Fix ComfyUI Model Load Failed
Troubleshoot ComfyUI model load failed errors including invalid header, shape mismatch, size mismatch, pickle data was truncated, file not found, and CUDA out of memory.
Quick answer
Use this guide when the model can be selected but ComfyUI fails while reading it or running the graph. If the dropdown is empty, go back to the empty-dropdown guide first.
Recommended workflow
- 01
Model load failed happens after discovery: Use this guide when the model can be selected but ComfyUI fails while reading it or running the graph. If the dropdown is empty, go back to the empty-dropdown guide first.
- 02
Use the error keyword table before changing settings: For invalid header, inspect the file size and redownload from a trusted source. A tiny .safetensors file is often an HTML page, failed download, or interrupted copy.
- 03
Use a 30-second load-failure pass: First confirm the file appears in the correct dropdown. Second read the last terminal keyword. Third check whether the file is plausible in size and extension. Fourth confirm model-family compatibility. Fifth reduce workload only if the keyword is CUDA out of memory.
- 04
Keep the repair chain linear: The stable order is: choose family, make the file appear in the correct dropdown, then debug load failures. This prevents the common beginner loop of downloading more models, moving folders randomly, and reinstalling ComfyUI without reading the actual error.
Full tutorial notes
Model load failed happens after discovery
Use this guide when the model can be selected but ComfyUI fails while reading it or running the graph. If the dropdown is empty, go back to the empty-dropdown guide first.
Read the terminal keyword before reinstalling. File not found, invalid header, shape mismatch, size mismatch, pickle data was truncated, and CUDA out of memory point to different causes.
- file not found: path or referenced filename problem.
- invalid header: damaged, incomplete, or wrong downloaded file.
- shape mismatch: SD1.5, SDXL, Flux, LoRA, or ControlNet family mismatch.
- CUDA out of memory: workload or VRAM pressure, not a folder problem.
Use the error keyword table before changing settings
For invalid header, inspect the file size and redownload from a trusted source. A tiny .safetensors file is often an HTML page, failed download, or interrupted copy.
For shape mismatch or size mismatch, check the whole model stack: checkpoint, LoRA, ControlNet, VAE, and workflow notes. A prompt cannot make an SD1.5 adapter compatible with an SDXL checkpoint.
- No such file or directory -> verify real paths and extra_model_paths.yaml.
- pickle data was truncated -> redownload or recopy the file.
- expected channel errors -> inspect workflow connections and model family.
- CUDA out of memory -> reduce resolution, batch, upscale, ControlNet branches, or model family.
Use a 30-second load-failure pass
First confirm the file appears in the correct dropdown. Second read the last terminal keyword. Third check whether the file is plausible in size and extension. Fourth confirm model-family compatibility. Fifth reduce workload only if the keyword is CUDA out of memory.
This keeps file corruption, family mismatch, and VRAM pressure separate. If the model file is corrupt, lowering resolution will not help. If the workflow is too heavy, moving folders will not help.
Keep the repair chain linear
The stable order is: choose family, make the file appear in the correct dropdown, then debug load failures. This prevents the common beginner loop of downloading more models, moving folders randomly, and reinstalling ComfyUI without reading the actual error.
Check before you run
- Confirm the model already appears in the correct dropdown before using this guide.
- Read the terminal keyword: file not found, invalid header, shape mismatch, or CUDA out of memory.
- Separate file damage, family mismatch, and VRAM pressure before reinstalling anything.
Common mistakes
- Calling every load error a broken model file.
- Redownloading files before reading the terminal traceback.
- Treating OOM as a path problem instead of reducing workload.
Success standard
- The same model loads without invalid header or shape mismatch.
- The workflow reaches sampling instead of stopping at model loading.
- If VRAM fails, the error changes after lowering size, batch, or model family.
What to do next
- Check SD1.5 vs SDXL when shape mismatch appears.
- Check model-dropdown-empty if the model cannot be selected at all.
- Check CUDA out of memory if the model loads but the workload is too large.
Need more context?
This English guide gives the direct working path first. The paired Chinese reference can provide extra screenshots, local download notes, and longer troubleshooting branches for the same topic.