clones this repo on first boot and fetches from it when checking for updates.
`REPO_BRANCH`	`main`	config.py	The Git branch to track. Default `main` for production pods. Set to `dev` to point a test pod at the development branch. Used by bootstrap (clone) and the update mechanism (fetch/reset).
`RUNPOD_API_KEY`	(auto)	RunPod	Automatically injected by RunPod into the container. Used by the telemetry system to query the RunPod API for network volume size (to calculate free disk space on the dashboard).
`RUNPOD_POD_ID`	(auto)	RunPod	The pod identifier. Used in telemetry API calls.
`RUNPOD_DC_ID`	(auto)	RunPod	The datacenter identifier.
`RUNPOD_VOLUME_ID`	(auto)	RunPod	The network volume identifier.

Build Arguments (Docker Build Time)¶

These are set when building the Docker image, not at runtime. They control what's included in the image.

Argument	Default	Description
`CUDA_VERSION`	`12.8.1`	NVIDIA CUDA base image version. Must match your target GPU architecture. See GPU Compatibility.
`PYTORCH_INDEX`	`cu128`	PyTorch wheel index tag. Must correspond to CUDA_VERSION: `cu128` for 12.8, `cu124` for 12.4, `cu121` for 12.1.
`PYTHON_VERSION`	`3.12`	Python interpreter version. 3.12 is recommended for ComfyUI.
`ENABLE_SAGE_ATTENTION`	`true`	Install SageAttention 2 + Triton for 2-3x faster attention. Requires Ampere (SM 80) or newer. Set to `false` for Turing/Volta GPUs.
`ENABLE_FLASH_ATTENTION`	`true`	Install FlashAttention for memory-efficient attention. Builds from source (20-30 min). Requires Ampere or newer.
`ENABLE_LLM`	`true`	Compile llama-server for local LLM inference. Compiles with CUDA for all GPU architectures (SM 75-100). Takes ~60 min on CI without GPU. Set to `false` to skip.
`LLAMA_CPP_VERSION`	`b8505`	The llama.cpp release tag to build. Pinned to a known-good version for reproducible builds. Set to `latest` for the newest code (may break).

Use the Build Configurator (docker/configure.sh) to automatically determine the right values for your GPU.