⚡ DreamShaper Interactive Workbench

A real-time generative prototyping tool exploring the trade-offs between inference latency and high-fidelity image synthesis.

(Click to try the live demo running on CPU)

---

💡 Project Overview

Built to mirror the commercial workflows of tools like Adobe Firefly, this project demonstrates how Latent Consistency Models (LCM) can be leveraged to create sub-second, interactive creative tools.

By distilling the standard Stable Diffusion inference process from 50 steps down to just 6, this workbench achieves a 10x reduction in latency (1.2s on T4 GPU) without sacrificing the high-frequency texture details required for professional design assets.

Key Features

• ⚡ Sub-Second Inference: Optimized pipeline achieving ~1.2s latency on consumer GPUs using LCM-LoRA.
• 🎨 Multi-Style Comparison Engine: Parallelized batch generation that allows users to instantly visualize a single concept across 5 distinct aesthetic styles (Cinematic, 3D Render, Anime, etc.) for rapid creative iteration.
• 🛠️ Full-Stack Interactive UI: A responsive Gradio frontend featuring real-time performance metrics, reproducible seed controls, and a "Sketch-to-Image" canvas powered by ControlNet.

---

📸 Performance & Visuals

1. The "Design Workbench" (Style Comparison)

Demonstrating the ability to rapidly explore aesthetic variations of a single concept; critical for speeding up the ideation phase in professional workflows.

Prompt: "Magical floating island in a light blue sky..." | Time: 7.6s (Total for 5 images)

2. High-Fidelity Texture Generation

Proving that speed does not compromise quality. The model successfully renders complex textures (frosting, light reflection) even at low step counts.

Prompt: "A cupcake with lemon frosting" | Time: 1.28s | Steps: 6

---

🛠️ Tech Stack & Engineering Decisions

• Core Frameworks: Python, PyTorch, Hugging Face Diffusers
• Models: * Checkpoint: Lykon/dreamshaper-8 (Selected for superior photorealism over base SD 1.5)
  • Speed Optimization: latent-consistency/lcm-lora-sdv1-5 (Enables 4-8 step inference)
  • Control: lllyasviel/control_v11p_sd15_scribble (Ensures strict geometric adherence to user input)
• Interface: Gradio (Custom CSS for responsive grid layouts)
• Deployment: Hugging Face Spaces (CPU/GPU agnostic architecture)

🚀 How to Run Locally

1. Clone the repository:

   git clone [https://github.com/rukmininazre/dreamshaper-workbench.git](https://github.com/rukmininazre/dreamshaper-workbench.git)
   cd dreamshaper-workbench

2. Install dependencies:

   pip install -r requirements.txt

3. Run the application:

   python app.py

   Note: The application defaults to CUDA if available. To run on CPU, ensuring torch_dtype=torch.float32 is handled automatically.