Test hardware and environments

Use this reference alongside the workload at the top of each model's benchmark page. Different hardware, drivers, cache state and memory limits can change both time and peak memory; these results are not performance guarantees.

Reference machine

Component Configuration
GPU NVIDIA GeForce RTX 3060 Ti
Dedicated GPU memory 8 GiB (8192 MiB)
CPU AMD Ryzen 5 5600X, 6 cores / 12 logical processors
Installed system RAM 64 GiB

The GPU and its capacity are recorded in the test records. CPU and installed RAM were confirmed on the reference Windows host on 2026-10-15; this inspection does not reconstruct a complete hardware inventory for every earlier run. Storage model, interface and measured disk throughput were not captured consistently, so no storage-equivalence claim is made.

Platform conditions

Platform Recorded environment and limits
Windows Same reference GPU; full host RAM rather than the WSL memory limit. Selected comparisons record successful standby purge before the transformer. Current host inspection identifies Windows 10 Enterprise, build 19045; not every result recorded its exact OS build.
Native Linux Ubuntu 26.04.1 x86_64 is documented in the native-Linux environment records. Selected protected NVMe observations use NVIDIA driver 595.99.02 and a different reboot/cache session from the earlier leaf/linear/fast-VMM comparisons.
WSL2 Ubuntu 22.04.3, kernel 6.18.33.2-microsoft-standard-WSL2; approximately 39 GiB guest RAM and 32 GiB configured swap. GPU access uses the Windows driver. Guest RAM is not the machine's full 64 GiB, and configured swap is not evidence that a run used it.

Windows and WSL share the physical GPU, not an identical execution environment. Native Linux is a separate target. Exact driver versions were not recorded uniformly across all rows; do not assign one version to the entire comparison. Primary comparisons use official Diffusers 0.41.0; software and component-route exceptions are described in each model's measurement details.

Match your comparison

Record your GPU/VRAM, CPU, installed RAM, OS, driver, storage and any VM/container memory limits. Then match resolution, precision, steps, seed, conditioning, encoder/VAE routes and runtime capacities before comparing results.

The model pages state resolution and generation settings separately for primary and supplementary workloads. Compare within the matching group, not across different resolutions, video durations or quantization formats.