The pitfalls that cost you a weekend

ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. is famous for a steep first week, and almost all of the pain is a handful of specific, avoidable pitfalls rather than genuine difficulty. Know them and the curve flattens fast.

The deepest one is architecture mismatch. The model families do not interoperate, and mixing them, an SDXL checkpointCheckpointA saved set of model weights. The .safetensors or .ckpt files you download to run a model. with an SD1.5 ControlNetControlNetA conditioning technique that lets you provide structural guidance (depth, edges, pose) to a diffusion model to control spatial layout., a Flux model with the wrong text encoderText encoderA neural network that converts text prompts into numerical embeddings that the diffusion model can use. Modern models use combinations including CLIP and T5., produces not a helpful message but a cryptic wall of tensor-shape numbers that says nothing about the real cause. The families differ underneath, in latent channel counts and text encoders, and a large share of beginner frustration is simply one family's part plugged into another's socket. The close cousins are the VAEVAE (Variational Autoencoder)A neural network that learns to compress images into a small numerical representation (a latent) and decompress them back. Used in latent diffusion models to make the generation process tractable. mismatch, which turns your output purple or grey, and the wrong resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. for a model, which gives you the classic two-headed, melted anatomy because you ran an SDXL model at 512 pixels or an SD1.5 model at 1024.

The second cluster is settings carried over from other tools. Flux in particular breaks completely at the guidance value that is normal in older models; it wants its CFGCFG (Classifier-Free Guidance)A technique that pushes a diffusion model to follow the prompt more strongly by contrasting its prediction with and without the prompt. Higher values obey the text more but can look oversaturated or 'fried'. near one, with a separate guidance control for adherence, and a beginner applying the old muscle memory of CFG seven will destroy every generation and not know why. The third cluster is custom nodes: you open a workflow someone shared, half the nodes are red, and 'installed' in the Manager does not mean the import succeeded, it means the files are on disk while a Python dependency clash quietly failed underneath. The fix is almost always the Manager's install-missing button followed by a full restart of the process, not just the browser.

The meta-pitfall, the one that compounds, is copying a sixty-nodeNodeA single block in a visual pipeline like ComfyUI that does one operation and wires to others to form a workflow. workflow you do not understand. When it breaks, and it will, you are helpless, because you cannot see which link carries what. The whole craft is built one node at a time, and the people who get good fast are the ones who resisted the urge to paste in someone else's masterpiece and instead built the small graph themselves until the data flow was obvious.