CelesteImperia/celeste-imperia-sdxl-openvino

🌌 Celeste Imperia: SDXL OpenVINO Trinity

The Celeste Imperia project provides high-performance, hardware-optimized ports of Stable Diffusion XL 1.0. By mathematically fusing the Latent Consistency Model (LCM) weights and utilizing Intel's OpenVINO toolkit, this suite achieves high-quality 1024x1024 image generation in just 4 to 8 steps on standard consumer CPUs and integrated GPUs.

🚀 The Trinity Flavors

We provide three distinct bakes of the model to fit different hardware constraints:

Flavor	Folder Name	Format	Best For
Master	sdxl-lcm-openvino-fp16	FP16	High-end CPUs / ARC GPUs (Maximum Fidelity)
Pro	sdxl-lcm-openvino-int8	INT8	Balanced performance (Standard Core i7/i9)
Lite	sdxl-lcm-openvino-int4	INT4	Maximum speed (Laptops / Systems with < 16GB RAM)

✨ Key Features

• Fused LCM Engine: No need to load LoRAs separately; the speed is baked into the weights.
• TinyVAE Injection: All versions use a specialized TinyVAE for near-instant pixel decoding and reduced memory overhead.
• 4-Step Magic: Generate 1024px images in a fraction of the time required by standard SDXL.
• Intel Optimized: Tailored for Intel Core processors, Iris Xe, and ARC Graphics.

🛠️ Usage (via Optimum Intel)

To run these models, ensure you have optimum-intel and openvino installed: pip install "optimum[openvino,diffusers]"

Python Inference Code

from optimum.intel import OVStableDiffusionXLPipeline
from diffusers import LCMScheduler

# Choose your flavor: sdxl-lcm-openvino-fp16, sdxl-lcm-openvino-int8, or sdxl-lcm-openvino-int4
model_id = "CelesteImperia/celeste-imperia-sdxl-openvino"
flavor = "sdxl-lcm-openvino-int4" 

# Load the model
pipe = OVStableDiffusionXLPipeline.from_pretrained(
    model_id, 
    subfolder=flavor, 
    device="CPU"
)

# Essential: Use the LCM Scheduler for 4-step generation
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)

# IMPORTANT: guidance_scale MUST be low (1.0 - 2.0) for LCM
image = pipe(
    prompt="A breathtaking futuristic sci-fi city in the clouds, neon lights, highly detailed",
    num_inference_steps=4,
    guidance_scale=1.0,
    width=1024,
    height=1024
).images[0]

image.save("celeste_output.png")

⚙️ Recommended Settings

• Steps: 4 - 8 steps.
• Guidance Scale (CFG): 1.0 to 2.0. Values higher than 3.0 will result in artifacts.
• Resolution: 1024x1024 (Native SDXL size).
• Inference Device: Works excellently on CPU. Use GPU if you have Intel ARC or integrated Iris Xe.

🤝 Credits

• Base Model: SDXL 1.0 by Stability AI.
• Speed Engine: LCM-LoRA-SDXL by SimianLuoxue.
• Optimized by: @itsabhishek19 for CelesteImperia.

☕ Support My Work

I develop and port open-source AI models and tools for the community. If you find my work helpful, consider supporting the development and compute costs!

Platform	Support Link
Global & India	Support via Razorpay

---

Scan to support via UPI (India Only):