How to reduce inference time in Diffusers on consumer GPUs?

[nhhai71]

I'm running Diffusers on a standard consumer GPU (RTX 3060 / 6GB).
What are some practical ways to optimize inference speed without sacrificing too much quality?

[nguyenhoanghai1605]

There are several ways you can speed up inference in Diffuser on the user's GPU:

1. Use fp16 (half-precision):

pipe = StableDiffusionPipeline.from_pretrained("model_name", torch_dtype=torch.float16)
pipe.to("cuda")

2. Enable xFormers memory-efficient attention:
   pipe.enable_xformers_memory_efficient_attention()

3. Use a smaller model (like stabilityai/sd-turbo or stabilityai/sdxl-turbo):
   These are optimized for speed at the cost of some image quality.

4. Reduce steps (try 20–30 instead of 50):
   pipe(prompt, num_inference_steps=25)

[nhhai71]

Thanks!