Loading...
Loading...
Plain-language fixes for 12 of the most common ComfyUI errors — what each one means, why it happens, and the steps to resolve it. Search by message or filter by category.
12 errors
A node received None where it expected a real value — usually because an upstream node produced nothing, or an expected input was not connected.
Generation finishes with no error but produces a black or garbled image. This is usually VAE numerical instability in half precision, not a broken prompt.
The browser lost its connection to the ComfyUI server. The banner persists because the backend crashed, froze, or the connection timed out mid-run.
ComfyUI ran out of GPU VRAM while loading a model or sampling. Almost always fixed by lowering memory pressure, not by buying a bigger GPU.
A custom-node package is installed but crashed while importing at startup, usually because a Python dependency is missing or a version conflicts.
A tensor dimension mismatch, almost always caused by pairing a text encoder, LoRA, or ControlNet from one model architecture with a base model of another.
A workflow references custom nodes that are not installed in your ComfyUI. The missing nodes render red and the graph cannot run until they are installed.
ComfyUI refused to queue the prompt because a node failed pre-run validation — an unconnected required input, an out-of-range value, or a missing file.
A node has a mandatory input with no incoming connection. ComfyUI names the input; connecting it resolves the error.
ComfyUI tried to load a model with the wrong architecture assumptions — the checkpoint’s weight shapes don’t match the model it is being loaded into.
An operation received tensors split across the GPU and CPU. Something in the graph forced part of the work onto a different device than the rest.
A loader references a checkpoint, LoRA, or VAE filename that ComfyUI cannot find in your models folders, so the value is no longer a valid dropdown option.