README.md

---
license: mit
library_name: diffusers
language:
- en
base_model:
- Layasaran/pixelora-1.0-xhigh
pipeline_tag: text-to-image
---

# 🌌 Pixelora-1.0-xhigh

Pixelora-1.0-xhigh is a high-performance, distilled latent diffusion engine engineered for **ultra-photorealistic** image synthesis. By leveraging advanced distillation techniques, this model achieves cinematic-quality results in just **4 to 8 sampling steps**, making it ideal for real-time production environments and high-throughput workflows.



---

## 🌟 Key Features
* **Zero-Shot Photorealism:** Specialized in rendering hyper-detailed skin pores, fabric textures, and complex lighting without "over-sharpening."
* **Lightning Fast:** Optimized for inference on mid-tier hardware (like 2xT4 or RTX 30-series) with a ~90% reduction in generation time compared to standard models.
* **Advanced Prompt Adherence:** High sensitivity to technical photography terms (e.g., focal length, aperture, film stock).
* **Balanced Latents:** Minimized "AI artifacts" and improved anatomical consistency in low-step counts.

---

## ⚙️ Technical Specifications

To achieve the intended aesthetic, please adhere to the following inference parameters:

| Parameter | Recommended Setting |
| :--- | :--- |
| **Resolution** | 1024 x 1024 (Native), 832 x 1216 (Portrait) |
| **Sampling Steps** | 4 — 6 steps (Sweet spot: 5) |
| **Guidance Scale (CFG)** | 1.0 — 2.0 (Strictly) |
| **Sampler** | `DPM++ SDE Karras` or `Euler A` |
| **VAE** | Use built-in SDXL VAE |

> **Note:** Setting the Guidance Scale above 2.0 may result in "burnt" images or color banding due to the lightning-distillation process.

---

# License & Credits

This model is provided under the **MIT** and **CreativeML Open RAIL++-M** licenses.  
It is intended for **ethical use**.  

Users are encouraged to share their generations and provide feedback for future fine-tuning iterations.

<p style="display:flex; gap:40px;">
  <img src="pixelora_output_1.png" alt="Model banner" width="50%">
  <img src="realvis_77eeceea.png" alt="Model banner" width="50%">
</p>



# My Awesome Model

Visualize the fantasy

## 🚀 Quickstart Usage

```python
import torch
from diffusers import StableDiffusionXLPipeline, EulerDiscreteScheduler

model_id = "Layasaran/pixelora-1.0-xhigh"

pipe = StableDiffusionXLPipeline.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
).to("cuda") 

pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")

import torch
import uuid

def generate_image(prompt: str, pipe, device="cuda:0"):
    """
    Args:
        prompt (str): The text description.
        pipe: The loaded StableDiffusionXLPipeline.
        device (str): Which T4 or any gpu.
    """

# use negative prompt if needed.
    negative_prompt = "(worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth"
    
    pipe.to(device)
    
    image = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=6,   
        guidance_scale=1.5,      
        width=1024,
        height=1024
    ).images[0]

    filename = f"pixelora.{uuid.uuid4().hex[:8]}.png"
    image.save(filename)
    
    print(f"Image saved as {filename} using {device}")
    return image

# Recommended Prompt Structure
prompt = "RAW photo, a close-up cinematic portrait of a sailor, weathered skin, salt-encrusted beard, soft morning light, 85mm lens, f/1.8, 8k uhd"

# Generate
generate_image(prompt, pipe)