Practical research · explicit evidence
ComfyUI RTX performance: fix slow runs and VRAM failures
Start with the stage that fails or takes longest. Model loading, sampling and decoding need different fixes. This guide gives you a small comparison you can run on your own RTX system, with a record sheet for results.
Locate the bottleneck first
Save your existing workflow before changing it. Run one representative job and note when the wait or error occurs. Keep the final output dimensions and acceptance criteria visible: making a smaller image may be useful, but it is a different workload.
| Observed problem | First comparison | What the result tells you |
|---|---|---|
| Long wait before sampling begins | Repeat the same graph once after models have loaded. Record the first and second run separately. | A large difference points toward setup or loading overhead; it does not prove the sampler improved. |
| Out of memory during sampling | Try batch size 1, keeping model, resolution, seed and sampler fixed. | Success shows that this smaller batch fits. It does not promise the same result for another model. |
| Failure after sampling finishes | Identify whether VAE decoding, an upscaler or saving is executing. Test the failing stage separately where the graph permits. | A decode failure should not be diagnosed solely by changing the number of sampling steps. |
| Slow execution with memory pressure | Compare a documented low-memory launch option with your original launch. | Record both completion time and memory. Offloading can trade speed for capacity. |
| Failure only with an extension | Use the official custom-node troubleshooting route and preserve the error report. | A smaller graph can isolate an extension dependency before you replace the entire environment. |
A practical, repeatable comparison
- Choose one output that matters to you. Save its graph, exact model filenames and hashes, seed, dimensions, batch, steps, sampler and scheduler.
- Record your GPU, driver, operating system, Python and PyTorch versions, ComfyUI commit and custom-node commits. Keep these fixed for the comparison.
- Record the cold run separately. Do one warm-up, then record five repeated attempts under the baseline condition, including failures.
- Change one setting. Restart between launch-option conditions, warm up again and repeat the five attempts. Use identical timing boundaries.
- Compare successful outputs visually against the same acceptance rule. Report the median and range alongside failures; do not keep only the fastest run.
For a manual installation, the maintained source exposes these separate options. Confirm them against python main.py --help in your installed version. Desktop and packaged launchers may configure them differently.
python main.py --lowvram
python main.py --preview-method noneThese are two alternative experimental conditions, not a command sequence to run simultaneously. Keep the original launch command so you can return to the baseline. Do not install an attention package simply because its name appears in another person’s graph.
Record memory without mixing metrics
On NVIDIA systems, a device log can help associate memory pressure with execution stages. This read-only example samples once a second until stopped:
nvidia-smi --query-gpu=timestamp,name,memory.used,utilization.gpu --format=csv -l 1That is sampled device memory, including other GPU work. A short peak between samples may be missed. PyTorch’s peak allocation describes its tensor allocator and must be measured inside the process running the workload. Running a separate Python process does not measure ComfyUI’s allocator peak.
Download the blank run-record CSV Full benchmark methodology
The CSV contains column headings only. It is a recording template, not a tested workflow or measured result.
A useful conclusion is specific: “Condition B completed five attempts of this graph with these versions; its median was lower, and all outputs passed our stated quality check.” Until those records exist, the outcome is unmeasured.
Need capacity rather than speed? Read the RTX 4090 capacity decision guide.