RTX 5060 Ti 16GB Stable Diffusion Setup

The simplest SD setup is HuggingFace Diffusers in a Python venv. Good for custom scripts or API wrappers on the RTX 5060 Ti 16GB at our hosting.

Contents

• Install
• First generation
• FLUX.1
• Expose as API

Install

uv venv --python 3.12 ~/.venvs/sd
source ~/.venvs/sd/bin/activate
uv pip install torch --index-url https://download.pytorch.org/whl/cu126
uv pip install diffusers transformers accelerate safetensors

First Generation (SDXL)

import torch
from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    use_safetensors=True,
    variant="fp16",
).to("cuda")

image = pipe(
    prompt="a golden retriever wearing glasses, hyperreal, studio lighting",
    num_inference_steps=30,
).images[0]

image.save("output.png")

Expected time: ~3-4 seconds for 1024×1024.

FLUX.1-schnell

from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-schnell",
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    prompt="a neon-lit cyberpunk street at night",
    num_inference_steps=4,
    guidance_scale=0.0,
).images[0]
image.save("flux.png")

Expected time: ~3.8 s at FP16. For FP8 quantised weights see FLUX benchmark.

Expose as API

from fastapi import FastAPI
from pydantic import BaseModel

app = FastAPI()

class PromptIn(BaseModel):
    prompt: str
    steps: int = 30

@app.post("/generate")
def generate(p: PromptIn):
    img = pipe(prompt=p.prompt, num_inference_steps=p.steps).images[0]
    # save or return base64
    ...

For production API, ComfyUI or A1111 give you the UI free – see ComfyUI setup or A1111 setup.

See also: ComfyUI, A1111, SDXL benchmark, FLUX benchmark.