← All reportsPublic report

comfy-org/comfyui

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.

130k stars Python View on GitHubprofiled 11d ago
Try GitZoid
207PRs this week
30Contributors
5Deps scanned
0Issues found
01 · Repo overview

How comfyui is put together

ComfyUI is a modular, node-graph AI content-creation engine that runs diffusion and other generative models locally, exposing a workflow graph executor over an HTTP/WebSocket server. The backend is Python built on an aiohttp web server (server.py/PromptServer) that queues and executes node graphs (execution.py, nodes.py) against PyTorch-based model inference code under the comfy/ package. Model, checkpoint, LoRA, VAE and text-encoder loading is orchestrated through folder_paths.py plus manager classes in app/, with persistence via SQLAlchemy/Alembic. It ships a bundled frontend package and optional paid API nodes (comfy_api_nodes/) that call external model providers.

Languages

Python

Frameworks

aiohttpPyTorchtransformersSQLAlchemyAlembicpydantic

Datastores

SQLite (via SQLAlchemy/Alembic)filesystem (models, outputs, user files)

Infrastructure

CUDA/ROCm/oneAPI GPU runtimeWebSocket serverOpenAPI (openapi.yaml)

Major components

PromptServer (server.py)

aiohttp HTTP/WebSocket server that registers routes, middleware, and dispatches workflow prompts.

Execution engine (execution.py, nodes.py)

Validates, queues, and runs node graphs, handling partial re-execution and node dispatch.

comfy/ core

PyTorch model loading, sampling, memory/VRAM management, CLIP/VAE/diffusion model inference.

app/ managers

User, model file, custom node, subgraph, and frontend management plus settings and DB access.

comfy_api_nodes/

Optional nodes that call external paid model providers (Anthropic, Gemini, BFL, ByteDance, etc.).

folder_paths.py

Central registry resolving model/input/output/user directories and extra configured paths.

app/database (Alembic)

Schema migrations and SQLAlchemy-backed persistence for assets and app data.

api_server / assets routes

Internal API routes and asset ingest/serving endpoints registered onto the web app.

Over the past few weeks ComfyUI has rapidly expanded its support for new AI models — especially MiniMax-H3 audio/video, LTX 2.5, and a steady stream of third-party 'partner' services (ByteDance, Bria, Gemini, Wan, Meshy) — while shipping better video-saving options (HDR, AV1, mkv/webm) and continuous memory and performance tuning. Alongside the new capabilities, the team put ongoing effort into stability fixes, security hardening, and housekeeping like dropping older Python/PyTorch versions. The result is a tool that keeps adding cutting-edge generation features while smoothing out reliability and setup for everyday users.

Week by week

2026-08-24A lighter week focused on adding new partner models and tidying up setup and defaults.latest6 changes

Feature

New Wan 3.0 video generation nodes

Added building blocks that let users create video with the Wan 3.0 model.

Feature

Meshy-7 3D model support

Users can now generate 3D content using the newer Meshy-7 service.

Fix

Correct database location for custom user folders

The app now stores its database in the right place when a custom user directory is set.

Chore

Python 3.10 end-of-life warning

A startup notice now reminds users that Python 3.10 will stop being supported after October 2026.

Chore

Retired an outdated Tripo 3D option

Removed an old Tripo 'Refine' node and its version 2.0 model that were no longer maintained.

Chore

Refreshed built-in workflow templates

Updated the ready-made example workflows to their latest version.

2026-08-17Big push on richer video saving (HDR, AV1, mkv/webm) plus many new partner nodes and multi-GPU handling on Windows.6 changes

Feature

Higher-quality and more flexible video saving

Added HDR color support, the AV1 codec, and mkv/webm formats when saving videos.

Feature

New 3D body and object generation support

Added support for Sam3d-body, Pixal3d, and TRELLIS2 for creating 3D content.

Feature

Several new partner service nodes

Added tools for ByteDance video enhancement, a BFL Flux video upscaler, a faster Seedream image node, and Fish Audio.

Fix

Smarter default GPU choice on Windows

Windows now defaults to the first CUDA graphics card and better limits which GPUs are visible in multi-card setups.

Feature

Friendlier error messages when loading models

The model loader now gives clearer explanations when something goes wrong with low-memory loading.

Chore

Gemini 3.7 Flash added to the text node

The Gemini text tool can now use Google's newer 3.7 Flash model.

2026-08-10Two major releases (v0.32 and v0.33) landed alongside new music generation and a wave of stability fixes.6 changes

Feature

MiniMax Music 3 music generation

Added the ability to generate music with MiniMax Music 3, including performance work using CUDA graphs.

Feature

LTX 2.5 model support

Added support and dedicated nodes for the LTX 2.5 generation model.

Chore

Released ComfyUI v0.32.0 and v0.33.0

Two tagged versions were published, bundling recent features and fixes for users.

Refactor

Faster Gemma4 text generation

Sped up text output from the Gemma4 model.

Fix

Fixed thinking-mode and music reliability issues

Corrected 'thinking' behavior on Gemma4 and made MiniMax music work on non-dynamic-VRAM setups.

Feature

New Bria editing nodes

Added Bria tools for generative fill, erasing, expanding, and increasing image resolution.

2026-08-03Released v0.31 with heavy MiniMax-H3 optimization, new compositing tools, and sampler improvements.6 changes

Refactor

MiniMax-H3 optimizations

Made the MiniMax-H3 image decoder run faster and use memory more efficiently.

Feature

Layered image compositing tools

Added an ImageCompositor node with layer editing plus a Seedream layer-separation node.

Chore

Released ComfyUI v0.31.0

Published a new tagged version and raised the minimum supported PyTorch to 2.7.

Feature

New partner models added

Added ByteDance SeeDance 2.5, a BFL Flux 3 video model, and Topaz AI Bloom 2 / Wonder 3.5 support.

Fix

Improved audio sampling stability

Fixed several sampler and offloading issues affecting MiniMax audio generation.

Refactor

Speed and memory tuning

Sped up LTX and Wan and reduced unnecessary memory pinning on some Linux systems.

2026-07-27Introduced MiniMax-H3 support with security hardening, more video/text saving options, and memory management work.6 changes

Feature

MiniMax-H3 model support

Added initial support for the MiniMax-H3 model, including higher 768P resolution.

Feature

More flexible saving options

The save-text node can now export CSV, and the save-video node gained a quality (crf) control.

Fix

Security fixes for previews and file access

Fixed an SVG preview cross-site scripting risk and restricted dataset nodes to a safe folder.

Refactor

Smarter memory loading

Improved how model weights are loaded into RAM and cleaned up memory pinning behavior.

Feature

New Ideogram P-Image model

Added support for Ideogram's new P-Image model in the partner nodes.

Fix

Various crash and reliability fixes

Resolved crashes in audio-less video sampling and restored a broken custom stylesheet load.

03 · Security check

Dependencies and code review

5 dependencies scanned

Dependency advisories

Security Watch

No known advisories across 5 scanned dependencies.

No known advisories in the scanned dependencies.

Code review

No concrete code-level issues confirmed.

Get this report every week for your repos.

GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.

$19 a month, flat · First 10 outputs free · No card required