bakhil-aissa/3d_interior_openvino_8bit

SD1.5 Interior LoRA + LCM-style Model

This repository contains a fine-tuned Stable Diffusion v1.5 model with interior design LoRA adapters merged. The model has been prepared for OpenVINO inference with optional quantization.

---

Model Overview

• Base model: Stable Diffusion v1.5

• Adapters applied:

  • UNet LoHa adapter for interior design
  • Text Encoder LoRA adapter for interior design

• Merged: All adapters merged into SD1.5 UNet and Text Encoder weights

• LCM-style: Optional LCM-LoRA applied for latent consistency guidance

• Quantization: INT8 quantization supported for OpenVINO CPU acceleration

---

Features

• Generates high-quality interior renderings of rooms, kitchens, and living spaces.
• Preserves your interior style across prompts.
• Supports OpenVINO INT8 acceleration.
• Works with both GPU and CPU (FP16/INT8).

---

Installation

pip install torch diffusers peft optimum[intel] transformers safetensors

Optional: For OpenVINO inference:

pip install openvino

---

Usage

Load pipeline (OpenVINO INT8)

from optimum.intel import OVStableDiffusionPipeline
from openvino.runtime import Core

core = Core()
core.set_property({"CACHE_DIR": "ov_cache"})
core.set_property({"INFERENCE_PRECISION_HINT": "int8"})
core.set_property({"PERFORMANCE_HINT": "LATENCY"})
core.set_property({"NUM_STREAMS": "1"})

pipe_ov = OVStableDiffusionPipeline.from_pretrained("bakhil-aissa/3d_interior_openvino_8bit", compile=True)

image = pipe_ov(prompt, negative_prompt=negative_prompt, num_inference_steps=16, guidance_scale=1.0, height=512, width=512).images[0]
image.save("output_ov.png")

---

Notes

• LCM-style LoRA applied for latent consistency; improves image coherence in fewer steps.
• INT8 quantization requires representative calibration images.
• Merged LoRA weights preserve interior style, but may reduce speed on CPU if not quantized.
• Recommended image sizes: 512×512 or 768×768. Higher resolutions may be slow.

---

Examples

Disclaimer: This model is intended for research and personal use only. Generated images may contain artifacts. Use responsibly.