Understanding and Fixing Workflow Errors

Solve the most common workflow loading and execution problems

#Common Workflow Errors

Error: "Node not found in workflow"

Cause: Missing custom node

Fix:

1. →Install ComfyUI Manager
2. →Load workflow → Manager will show "Install Missing Nodes"
3. →Click install → Restart

Error: "Input type mismatch"

Cause: Wrong model type for workflow

Example:

• →SDXL workflow + SD 1.5 checkpoint = fails

Fix: Check workflow description for required model type

Quick reference:

• →SD 1.5 workflows → Use SD 1.5 checkpoints
• →SDXL workflows → Use SDXL checkpoints
• →Flux workflows → Use Flux models

Error: Workflow loads but produces black images

Causes:

1. →Wrong VAE
2. →Incompatible sampler
3. →Resolution too high

Fixes:

1. →Add VAE Loader node with vae-ft-mse-840000-ema-pruned.safetensors
2. →Try different sampler (euler, dpm++ 2m)
3. →Reduce resolution to 512×512 (SD1.5) or 1024×1024 (SDXL)

Error: "VRAM out of memory" mid-generation

Fix: See the GPU errors guide.

Quick solution:

• →Reduce resolution
• →Add --lowvram flag
• →Install Tiled VAE node

#Workflow Best Practices

1. →Always use .safetensors models (safer than .ckpt)
2. →Match LoRAs to base model version
3. →Save workflows frequently
4. →Test with default workflow first before loading complex ones
5. →Check custom nodes are installed before importing workflows

#Related Guides

• →Model folders
• →Custom nodes
• →Model types

#Capture a useful failure record

Before reinstalling or changing several packages, save the full error text and the workflow JSON. Record the ComfyUI revision, Python and PyTorch versions, GPU/driver, model filename, custom-node revision, input dimensions, batch size, and the last node that completed. This separates a missing dependency from a model-path mismatch or a memory failure.

#Safe recovery order

1. →Reopen the original graph and identify the first missing or failed node.
2. →Verify the exact model filename and folder expected by that node.
3. →Confirm the custom-node repository and version from its upstream instructions.
4. →Retry with a small input and batch size of one.
5. →Change only one dependency or setting per retry, retaining the error output if it fails.

These steps do not prove that a workflow is compatible with your hardware. They make the next troubleshooting decision observable and reproducible.