ComfyUI Glossary

How ComfyUI Glossary works on Windows

Plain-English definitions of every term on the ComfyUI canvas: KSampler, CFG scale, Euler, denoise, VAE, CLIP, and FP8. Understanding what each pin and parameter on a ComfyUI node actually computes during a diffusion pass.

Every page in this directory focuses on practical local execution on Windows 10 and Windows 11, keeping your checkpoints, LoRAs, and generated outputs entirely on your own SSD.

Start by grabbing the current Windows build from /download, or browse the guides and node blueprints linked below.

Key workflow principles

• Keep large `.safetensors` and `.gguf` weights inside organized `models/` subfolders or map shared directories through extra_model_paths.yaml.
• Match your model precision (FP16, FP8 e4m3fn, or GGUF Q4/Q8) to your card's VRAM tier so sampling stays in GPU memory without paging to disk.
• Drag any generated PNG directly onto the ComfyUI Studio canvas to reload its exact node wiring, seed, and sampler values.

Featured pages in ComfyUI Glossary

• Latent Space & LATENT Pin — The compressed 8× smaller tensor representation where diffusion sampling happens before VAE decoding.
• CFG Scale (Classifier-Free Guidance) — Controls how strongly the sampler forces the image toward your positive prompt and away from the negative prompt.
• Sampler vs Scheduler — The sampler is the mathematical solver (euler, dpmpp_2m) while the scheduler spaces the noise levels across steps (karras, simple, normal).
• Denoise Strength — Determines how much noise is added and removed during sampling (1.00 for text-to-image, 0.30–0.65 for img2img and upscaling).
• Safetensors Format (.safetensors) — Hugging Face's zero-copy tensor serialization format that cannot execute arbitrary Python pickle code.
• GGUF Quantization (Q4 / Q5 / Q8) — Block-quantized weight format that shrinks 24 GB Flux and Hunyuan Video models to fit 8 GB and 12 GB GPUs.
• FP8 (e4m3fn / e5m2) Precision — 8-bit floating point weight storage that halves checkpoint memory compared to FP16.
• VAE (Variational Autoencoder) — The neural encoder/decoder that converts between RGB pixel images and compressed diffusion latents.
• CLIP Skip (CLIPSetLastLayer) — Stops text encoding one or more layers early in the CLIP transformer, common in anime and Pony checkpoints.
• LoRA (Low-Rank Adaptation) — Small 20 MB–300 MB patch weights trained to add a specific character, art style, or concept to a base checkpoint.
• ControlNet Conditioning — Spatial guidance that locks pose, depth, edges, or line art during diffusion generation.
• Embedded Workflow JSON in PNG — ComfyUI saves the entire node graph inside the tEXt chunk of every generated PNG image.