ComfyUI

Where does your data go?

Nowhere — with one optional exception you control:

• Locally: fully offline. ComfyUI's core "will never download anything unless you want it to" — no account, no login, no telemetry requirement. Prompts, uploads, and outputs stay on your machine.
• The optional exception is API nodes — paid access to closed models through Comfy's cloud, disabled with a documented flag (--disable-api-nodes). Default-on, so a business owner who wanted strictly local should know it exists.
• The catch: "local" ends when you expose the port. The web interface listens unauthenticated by default, and roughly a thousand instances are publicly reachable on the internet right now — the exposure the security findings below feed on.

What happens when it's wrong?

ComfyUI publishes no quality benchmarks — output quality belongs to the model you load, not the tool. Its documented failure modes are operational:

• It's a workflow engine, not an assistant. Every step (load checkpoint, encode prompt, sample, decode) is a visible, wirable node — though the V1 UI overhaul added template workflows, guided setup, and an essentials category that let beginners start without touching the graph. The learning curve is shallower than it was; it still exists.
• Output quality is the model's, and model licensing is the trap. FLUX.1 [dev] is freely downloadable but Non-Commercial — using it for client deliverables without a paid Black Forest Labs license is a violation. This is "the single most common mistake in run-Flux-locally guides." License-clean options: FLUX.1 [schnell] and FLUX.2 klein 4B (Apache 2.0), SDXL (no revenue cap), SD 3.5 (free under $1M revenue).
• Documented security failures — the heaviest caveats of any listing. An active campaign (March 2026) scanned roughly a thousand internet-exposed ComfyUI instances, exploited custom nodes that execute arbitrary Python, and installed cryptominers plus a proxy botnet with persistence that survives reboots. Separately, malicious nodes in the official Comfy Registry distributed Akira Stealer (credential and wallet theft) disguised as an "Upscaler" node — around 779 installs. CVE-2025-67303 (ComfyUI-Manager, CVSS 7.5, fixed in v3.38) let unauthenticated attackers install malicious nodes remotely.
• What the failure teaches: the custom-node ecosystem is a third-party software supply chain with no enforcement boundary — the openness that makes ComfyUI powerful is what makes it attackable. The core project patches promptly; the defaults trust the operator.

What's it actually bad at?

• The interface has genuinely improved — but the mental model still assumes you care. The V1 frontend overhaul shipped a desktop app, tabbed workflows, a template browser with 400+ ready-to-run workflows, automatic model downloads for missing templates, and drag-and-drop model loading; the June 2026 "Comfy Desktop" rework added multi-instance management with snapshot rollback. Templates and an "essentials" category now soften the onboarding substantially — one-click workflows exist. What's still missing is the Canva layer: brand kits, an asset library, a layout canvas. The assumption is no longer that you know what a VAE is — it's that you want to control what a VAE does.
• Hardware is the real price, steeper than Ollama. 12GB VRAM is the practical minimum for a good all-round experience; 16GB runs full-precision Flux comfortably; video generation wants 24GB. Plus 200–500GB NVMe for models. CPU-only generation runs 5–15 minutes per image; video runs 4–20 minutes per clip on high-end cards.
• Custom nodes are the supply chain. More than 1,300 community packages, several found executing arbitrary code, one confirmed malware distribution through the official registry. The registry response was reactive — malicious nodes stayed live with growing install counts.
• Model licenses vary wildly — "free download" and "free to use commercially" are different things, and no gate warns you.

What it does that the cloud tools can't

• Named model support, day one. Every major release ships as a ComfyUI workflow first — the FLUX and SD 3.5 model cards publish ComfyUI graphs as the canonical way to run them. Image: SD 1.5, SDXL, SD 3.5, Flux.1/2, Qwen Image, Z-Image. Video: Wan 2.1/2.2, LTX-Video, HunyuanVideo 1.5, Mochi. Audio: ACE-Step, Stable Audio 3, MiniMax Music. 3D: Hunyuan3D. Text: Gemma, Qwen3.
• The advanced workflow stack: ControlNet (pose/depth/edges as guardrails), IPAdapter (face and style consistency across a set), regional prompting, multi-stage refiners, upscalers — and LoRA training, with 10,000+ community LoRAs. A business can train a brand LoRA per product line and keep the file.
• Model-licensing map, briefly: FLUX.1 schnell and FLUX.2 klein 4B (Apache 2.0) and SDXL (no revenue cap) are clean for client work; SD 3.5 is free under $1M revenue; FLUX.1 [dev] needs a paid BFL license. FLUX.2 klein 4B generates in about 13GB VRAM with a claimed sub-second time (vendor figure).
• Performance engineering: only re-executes the nodes that changed between runs; smart memory management runs large models on cards as small as 1GB VRAM via offload; NVFP4 quantization, async offload, and pinned memory (2026 additions) cut video-workflow wait times.
• App Mode: a finished workflow can be exposed as a simple form for non-technical staff — build the pipeline once, hand the team a one-screen tool. The closest thing in the local world to a Canva hand-off.

What does it cost really?

The traps: "free download" is not "free license" — the asterisk above is where most local-generation guides lose their readers. Comfy Cloud exists as the official paid option for those without hardware — optional; local use never touches it. The honest economics: $0 per output after hardware is a volume argument that pays off only at sustained high output, or if the GPU already exists. Video is the heavy end: 4–20 minutes per clip on an RTX 4090.

What does it replace — and what do you still do?

It replaces the API bill and the vendor relationship across every modality: brand visuals, short video clips, audio, 3D assets, and text via open models — all on your own GPU, fully offline, no rate limits, and with license-clean models, clean commercial rights. For a business generating serious volume, the brand LoRA you train becomes an asset, not a rental. You still own: the node-graph learning curve, model selection and licensing diligence, software patching, custom-node vetting, and network security.

What the ecosystem offers

• Every new model, day one: when the big labs release a model, its official instructions are usually a ComfyUI workflow. Recent weeks added Wan 3.0 video, Seedance 2.5, Meshy-7 3D, Flux video upscaling, and Meta's Muse Image — which can edit a picture using up to ten reference images at once.
• The closed models, rented: alongside the free open models, ComfyUI now brokers the big labs' closed models (Kling, NanoBanana, Gemini, Grok) at metered credit pricing — a documented setting turns that off if you want strictly local.
• Security response, re-checked: the July security release fixed the four published vulnerabilities (with tests added so they stay fixed), and content-security headers landed on the API. The unpatched Windows updater and exposed-server landscape remain the standing caveats.
• Performance keeps climbing: better memory handling on AMD cards, Windows multi-GPU controls, and faster text generation on the newest open models.

If I stop paying, what do I lose?

• Nothing — there's nothing to cancel. GPL-3.0 open source, no account, no server-side state. Your workflows are JSON files; your models are files on disk.
• Portability is real: workflows load from PNG metadata, JSON export works, and the same models run in alternative interfaces. No exit cliff.
• The lock-in is skill, not vendor: node-graph expertise transfers to every diffusion pipeline, and the workflows you've built are assets. The healthiest exit profile in the design category.

Who is this NOT for?

Anyone who wants a design tool rather than an engine — Canva and Adobe Express own that job and do it well. Businesses without GPU hardware — the $0 software runs on a $1,500-plus machine for comfortable production. And anyone unwilling to manage a software supply chain: unpatched, internet-exposed ComfyUI is actively being farmed by botnets.