ComfyUI Error Decoder
Paste a redacted ComfyUI or PyTorch log to classify the first causal error, extract CUDA OOM telemetry, identify the likely workflow stage, and get an ordered fix plan.
Runs locally in your browser. The pasted log is not uploaded by this tool.
Paste a redacted log to begin.
What the decoder checks
The classifier distinguishes CUDA out-of-memory, unavailable CUDA, failed custom-node imports, missing model paths, corrupt safetensors headers, NaN/black output, and tensor-shape mismatches. For PyTorch OOM messages it parses requested, free, process-used, allocated, and reserved-but-unallocated memory when those values are present.
Why it does not guess a universal fix
The final traceback line can be a cascade. The same “CUDA out of memory” text may happen while loading a model, sampling, decoding a VAE, or processing video frames; each stage changes the useful next step. The tool reports the evidence it matched and says “unknown” when the pasted text cannot support a diagnosis.
Privacy and limitations
Analysis is deterministic and local. It cannot inspect your workflow graph, GPU processes, model checksums, installed node versions, or live peak memory. Redact usernames, file paths, prompts, access tokens, and private model names before sharing a result.
Detailed guides
- Fix ComfyUI CUDA out of memory
- Fix black FLUX output in ComfyUI
- VRAM pressure and OOM headroom planner