VRAM Model Fit & OOM Headroom Calculator
Screen a ComfyUI workload for VRAM pressure or parse a real PyTorch OOM line without pretending that every model has one exact memory requirement.
Describe a scenario to begin.
Why this is a pressure screen, not a fake benchmark
Peak GPU memory depends on the full graph: loaded diffusion and text-encoder models, VAE, ControlNet and LoRA nodes, attention backend, ComfyUI/PyTorch versions, preview generation, offload behavior, canvas size, batch and frame count. A model name plus “12GB” cannot prove that a workflow fits.
Two useful modes
- Planning mode: describes relative pressure and the missing inputs that still require a real benchmark.
- Measured OOM mode: parses requested allocation, free memory, process use, PyTorch allocation and reserved memory from an actual error line.
Measure a real threshold
Start from batch 1 and a known-good workflow. Record the successful dimensions and peak memory, then change one setting at a time. If the failure happens during VAE decode, tiled decode can be more relevant than changing the sampler. If previews contribute to the peak, ComfyUI documents --preview-method none. Use --cpu only as a slow last resort.