📍 Home › Local AI

ComfyUI on a Home GPU: the Node Graph for Local Image Models

Local AI  Updated 2026-11-04 · 5 min read · #comfyui #image #flux #local #gpu

Status board (checked 2026-11-04)

What it is

ComfyUI is the local workbench for image models: you wire loaders, samplers, VAEs and (optional) custom nodes into a graph, queue a run, and get files in ComfyUI/output/. It is not a chat engine — that side of the home-rig story is Strata / Maya. Same PC, different job: VRAM for pixels, not MoE experts in RAM.

The project also sells Comfy Cloud and enterprise hosting. This page is about running on your own GPU. Partner / cloud nodes that bill per image are optional; you can stay fully offline after models are downloaded.

Free tier

• Local app: free. GPL-3.0, no feature gate on the core engine, no “pro” unlock for basic generation (download FAQ).
• Models keep their own licenses. Example: FLUX.3 [klein] weights are under the FLUX Kommunity License on a gated Hugging Face repo — accept that agreement before you download; commercial terms are in that license, not in ComfyUI’s GPL.
• Comfy Cloud / Partner Nodes: paid compute or per-image APIs when you choose them. Skip them for a pure home-rig setup.

Getting started (3 steps)

1. Install the right package for your OS from comfy.org/download or the system-requirements page: Desktop if you want auto-updates; Windows Portable if you want a self-contained folder on the latest commits; manual install if you need a specific CUDA / ROCm / MPS stack.
2. Pick one starter model family and one precision, not “every checkpoint on Civitai.” For a current open-weights path, Comfy’s own templates cover FLUX.3 Image Klein (needs ComfyUI 0.41.0+): accept the gated HF license, drop the int8 (or bf16) files into models/diffusion_models, models/text_encoders, and models/vae as the Klein tutorial lists, then open the matching Workflow Template.
3. Queue once, then read VRAM + system RAM in your OS monitor. Comfy will happily fill the card; system RAM still holds weights. If the machine pages hard, turn the prompt upsampler off or drop resolution before buying another GPU.

What fits which card (official guidance, Nov 2026)

Comfy’s FLUX.3 Klein docs ship int8 templates “to keep the memory footprint small,” and say that with more than 24 GB VRAM you can swap in bf16 checkpoints (and the bf16 prompt upsampler) for closer prompt following. NVFP4 builds need a Blackwell GPU. Exact peak MiB depends on resolution, batch, and whether the Qwen3.5 upsampler is on — treat the table as a starting map, not a promise.

Disk: the app is light; models are not. Plan a dedicated folder (tens of GB per family). Python: docs recommend 3.13 for manual installs; Desktop/Portable bundle their own env. Apple Silicon: supported via PyTorch MPS (not MLX) per Comfy’s system-requirements FAQ.

Best use

• Reproducible local image / edit pipelines you can version as JSON workflows (same graph on Desktop, Portable, or later Cloud).
• FLUX.3 Klein text-to-image and instruction edits on your GPU with no per-image API bill after download (official tutorial).
• Same machine as your LLM stack when you schedule jobs: run Comfy when you need pixels; run Strata / Maya when you need tokens — a 16 GB card rarely does both at full tilt.

Honest weaknesses

• The node graph is the product and the tax. Power users love it; “just make a picture” users bounce until they load a template.
• Custom-node entropy. Thousands of extensions exist; broken graphs after updates are normal. Prefer official templates until one workflow is stable.
• VRAM numbers lie without context. Peak usage tracks resolution, batch, upsampler, and whether weights sit in RAM. Blog “8 GB works” posts often assume a specific int8 stack and ~32 GB system RAM.
• License stacking. GPL for the app ≠ free commercial use of every weight file you drop in models/.

FAQ

ComfyUI or AUTOMATIC1111 / Forge?

ComfyUI won the “serious local pipeline” niche: graphs, templates, and fast model support. A1111-style UIs are simpler for one-off prompts. If you already live in workflows, start here.

Does this replace Strata or Maya?

No. Those run chat / code MoE models. ComfyUI runs image (and video) diffusion graphs. Link them only if you build a tool that calls both.

Do I need Comfy Cloud?

Only when your GPU is too small for the job or you want a partner model that never ships weights. Desktop stays free for local open models.

Will you publish speeds for the 4070 Ti SUPER box?

Yes — when we have a dated, stopwatch-backed table (model, precision, resolution, steps, sec/image). Until then, trust Comfy’s docs for fit, not influencer clips for timing.