# ComfyUI low-VRAM benchmark methodology v1.0

Status: methodology published; no qualifying cross-hardware result rows yet.  
Reviewed: 2026-08-05  
Canonical HTML: https://42.uk/comfyui-low-vram-benchmarks.html

## Contract

1. Freeze the workflow, models, seed, prompts, dimensions, batch size, steps, sampler, scheduler and output format.
2. Record SHA-256 hashes for workflow, models and output plus immutable ComfyUI and custom-node commits.
3. Capture GPU identity and installed memory in bytes, driver/backend, operating system, system RAM, Python and PyTorch versions, launch command and memory-affecting environment variables.
4. Compare one declared change with the baseline. Supported run families are baseline, `--lowvram`, `--preview-method none`, one reduced-workload condition and `--cpu` as a compatibility fallback.
5. Run one unreported warm-up and five measured repetitions. Restart ComfyUI between run families; record all failures.
6. Report device-memory peak and PyTorch allocator peak as separate metrics, including the device sampling interval.
7. Publish every raw sample, the median and range, success rate, logs, artefact hashes and the exact outcome check.
8. Label fewer than three successful repetitions as incomplete. Call a row reproduced only after a second clean run reaches the same outcome category.

Hardware tiers are organizational labels, not performance promises: A ≤4 GiB; B >4–8 GiB; C >8–12 GiB; D >12–24 GiB; E >24 GiB. GiB means 1,073,741,824 bytes.

## Primary sources

- ComfyUI troubleshooting and memory flags: https://docs.comfy.org/troubleshooting/overview
- ComfyUI system requirements: https://docs.comfy.org/installation/system_requirements
- PyTorch `max_memory_allocated`: https://docs.pytorch.org/docs/stable/generated/torch.cuda.max_memory_allocated.html
- NVIDIA System Management Interface: https://docs.nvidia.com/deploy/nvidia-smi/index.html