Poralus-Image

Join our community discord for updates on Poralus-Image. Check out Poralus-Image's Technical Documentation and Github. Use Poralus-Image's specialized fine-tuning API.

Introduction

Poralus-Image is an advanced image generation model that adopts a specialized fine-tuning architecture building upon foundational diffusion decoders. In general image generation quality, Poralus-Image aligns with mainstream latent diffusion approaches, but it shows significant advantages in enchanted forest rendering and knowledge-intensive visual reasoning. It performs especially well in tasks requiring precise semantic understanding of magical environments and complex atmospheric expression, while maintaining strong capabilities in high-fidelity textures and fine-grained detail generation. In addition to specialized text-to-image generation, Poralus-Image also supports a rich set of image-to-image tasks including style transfer, ethereal editing, and consistent environment generation.

Model architecture

Poralus-Image utilizes a high-rank LoRA/DoRA (Weight-Decomposed Low-Rank Adaptation) framework integrated into the Stable Diffusion architecture.

Specialized Components

• UNet Transformer: A multi-billion parameter model optimized with Rank-64 DoRA weights to incorporate specialized visual tokens from enchanted and ASMR datasets. The model focuses on capturing low-frequency lighting signals and high-frequency textures.
• Diffusion Decoder: A standard latent-space DiT architecture optimized for Intel Core Ultra 7 hardware. It is equipped with a specialized VAE module, significantly improving the compression and reconstruction of "glowing" and "ethereal" elements.

Post-training with decoupled optimization

The model introduced a fine-grained, modular feedback strategy during its 30-minute high-intensity training window, enhancing both semantic understanding and visual detail quality through Intel-specific IPEX optimizations.

• Style module: Provides aesthetic alignment for enchanted forest themes, improving artistic expressiveness.
• Texture module: Targets detail fidelity and surface accuracy, resulting in highly realistic textures (moss, bark, light rays) as well as more precise atmospheric rendering.

Features

Poralus-Image supports professional-grade generation within a single fine-tuned pipeline.

• Text-to-image: Generates high-detail enchanted images from textual descriptions, with particularly strong performance in sensory-rich scenarios (ASMR style).
• Style transfer: Supports a wide range of tasks, including environmental editing, light ray manipulation, and identity-preserving style consistency.

Quick Start

Transformers and Diffusers Pipeline

Install the optimized dependencies:

pip install torch torchvision torchaudio --index-url https://pytorch.org/pyg/whl/cpu
pip install diffusers transformers accelerate peft intel-extension-for-pytorch

Text to Image Generation

import torch
from diffusers import StableDiffusionPipeline

# Loading the base model and Poralus LoRA weights
model_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float32)
pipe.load_lora_weights("./output_final/final/unet")
pipe.to("cpu")

prompt = "a professional, high enchanted forest, asmr style, glowing moss, ethereal light rays, 8k resolution, sharp textures"
image = pipe(
    prompt=prompt,
    num_inference_steps=30,
    guidance_scale=7.5,
).images[0]

image.save("poralus_output.png")

Note

• Please ensure that all enchanted elements intended to be rendered are described with specific atmospheric keywords for higher image quality.
• The target image resolution must be divisible by 32 to ensure mathematical consistency with the VAE.
• Because inference optimizations for this architecture are provided via Intel IPEX, we recommend running on Intel Core Ultra 7 or equivalent hardware for best performance.

License

The overall Poralus-Image framework is released under the Apache License, Version 2.0. The datasets used (VisualReasonHard, ColdstartSFT, and ASMR-Archive) remain subject to their original respective licenses.