Top-level files

These aren't a package, but they hold the runtime entry points and the standard node library. They live at the repository root and import everything else.

Purpose

main.py boots the process. server.py is the HTTP/WebSocket layer. execution.py runs queued prompts. nodes.py is the standard node library that ships with ComfyUI. folder_paths.py is the model/path registry. The rest are small focused helpers.

Files

File	Lines	Role
main.py	531	Entry point. Parses args, sets device env vars, starts the prompt worker thread
server.py	1,288	The aiohttp PromptServer — /prompt, /queue, /history, /view, the WebSocket
execution.py	1,355	The PromptExecutor and validation: builds the DAG, traverses caches, runs nodes
nodes.py	2,513	The standard node library: LoadCheckpoint, KSampler, VAEEncode, SaveImage, …
folder_paths.py	498	folder_names_and_paths registry; get_folder_paths, add_model_folder_path
latent_preview.py	137	TAESD/Latent2RGB preview rendering during sampling
cuda_malloc.py	101	cudaMallocAsync detection and blacklist
node_helpers.py	91	Small helpers used by built-in nodes (image loading, etc.)
new_updater.py	35	Windows standalone build updater
protocol.py	7	BinaryEventTypes constants for the WebSocket binary protocol
hook_breaker_ac10a0.py	17	Save/restore monkey-patched core functions around node execution
comfyui_version.py	3	The version constant; auto-generated from pyproject.toml

main.py — startup sequence

main.py does, in order:

1. import comfy.options; comfy.options.enable_args_parsing() — must run before any module that imports comfy.cli_args.
2. Disable HF/Posthog telemetry env vars.
3. Configure device env vars: CUDA_VISIBLE_DEVICES, HIP_VISIBLE_DEVICES, ASCEND_RT_VISIBLE_DEVICES, MIMALLOC_PURGE_DELAY, OCL_SET_SVM_SIZE. These must be set before torch is imported.
4. Conditionally bring up comfyui_manager if --enable-manager is set.
5. Apply custom paths from extra_model_paths.yaml, register output directories.
6. Run prestartup_script.py from each custom node directory.
7. Import torch + the rest of comfy/. If torch was already imported, log a warning.
8. Call start_comfyui() which builds the PromptServer, initializes nodes, sets up the DB, and launches the prompt worker thread.
9. Run the asyncio loop.

The prompt_worker function is the per-request handler — see Prompt execution.

server.py — the aiohttp app

server.py defines PromptServer. Key endpoints:

Path	Purpose
POST /prompt	Validate and enqueue a prompt
GET /queue	Inspect the queue
GET /history, /history/<prompt_id>	Past executions
POST /interrupt	Cancel the current execution
GET /view?filename=...	Read an output/input/temp file
GET /system_stats	Memory and device state
GET /object_info	Schema dump of every registered node
GET /ws	WebSocket for progress, status, executed events
GET / and asset paths	Serve the frontend bundle from web/
GET /internal/*	Routes mounted from api_server/routes/internal/

Middleware stack (composable via flags):

• cache_control — from middleware/cache_middleware.py
• compress_body — gzips JSON when --enable-compress-response-body is set
• cors_middleware — when --enable-cors-header ORIGIN is set
• origin_only_middleware — blocks cross-site fetches when listening on loopback (default)
• block_external_middleware — sets a Content-Security-Policy header
• deprecation_warning — logs once when legacy /scripts/ui paths are hit

The WebSocket binary protocol is defined in protocol.py (BinaryEventTypes.UNENCODED_PREVIEW_IMAGE etc.).

execution.py

PromptExecutor.execute(prompt, prompt_id, extra_data, ...) is the main entry. Internally it:

1. Validates the prompt via validate_prompt.
2. Builds an ExecutionList from comfy_execution.graph.
3. Walks output nodes backward, checking the CacheSet (outputs, objects).
4. For each node, calls _async_map_node_over_list which handles list-input batching, lazy inputs, IS_CHANGED, V3 fingerprint_inputs, and node expansion via DynamicPrompt.
5. Accumulates history_result and ships progress to the server.

See Prompt execution for the full flow.

nodes.py

The standard node library. Roughly 100 nodes covering text encoding, sampling, model loading, image I/O, latent ops, conditioning ops, model merging, ControlNet loading, and the LoRA/IP-Adapter family. nodes.NODE_CLASS_MAPPINGS is the global registry; init_extra_nodes populates it further from comfy_extras/ and custom_nodes/.

Notable classes:

• CheckpointLoader, CheckpointLoaderSimple, UNETLoader — load model files
• CLIPLoader, DualCLIPLoader, TripleCLIPLoader — load text encoders
• VAELoader, VAEEncode, VAEDecode
• CLIPTextEncode — text → conditioning
• KSampler, KSamplerAdvanced — the canonical samplers
• LoraLoader, LoraLoaderModelOnly — apply a LoRA
• LoadImage, SaveImage, PreviewImage — image I/O
• ConditioningCombine, ConditioningSetArea, ConditioningSetMask — conditioning ops

folder_paths.py

The map from logical model categories to disk paths and accepted extensions. Used everywhere instead of hardcoded paths. Custom nodes call folder_paths.add_model_folder_path("checkpoints", "/some/path") to register additional locations. The default registry is set up at module import; extra_model_paths.yaml is layered on top by apply_custom_paths in main.py.

Supported model file extensions: {'.ckpt', '.pt', '.pt2', '.bin', '.pth', '.safetensors', '.pkl', '.sft'}.

latent_preview.py

Provides three preview methods: NoLatentPreviewer (off), Latent2RGBPreviewer (fast, low-quality, no extra weights), and TAESDPreviewer (loads small TAESD decoders from models/vae_approx/). Selected via --preview-method.

cuda_malloc.py

Detects whether the GPU supports cudaMallocAsync and maintains a blacklist of cards that crash with it (the warning is printed at startup if you have one). Imported early in main.py so the env var is set before torch initializes the CUDA allocator.

protocol.py

Tiny module:

class BinaryEventTypes:
    UNENCODED_PREVIEW_IMAGE = 1
    PREVIEW_IMAGE_WITH_METADATA = 2

These are message-type bytes for the WebSocket binary frames the server sends to push preview images.

hook_breaker_ac10a0.py

Saves references to a small set of core functions before custom-node imports run, and restores them after each prompt execution. The cryptic suffix avoids name collisions with custom-node modules. See Debugging.