Stable Diffusion 1.5 is 5+ years old but still massively used because of LoRA ecosystem depth and raw speed. On the RTX 5060 Ti 16GB at our hosting, SD 1.5 is effectively CPU-bound – the GPU is barely working.
Contents
Setup
- Diffusers 0.30, PyTorch 2.5, xFormers
- Model: runwayml/stable-diffusion-v1-5
- Sampler: DPM++ 2M Karras
512×512
| Steps | Time | VRAM |
|---|---|---|
| 20 | 0.65 s | 3.0 GB |
| 30 | 0.95 s | 3.0 GB |
| 50 | 1.55 s | 3.0 GB |
768×768 (via img2img or fine-tuned checkpoints)
| Steps | Time | VRAM |
|---|---|---|
| 30 | 2.2 s | 4.5 GB |
Batch Throughput
512×512, 30 steps:
| Batch | Total time | Time per image |
|---|---|---|
| 1 | 0.95 s | 0.95 s |
| 4 | 1.8 s | 0.45 s |
| 8 | 3.1 s | 0.39 s |
| 16 | 5.4 s | 0.34 s |
| 24 | 7.8 s | 0.33 s |
At batch 24, throughput = ~180 images per minute. Aggregate rate on 16 GB is outstanding.
When to Use SD 1.5
- Anime / stylised art where you have specific LoRAs / Dreambooth models
- High-volume thumbnail generation
- Interactive tools where 0.4s/image matters
- Fine-tuning where SD 1.5’s lightweight size speeds training
For new projects at photorealism, skip to FLUX.1 or SDXL. For existing LoRA-heavy workflows, SD 1.5 still has no equal on speed.
SD 1.5 on Blackwell 16GB
180 images/min at 512, LoRAs galore. UK dedicated hosting.
Order the RTX 5060 Ti 16GBSee also: SDXL benchmark, FLUX.1 schnell, SD setup, A1111 setup, ComfyUI setup.
