mrkeyoor.com_
Tue 01 Sept 17:44 UTC
AI Toolsevaluationupdated 29 Aug 2026

ComfyUI review

ComfyUI is a visual, node-based engine for building AI media workflows without writing the whole pipeline in code. It lets creators connect models and processing steps for images, video, audio, 3D, and text, then save, reuse, or expose those workflows through an API or simplified App Mode interface.

+366 / 2dstars / 7d
Verdict

Our build finished in 8 seconds, but the test run ended with 44 failures and 65 collection or setup errors. ComfyUI is still the strongest choice here when visible, reusable graph control is the actual requirement, particularly across several media types. Use a stable release and validate the exact custom nodes your workflows need; choose a simpler alternative if your job is mostly prompt, generate, and export.

We ran it

Install✓ · 88s126 packages · 5926 MB
Build✓ · 8s
Tests✗ · 275s1447 passed · 44 failed · 10 skipped · 65 errors of 1556 (pytest)
Known vulns0(pip-audit)
Repo1068 files~289,473 lines of source · 45.3 MB · 28 CI workflows · tests dir

Answers from our run

Does ComfyUI build from source?

Dependencies installed in 88 seconds (126 packages), and the build succeeded in 8 seconds. We cloned commit 0a33ed6 into a clean Debian container with 3 CPUs and no project-specific setup.

Do ComfyUI's tests pass?

Not all of them: 1447 of 1556 passed and 44 failed when we ran the project's own test command (pytest), with 65 collection errors. Some failures need services or credentials a bare container does not have.

Does ComfyUI have known vulnerabilities in its dependencies?

pip-audit found none in the dependency tree at the time of our run.

Who should not use ComfyUI?

Beginners who want a single prompt box and no graph concepts

What are the alternatives to ComfyUI?

AUTOMATIC1111 Stable Diffusion WebUI, Fooocus, InvokeAI. Our build finished in 8 seconds, but the test run ended with 44 failures and 65 collection or setup errors.

Setup3/5Install worked in 88s, but used 5,926 MB and tests were not clean
Docs4/5Clear install paths, feature map, examples, and release guidance
Community5/5130,552 stars, a recent push, and an active release cycle
Maturity4/5Broad capability and stable releases, with custom-node compatibility risk

Who it’s for

Visual creators who want precise control over models, parameters, and processing steps
Technical artists building reusable image, video, audio, or 3D workflows
Developers who need a local API behind an existing application
Teams willing to maintain model files, custom nodes, and version compatibility

Who it’s NOT for

Beginners who want a single prompt box and no graph concepts
Teams expecting a tiny install, our environment reached 5,926 MB after installing dependencies
Operators who require every upstream test to pass in a generic fresh container
Users who do not want to manage model storage, GPU compatibility, or custom-node breakage

Setup reality

In our run, installation succeeded in 88 seconds with 126 packages, the build succeeded in 8 seconds, and the checkout occupied 5,926 MB after installation. Tests did not pass cleanly: pytest reported 1,447 passed, 44 failed, 10 skipped, and 65 collection or setup errors out of 1,556 in 275 seconds. The shown failures repeatedly said async functions were not natively supported, so the README's easy desktop or portable paths may be friendly, but a source checkout and validation still demand Python environment discipline, substantial disk space, and attention to test tooling.

It is a workflow engine, not just another image generator

ComfyUI turns generative-media pipelines into visual graphs. Instead of hiding the process behind one prompt field, it exposes models, encoders, samplers, conditioning, masks, upscalers, and output steps as connected nodes. That makes the project useful when repeatability matters: a creator can save a workflow as JSON, reopen it later, or recover supported workflow data and seeds from generated media. The repository has 130,552 stars, but popularity is less important than the core idea: the graph is both the editing surface and an executable description of the work.

The scope is unusually broad. The README documents image generation and editing, video, audio, 3D, vision, and text workflows, with native support spanning families such as Stable Diffusion, Flux.1, Qwen Image, Wan 2.2, LTX-Video 2.3, ACE-Step 1.5, and Hunyuan3D 2.1. It can load checkpoints plus diffusion models, VAEs, text encoders, LoRAs, ControlNets, adapters, and upscalers. Built-in operations include inpainting, outpainting, compositing, segmentation, depth estimation, frame interpolation, model merging, and media processing.

Our install worked, while the full test run did not

We cloned commit 0a33ed6 into a fresh Debian container with 3 CPUs and 8 GB of RAM. The repository contained 1,068 files, about 289,473 lines of source, and occupied 45.3 MB when checked out. Installation succeeded in 88 seconds and brought in 126 packages. The build then succeeded in 8 seconds. Those are encouraging results, although the installed environment grew to 5,926 MB before adding model files.

The tests were not clean on our box. Pytest ran for 275 seconds and reported 1,447 passed, 44 failed, 10 skipped, and 65 collection or setup errors out of 1,556. The final failures shown were concentrated in cache-control tests, and the repeated message was that async functions were not natively supported. We are not assigning a cause beyond that evidence. A successful install and build do not guarantee that the validation suite will run correctly in a generic environment without further test configuration.

Pip-audit found 0 known vulnerabilities among the installed Python dependencies. The repository includes a tests directory and 28 CI workflow files, but no Dockerfile. Teams wanting container deployment must define that layer themselves, including GPU access, model volumes, media tools, and dependencies.

Control and reuse are the decisive strengths

ComfyUI's best feature is selective execution. Its asynchronous queue, partial graph re-execution, model offloading, and VRAM and RAM management aim to avoid unnecessary work when only part of a graph changes. Quantized-model support can help fit constrained hardware, although we did not benchmark generation speed or memory consumption. Reusable subgraphs and templates reduce duplication, while App Mode can hide a sophisticated workflow behind a simpler interface.

It also fits automation better than many creator-first interfaces. Local API endpoints let an application submit work without reproducing pipeline logic elsewhere. Additional model paths can be configured through extra_model_paths.yaml, useful when tools share a model library. Core does not download anything unless requested, and --disable-api-nodes disables optional paid API nodes. Output support includes 16-bit PNG, 32-bit EXR, 10-bit AVIF, and HDR media.

Flexibility creates operational rough edges

A node graph offers precision, but it also exposes complexity. New users must understand data flow, compatible types, model components, and invalid connections. Large graphs can be harder to review than a concise script or preset. The README warns that commits outside stable tags may be very unstable and can break many custom nodes. Pinning ComfyUI, node extensions, models, and workflow JSON together is the safer production pattern.

Storage and hardware planning are part of setup, even though the desktop application is presented as the easiest route. Our dependency environment alone reached 5,926 MB. Models can add much more, and supported GPU types do not mean every workflow will be practical on every device. Partner nodes add closed-source services, while local mode and the API-node disable flag support a stricter offline boundary. Decide that boundary before importing community workflows or custom nodes.

The project is active, but the issue queue needs triage

Project activity is unmistakable: the latest release is v0.34.0 from August 26, 2026, and the last push was August 29, 2026. The documented process targets roughly weekly releases, with stable core versions around every 2 weeks and frontend updates on a separate rhythm. That cadence is healthy for fast-moving model support, but upgrades deserve workflow-level regression checks.

There are 4,736 open issues. That number can reflect adoption, feature breadth, extension interactions, and genuine defects at once, so it is neither proof of poor maintenance nor something to ignore. The combination of a same-day push, a release 3 days earlier, 28 CI workflows, and high adoption signals an active project. The large queue means users should search reports carefully and expect some environment-specific problems to require their own diagnosis.

It belongs behind a controlled creative or application layer

In a real stack, ComfyUI works best as the execution layer behind curated workflows. Technical artists build and test graphs, operators pin versions and model assets, and an application calls the local API or exposes selected inputs through App Mode. Store workflow JSON with the project, keep generated media in managed storage, and separate experimental custom nodes from the stable environment. A one-person studio can use the desktop or Windows portable package to remove some overhead.

Choose ComfyUI when the graph itself is valuable: you need inspectable routing, reusable stages, several model families, or media processing beyond one-click generation. Do not choose it merely because it is popular. A conventional interface can reduce cognitive load, while an artist-facing canvas may suit other teams. ComfyUI earns its place when control and reproducibility justify maintaining the graph and its dependencies.

Alternatives

ProjectWhat it isPick it when
AUTOMATIC1111 Stable Diffusion WebUI gh↗A familiar browser interface centered on Stable Diffusion generation and extensions.Pick this instead when you prefer conventional panels and presets over wiring node graphs.
FooocusA streamlined image-generation interface designed to reduce configuration work.Pick this instead when fast, guided image creation matters more than exposing every pipeline step.
InvokeAIA creative AI platform combining canvas-oriented work with workflow tooling.Pick this instead when an artist-facing canvas and asset workflow are more important than ComfyUI's graph-first flexibility.

What people are saying

  1. [velocity-scout] Comfy-Org/ComfyUI
  2. [github-trending] WhatDreamsCost/WhatDreamsCost-ComfyUI
  3. [velocity-scout] LeonQ8/ComfyUI-ALLinONE-MinimaxH3
  4. [hf-trending] drbaph/MiniMax-H3-Turbo-Lora-ComfyUI (trending model on Hugging Face)
  5. [github-trending] Comfy-Org/ComfyUI
  6. [hf-trending] ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot (trending model on Hugging Face)

Sources

  1. ComfyUI GitHub repository
  2. ComfyUI official website

More ai tools reviews

claudian · SkillSpector · robin · mjlab · MoGe · awesome-design-md · the whole board →