A set of benchmarks targeting different stable diffusion implementations to have a better understanding of their performance and scalability.
Running on an A100 80G SXM hosted at fal.ai. If you want to see how these models perform first hand, check out the Fast SDXL playground which offers one of the most optimized SDXL implementations available (combining the open source techniques from this repo).
Note
Most of the implementations here are also based on Diffusers, which is an amazing library that pretty much the whole industry is using. However, when we use 'Diffusers' name in the benchmarks, it means the experience you might get with out-of-box Diffusers (w/applying necessary settings).
| mean (s) | median (s) | min (s) | max (s) | speed (it/s) | |
|---|---|---|---|---|---|
| Diffusers (torch 2.1, SDPA) + OpenAI's consistency decoder** | 2.230s | 2.229s | 2.220s | 2.238s | 22.43 it/s |
| Diffusers (torch 2.1, xformers) | 1.729s | 1.728s | 1.720s | 1.747s | 28.94 it/s |
| Diffusers (torch 2.1, SDPA) | 1.604s | 1.603s | 1.589s | 1.618s | 31.19 it/s |
| Diffusers (torch 2.1, SDPA, tiny VAE)* | 1.567s | 1.562s | 1.547s | 1.602s | 32.02 it/s |
| Diffusers (torch 2.1, SDPA, compiled) | 1.354s | 1.354s | 1.351s | 1.356s | 36.93 it/s |
| Diffusers (torch 2.1, SDPA, compiled, NCHW channels last) | 1.058s | 1.057s | 1.056s | 1.060s | 47.29 it/s |
| OneFlow | 0.950s | 0.950s | 0.947s | 0.961s | 52.65 it/s |
| Stable Fast (torch 2.1) | 0.901s | 0.901s | 0.900s | 0.903s | 55.51 it/s |
| TensorRT 9.0 (cuda graphs, static shapes) | 0.819s | 0.818s | 0.817s | 0.821s | 61.14 it/s |
| mean (s) | median (s) | min (s) | max (s) | speed (it/s) | |
|---|---|---|---|---|---|
| minSDXL (torch 2.1) | 8.146s | 8.146s | 8.137s | 8.155s | 6.14 it/s |
| Diffusers (torch 2.1, SDPA) | 5.932s | 5.932s | 5.924s | 5.940s | 8.43 it/s |
| minSDXL+ (torch 2.1, SDPA) | 5.887s | 5.887s | 5.872s | 5.897s | 8.49 it/s |
| Comfy (torch 2.1, xformers) | 5.779s | 5.772s | 5.748s | 5.824s | 8.66 it/s |
| Diffusers (torch 2.1, SDPA, tiny VAE)* | 5.739s | 5.738s | 5.722s | 5.767s | 8.71 it/s |
| Diffusers (torch 2.1, xformers) | 5.719s | 5.717s | 5.710s | 5.732s | 8.75 it/s |
| minSDXL+ (torch 2.1, flash-attention v2) | 5.323s | 5.322s | 5.313s | 5.340s | 9.39 it/s |
| Diffusers (torch 2.1, SDPA, compiled) | 5.217s | 5.216s | 5.213s | 5.220s | 9.59 it/s |
| Diffusers (torch 2.1, SDPA, compiled, NCHW channels last) | 5.136s | 5.137s | 5.125s | 5.147s | 9.73 it/s |
| OneFlow | 4.300s | 4.301s | 4.282s | 4.316s | 11.62 it/s |
| Stable Fast (torch 2.1) | 4.150s | 4.149s | 4.138s | 4.168s | 12.05 it/s |
| TensorRT 9.0 (cuda graphs, static shapes) | 4.102s | 4.104s | 4.091s | 4.107s | 12.18 it/s |
Generation options:
prompt="A photo of a cat"num_inference_steps=50- For SD1.5, the width/height is 512x512 (the default); for SDXL, the width/height is 1024x1024.
- For all other options, the defaults from the generation systems are used.
- Weights are always half-precision (fp16) unless otherwise specified.
- Generation on benchmarks with a
*/**means the used techniques might lead to quality degradation (or sometimes improvements) but the underlying diffusion model is still the same.
Note
All the timings here are end to end, and reflects the time it takes to go from a single prompt to a decoded image. We are planning to make the benchmarking more granular and provide details and comparisons between each components (text encoder, VAE, and most importantly UNET) in the future, but for now, some of the results might not linearly scale with the number of inference steps since cost of certain components are one-time only.
Environments (like torch and other library versions) for each benchmark are defined under benchmarks/ folder.
