Benchmark results
Median seconds per image on rented home GPUs, measured 2026-10-23. Same prompt and seeds on every machine. * = fewer than 5 machines, reference only. Download CSV · See the images · Method
Speed by GPU and model
Seconds per 1024×1024 image, warm. Lower is faster. Small text = machines measured. Rows are ordered by SDXL speed.
| GPU | SDXL | FLUX.2 klein | FLUX.3 klein | FLUX.1 schnell |
|---|
| RTX 5070 12GB | 8.8 n=5 | 3.3* n=2 | 11.8* n=1 | 7.9* n=2 |
| RTX 4070 12GB | 9.5 n=5 | 3.1* n=2 | — | 7.2* n=1 |
| RTX 5060 8GB | 13.9 n=5 | — | — | — |
| RTX 4060 8GB | 18.0* n=4 | — | — | — |
| RTX 3070 8GB | 18.6* n=1 | — | — | — |
| RTX 3060 12GB | 21.0 n=6 | — | — | 14.1* n=2 |
Details per model
Stable Diffusion XL 1.0 (base)
1024×1024, 30 steps, euler / normal, cfg 7
| GPU | Median s/image | Min–max | Nodes | Peak VRAM / total GB | Cost per image | First image |
|---|
| RTX 5070 12GB | 8.80 | 8.51–8.91 | 5 | 11.1 / 12 | $0.00040 | 72 s |
| RTX 4070 12GB | 9.52 | 9.40–11.12 | 5 | 11.0 / 12 | $0.00040 | 636 s |
| RTX 5060 8GB | 13.91 | 13.58–15.03 | 5 | 7.5 / 8 | $0.00043 | 82 s |
| RTX 4060 8GB | 18.01* | 17.70–18.73 | 4 | 7.5 / 8 | $0.00040 | 286 s |
| RTX 3070 8GB | 18.59* | 18.59–18.59 | 1 | — | $0.00052 | 55 s |
| RTX 3060 12GB | 20.98 | 20.45–22.01 | 6 | 10.1 / 12 | $0.00047 | 72 s |
FLUX.2 [klein] 4B (distilled)
1024×1024, 4 steps, euler, cfg 1
| GPU | Median s/image | Min–max | Nodes | Peak VRAM / total GB | Cost per image | First image |
|---|
| RTX 4070 12GB | 3.06* | 2.92–3.21 | 2 | 11.5 / 12 | $0.00013 | 156 s |
| RTX 5070 12GB | 3.30* | 3.18–3.43 | 2 | 11.2 / 12 | $0.00015 | 92 s |
FLUX.3 [klein] (distilled, bf16)
1024×1024, 15 steps, er_sde, cfg 1
| GPU | Median s/image | Min–max | Nodes | Peak VRAM / total GB | Cost per image | First image |
|---|
| RTX 5070 12GB | 11.82* | 11.82–11.82 | 1 | 11.4 / 12 | $0.00054 | 186 s |
FLUX.1 [schnell] (fp8)
1024×1024, 4 steps, euler / simple, cfg 1
| GPU | Median s/image | Min–max | Nodes | Peak VRAM / total GB | Cost per image | First image |
|---|
| RTX 4070 12GB | 7.22* | 7.22–7.22 | 1 | 11.6 / 12 | $0.00030 | 77 s |
| RTX 5070 12GB | 7.87* | 6.53–9.21 | 2 | 11.5 / 12 | $0.00036 | 58 s |
| RTX 3060 12GB | 14.06* | 13.91–14.21 | 2 | 11.2 / 12 | $0.00031 | 792 s |
Raw GPU speed
Our own short PyTorch test (matrix multiplication, attention, memory copy), mean of the machines listed. Machines of the same card differ by up to ~50% on some numbers; see the note.
| GPU | Machines | fp16 matmul TFLOPS | fp32 matmul TFLOPS | Attention fp16 TFLOPS | Memory GB/s | SDXL s/image |
|---|
| RTX 5070 12GB | 3 | 63.6 | 19.8 | 66.0 | 576 | 8.8 |
| RTX 4070 12GB | 3 | 62.8 | 18.1 | 59.7 | 438 | 9.5 |
| RTX 5060 8GB | 3 | 39.1 | 11.9 | 39.8 | 387 | 13.9 |
| RTX 4060 8GB | 3 | 31.0 | 7.7 | 30.7 | 247 | 18.0 |
| RTX 3070 8GB | 2 | 36.7 | 9.7 | 37.5 | 421 | 18.6 |
| RTX 3060 12GB | 3 | 25.6 | 6.6 | 24.1 | 340 | 21.0 |
Square vs portrait
Same machine, same model, same prompt. Portrait has about the same pixel count as square.
| Model | GPU | Square 1024×1024 s | Portrait 832×1216 s | Portrait vs square |
|---|
| SDXL 1.0 | RTX 5070 12GB | 12.26 | 11.43 | -7% |
| FLUX.2 klein | RTX 5070 12GB | 4.55 | 4.37 | -4% |
Numbers are rough: rented machines are owned by individuals and vary. Cost uses the rental hourly price at the time. Peak VRAM is the nvidia-smi value and includes memory PyTorch reserves.
