ComfyUI FAQ: Common Problems Solved

Q: My workflow shows "Missing Node" errors. How do I fix this?

Missing nodes mean the workflow requires custom nodes that are not installed. In Thumper-Run, open the Custom Nodes panel and search for the missing node name. Most popular nodes are available for one-click install. If you loaded a workflow from the internet, check the author's description for required node packs.

See the ComfyUI guide for managing custom nodes.

Q: I get "CUDA out of memory" errors. What do I do?

Your model does not fit in your available VRAM. Options:

• Reduce resolution – generate at 768×768 instead of 1024×1024
• Use a lighter model – SD 1.5 instead of SDXL
• Enable tiled VAE – processes the image in chunks instead of all at once
• Close other GPU apps – browsers, games, and video players use VRAM

See the VRAM guide for per-model requirements.

Q: Generation is very slow. How do I speed it up?

Common causes of slow generation:

• Wrong PyTorch build – if PyTorch does not detect your GPU, it falls back to CPU. Check with nvidia-smi or rocm-smi.
• No xformers – xformers provides memory-efficient attention. Thumper installs it automatically.
• High step count – reduce from 50 to 20–30 steps. Quality loss is minimal.
• Large batch size – generate one image at a time if VRAM is tight.

Q: How do I use LoRA weights in ComfyUI?

Add a Load LoRA node between the checkpoint loader and the CLIP encoder. Point it at your LoRA file and set the strength (0.5–1.0). You can stack multiple LoRAs by chaining Load LoRA nodes.

Thumper's model management stores LoRAs in a shared location that ComfyUI auto-detects.

Q: My images have artifacts or look blurry. Why?

• Wrong resolution – each model has a native resolution. SDXL works best at 1024×1024, SD 1.5 at 512×512.
• Low CFG scale – try 7–8 for most models.
• Bad VAE – ensure you are using the correct VAE for your checkpoint. Wrong VAE produces washed-out colors.
• Too few steps – minimum 20 steps for decent quality.

Q: How do I use ControlNet?

1. Install the ControlNet Auxiliary Preprocessors custom node pack
2. Add a preprocessor node (Canny, Depth, Pose) connected to your reference image
3. Add an Apply ControlNet node connected to the preprocessor output and your KSampler
4. Adjust strength (0.5–0.8 is a good starting range)

See the ComfyUI guide for detailed ControlNet workflows.

Q: Can I run ComfyUI on CPU only?

Yes, but it is very slow. A single 512×512 image takes 5–15 minutes on CPU vs 5–15 seconds on GPU. CPU mode is useful for testing workflows before committing to a full render. Add --cpu to the launch arguments or let Thumper's GPU detection handle it.

Q: How do I share workflows with others?

ComfyUI workflows are JSON files. You can:

• Export from the ComfyUI menu (Save as JSON)
• Share the JSON file directly
• Import by dragging the JSON onto the ComfyUI canvas

If the recipient does not have the required custom nodes or models, they will see errors. Include a list of dependencies when sharing. Thumper's workflow library bundles workflows with their requirements.

Ready to try it? Download Thumper-Run free →