ComfyUI & Stable Diffusion error fixes
Pasted an error and just want it fixed? Pick the message you hit. Every page has the cause, the concrete fix, and — when it's a VRAM ceiling — the cheapest GPU that actually holds the model.
GPU Memory Exhausted
A "CUDA out of memory" error means the model needed more GPU VRAM than the card has. It is almost always a VRAM ceiling, not a bug in your workflow — the same graph runs fine on a bigger GPU.
Model Not Found
ComfyUI couldn't find the checkpoint/LoRA the workflow references. The filename in the graph doesn't match any file in the models directory — often an exact-name or subfolder mismatch.
Missing Python Package
A ModuleNotFoundError at ComfyUI startup almost always means a custom node's Python dependency didn't install — the node folder is there, but a package it imports is missing.
Disk Space Exhausted
This is a container DISK limit, not VRAM. Model weights download to the container disk, and large checkpoints (a 62GB LTX-2 or 31GB Z-Image) overflow a small volume before the download finishes.
Gated Model Access
A 401 on a model download means the weights are gated on HuggingFace (FLUX.1-dev, some Llama/Gemma repos) and the download had no accepted-license token. The node isn't missing — the download is being refused.
missing custom nodes (node type not found)
This means the workflow uses a custom node that is not installed in your ComfyUI — the graph references a node class ComfyUI has not registered, so the node shows up red and the graph will not run. It is a missing node package, not a broken workflow. (It is different from a ModuleNotFoundError, which is a node that IS installed but is missing a Python dependency.)
MetadataIncompleteBuffer / HeaderTooLarge (truncated safetensors)
This means the .safetensors file is truncated — the download did not finish. safetensors reads an 8-byte header-length prefix and then expects exactly that many bytes; when the file is cut short it fails at load with MetadataIncompleteBuffer or HeaderTooLarge. The file is on disk and looks present, but it is incomplete.
Value not in list on a text encoder
A "Value not in list" error on a CLIP/encoder loader means the encoder file the workflow expects is not in your text_encoders folder, so the dropdown cannot select it. On modern models this is almost always the large LLM/VL encoder people forget — newer models pair a small CLIP with a big language encoder, and shipping one short triggers this.
a .gguf model missing from the Unet loader
You downloaded a GGUF quant but UnetLoaderGGUF shows an empty dropdown or will not list it. This is almost always a folder problem: ComfyUI-GGUF reads a dedicated folder key, and a GGUF dropped in the wrong models subfolder will not appear.
WanVideoWrapper crash at load (fp16_fast / sageattention)
A Kijai WanVideoWrapper "advanced" graph that is supposed to be tested crashes at model load. The cause is usually a backend the graph assumes but your install does not have: base_precision=fp16_fast needs torch 2.7+, and attention_mode=sageattn needs the sageattention package. They are gated separately, so a graph moved between machines breaks when the backend is absent.
a deploy stuck installing custom nodes
A ComfyUI startup that hangs at "installing custom nodes" and never reaches model download is usually one custom node triggering a heavy from-source compile. Packages like llama-cpp-python and sam2 build native/CMake code, and with no timeout one doomed compile — or a pip resolver backtrack — blocks the whole install indefinitely.
LTX-2 gemma_path not in list (gated Gemma encoder)
LTX-2 (19B) uses a Gemma-3 LLM as its text encoder, loaded via LTXVGemmaCLIPModelLoader as a directory — and it shows an empty gemma_path list when that directory is missing the gated config + tokenizer files. The diffusion weights come from the open Comfy-Org/ltx-2 repo, but config.json and tokenizer.model come from the gated google/gemma-3-12b-it repo. Without an HF token + accepted license, that sidecar 403s and the loader has nothing to list.
DWPose / SAM2 preprocessor models not found
DWPose and SAM2 preprocessor nodes fail to find their models even though you downloaded them, because these nodes read from nested, node-specific folders that the generic models/ tree is not scanned for — and SAM2 also needs a specific re-serialized format, not Meta’s raw checkpoint.
Wan slow / low GPU use after a ComfyUI update
Wan (or any diffusion model) that was fast suddenly crawls at 0–20% GPU after a ComfyUI "Update All". The usual cause is that the update broke the compiled attention backend — sage/flash-attention was built against the old torch — so ComfyUI silently fell back to slow PyTorch attention. Low GPU utilization on a card that fits the model means it is CPU-bound, not out of VRAM.
Catch errors before they happen
Paste your workflow JSON into the free checker to see missing nodes, model requirements, and the GPU you need — before you deploy. No signup.
Check my workflow free