ComfyUI Not Working? Common Errors and How to Fix Them

ComfyUI is a powerful tool, but like any complex software it can behave unexpectedly — particularly when combining custom nodes, different model types, and varying hardware configurations. Most problems that users encounter have known causes and straightforward fixes. This guide covers the most common ComfyUI errors and how to resolve them, from GPU memory issues and black output images to nodes showing red error states and the server refusing to start.

Out of Memory and CUDA OOM Errors

An out-of-memory (OOM) error means your GPU ran out of VRAM during generation. The error message will typically include something like CUDA out of memory or RuntimeError: CUDA error. This is one of the most frequent issues, especially on GPUs with 8 GB of VRAM or less when working with SDXL or Flux models.

Solutions

Launch ComfyUI with the --lowvram flag to enable a mode that offloads model components to system RAM when not actively in use:

python main.py --lowvram

If --lowvram is not sufficient, try --novae (also written as --cpu-vae in some versions). This runs the VAE decode step on the CPU rather than the GPU, which can save several hundred megabytes of VRAM at the cost of slower decoding:

python main.py --lowvram --cpu-vae

Reduce your output resolution. SDXL’s native resolution is 1024×1024, and generating above this consumes significantly more VRAM. Try 768×768 or 896×896 as a starting point. You can upscale afterwards using a separate upscaling step.

Reduce the batch size if you are generating multiple images simultaneously. Batch size multiplies VRAM usage almost linearly.

Black or Grey Output Images

Generating an image only to receive a completely black or flat grey result is a confusing error because ComfyUI does not usually display an explicit error message — it simply completes the generation with a useless output.

Wrong VAE

The most common cause is a mismatched or corrupted VAE (Variational Autoencoder). The VAE is the component responsible for decoding the latent representation into a visible image. Many checkpoints have a built-in VAE, but some do not, and using the wrong external VAE will produce black images.

Try loading a known-good VAE file via a Load VAE node and connecting it explicitly to your KSampler or VAEDecode node. For SD 1.5 models, vae-ft-mse-840000-ema-pruned.safetensors is reliable. For SDXL, use the SDXL VAE from Stability AI.

CFG Scale Too High for Flux

Flux models behave differently from SD 1.5 and SDXL. Flux uses a guidance scale of 1.0 to 3.5. If you load a Flux model into a workflow designed for SD 1.5 with a CFG value of 7.0 or higher, you will typically get a grey or overexposed image. Reduce the CFG (also called guidance scale) to between 2.0 and 3.5 when using Flux.

Model Not Found Errors

If ComfyUI reports that it cannot find a model file, the most likely cause is that the file is in the wrong folder. ComfyUI looks for models in specific subdirectories:

• Checkpoints: ComfyUI/models/checkpoints/
• LoRAs: ComfyUI/models/loras/
• VAE files: ComfyUI/models/vae/
• ControlNet models: ComfyUI/models/controlnet/
• Upscale models: ComfyUI/models/upscale_models/

If you moved a file into the correct folder while ComfyUI was already running, click the refresh icon in the model selector node or restart ComfyUI. The file list is not updated live in older versions.

Nodes Showing Red or Error State

A red node means it failed to load or execute. There are two distinct causes: load-time errors (the node could not be imported when ComfyUI started) and runtime errors (the node failed during a generation).

Missing Custom Node or Dependency

If a workflow you downloaded uses a custom node you have not installed, that node will appear as a red placeholder. Open ComfyUI Manager, go to Install Missing Custom Nodes, and Manager will detect and offer to install the missing nodes automatically.

If the node is installed but still shows red at startup, check the terminal output for Python import errors. The most common fix is activating the ComfyUI virtual environment and installing the node’s requirements:

source /path/to/ComfyUI/venv/bin/activate
pip install -r /path/to/ComfyUI/custom_nodes/node-name/requirements.txt

ComfyUI Will Not Start

Port Already in Use

ComfyUI defaults to port 8188. If another process is already using this port, ComfyUI will fail to start. Find and kill the process using the port, or start ComfyUI on a different port:

python main.py --port 8189

Python Path or Version Issues

If you see ModuleNotFoundError or similar import errors at startup, ensure you have activated the correct virtual environment before running ComfyUI. Running with the system Python rather than the venv Python is a common mistake:

source /path/to/ComfyUI/venv/bin/activate
python main.py

Slow Generation Speed

If generation is unexpectedly slow, the most common cause is that ComfyUI is not using your GPU. Check the terminal output when a generation runs — it should mention CUDA or your GPU. If you see it running on CPU, reinstall PyTorch with the correct CUDA index URL for your CUDA version.

On Windows, exclude the ComfyUI folder from Windows Defender real-time scanning. Antivirus software scanning model files during generation can halve generation speed or worse.

Another common culprit is the number of steps. SD 1.5 workflows often default to 20 steps, which is reasonable. SDXL and Flux workflows are sometimes shared with step counts of 30 or 40 — reducing to 20–25 steps significantly cuts generation time with minimal quality loss for most samplers. DPM++ 2M Karras at 20 steps is a reliable fast option for SD 1.5 and SDXL.

If you are on Linux and generation speed seems lower than expected compared to Windows benchmarks, ensure you are using the correct CUDA version and that torch.backends.cudnn.benchmark is enabled. ComfyUI handles this automatically in most cases, but a mismatch between the CUDA toolkit and the PyTorch CUDA version can reduce throughput.

Queue Stuck or Frozen

If a generation appears to be running indefinitely with no progress, the most reliable fix is to stop the ComfyUI server process (Ctrl+C in the terminal) and restart it. Before restarting, clear your browser cache and reload the interface — stale browser state occasionally causes the queue display to misrepresent what the server is actually doing.

If the queue freezes regularly during specific workflows, the issue may be a custom node that is hanging rather than failing visibly. Disable custom nodes one at a time via ComfyUI Manager to isolate the cause.

For a full index of every ComfyUI guide on Serverman, see the ComfyUI complete guide and hub.