How to Use ComfyUI: A Step-by-Step Tutorial

No, but it's the fastest path. Nvidia (CUDA) gives the best performance. Apple Silicon (M1/M2/M3) is well-supported via Metal. AMD GPUs work via ROCm on Linux. CPU-only mode is possible but very slow (minutes per image vs. seconds).

A1111 is simpler upfront, with one big settings panel. ComfyUI has a steeper learning curve but is far more powerful once you are past the basics. ComfyUI supports newer architectures (FLUX, video, audio) that A1111 does not. Most power users have migrated to ComfyUI.

A checkpoint is the full base model (2-7 GB). A LoRA is a small file (50-200 MB) that adjusts a checkpoint toward a specific style, character, or concept, like a filter layered on top. An embedding (textual inversion) is even smaller and teaches the model new concepts via the text encoder.

Three most common causes: CFG Scale too high or too low (try 7-9 for SDXL); wrong aspect ratio for your model (SDXL prefers 1024x1024, SD 1.5 prefers 512x512); or a model-LoRA mismatch, so ensure LoRAs match the base checkpoint version.

FLUX requires a different node setup than SDXL or SD 1.5, specifically the FLUX CLIP loader (dual CLIP: T5XXL + CLIP-L) and an appropriate KSampler configuration. The ComfyUI template library includes a ready-made FLUX workflow. Download FLUX.1 Dev or Schnell from Hugging Face / Black Forest Labs and use the built-in FLUX template as your starting point.

The core software is completely free and open-source (GPL-3.0). Running it locally costs nothing beyond electricity. Comfy Cloud (the cloud-hosted version) uses a credit-based pricing model for GPU time.