Zhare-AI

Privacy-preserving text-to-image generation in your browser with WebGPU acceleration This is a browser-optimized implementation of Stable Diffusion v1.5, specifically converted and optimized for client-side deployment using WebGPU acceleration. Developed by Zhare-AI, this model enables high-quality image generation directly in web browsers without requiring server infrastructure, ensuring complete user privacy and data sovereignty. Democratizing AI through distributed computing and privacy-preserving technology - 🌐 Fully Client-Side: Complete image generation in the browser, no data leaves your device - ⚡ WebGPU Accelerated: Hardware-accelerated inference with automatic WebAssembly fallback - 🔒 Privacy-First: All processing happens locally, protecting user prompts and generated content - 📱 Cross-Platform: Compatible with desktop and mobile browsers - 🛠️ Production-Ready: Optimized for real-world web applications | Component | Description | Approximate Size | |-----------|-------------|------------------| | Text Encoder | CLIP ViT-L/14 for text understanding | ~500MB | | UNet | Core diffusion model for image generation | ~3.4GB | | VAE Decoder | Converts latents to final images | ~160MB | | VAE Encoder | Encodes images to latent space | ~160MB | | Safety Checker | Content filtering (optional) | ~600MB | Total Model Size: ~4.8GB (without safety checker: ~4.2GB) Generation time for 512×512 images with 20 inference steps: | Hardware Category | Example Device | Typical Performance | |------------------|----------------|-------------------| | High-End Desktop | RTX 4090, RTX 4080 | 3-8 seconds | | Gaming Desktop | RTX 3080, RTX 3070 | 8-15 seconds | | Intel Arc GPUs | Arc A750, Arc A770 | 8-15 seconds | | AMD High-End | RX 7900 XT/XTX | 6-12 seconds | | Apple Silicon | M2 Max, M1 Ultra | 10-20 seconds | | Integrated GPUs | Intel Iris Xe | 25-50 seconds | | WebAssembly Fallback | CPU-only devices | 2-10 minutes | - Minimum VRAM: 4GB (recommended: 6GB+) - System RAM: 8GB minimum, 16GB recommended - Storage: 5GB free space for model files - Browser: Chrome 113+, Edge 113+ (WebGPU), or any modern browser (WebAssembly fallback) | Browser | WebGPU Support | Performance Level | Notes | |---------|---------------|------------------|-------| | Chrome 113+ | ✅ Full Support | Excellent | Primary recommendation | | Microsoft Edge 113+ | ✅ Full Support | Excellent | Primary recommendation | | Firefox 141+ | ✅ Stable Support | Very Good | Recent WebGPU implementation | | Safari 17.4+ | 🔶 Experimental | Good | Behind feature flag | | Mobile Chrome 121+ | 🔶 Limited | Fair | Android only, limited memory | All browsers support WebAssembly fallback for universal compatibility This model is based on Stable Diffusion v1.5 with the following training characteristics: - Base Dataset: LAION-5B filtered subset (~590M image-text pairs) - Training Resolution: 512×512 pixels - Architecture: Latent Diffusion Model with CLIP ViT-L/14 text encoder - Precision: Originally trained in FP32, optimized to FP16 for browser deployment - ONNX Conversion: Optimized computational graph for web inference - WebGPU Kernels: Custom compute shaders for GPU acceleration - Memory Efficiency: Attention slicing and dynamic memory management - Cross-Platform: WebAssembly fallback ensures universal browser support - Content Filter: Optional NSFW detection and filtering - Prompt Sanitization: Basic filtering of potentially harmful prompts - Local Processing: No data transmission ensures privacy protection ✅ Encouraged Uses: - Creative art and design projects - Educational demonstrations of AI capabilities - Rapid prototyping for applications - Personal creative exploration - Research and development ❌ Prohibited Uses: - Creating harmful, offensive, or illegal content - Generating misleading information or deepfakes - Violating copyright or intellectual property rights - Any use that violates the CreativeML OpenRAIL-M license terms - Zero Data Collection: All processing occurs locally in your browser - No Server Communication: Model runs entirely offline after initial download - User Control: Complete control over generated content and prompts - GDPR Compliant: No personal data processing or storage - Resolution: Optimized for 512×512 (other resolutions may reduce quality) - Batch Size: Single image generation only in browser environment - Memory Constraints: Limited by browser and device VRAM/RAM - Generation Speed: Slower than dedicated server hardware - Language Bias: Best performance with English prompts - Cultural Representation: Training data may reflect Western/English-speaking biases - Artistic Style: Tendency toward photorealistic and digital art styles - Consistency: Multiple generations from same prompt may vary significantly - WebGPU Availability: Limited to supporting browsers and devices - Memory Management: Browser security limits may affect large model loading - Performance Variance: Significant variation across different devices and browsers This model is released under the CreativeML OpenRAIL-M license, which allows for: ✅ Permitted: - Commercial and non-commercial use - Distribution and modification - Creation of derivative works - Integration into applications and services 🚫 Restrictions: - Must not be used to generate harmful content - Cannot be used for illegal activities - Must include license terms in any distribution - Derivative works must maintain the same license restrictions Full License Text: Available at CreativeML OpenRAIL-M License When using this model: 1. Include License: Provide license terms to end users 2. Respect Restrictions: Ensure use cases comply with content restrictions 3. Derivative Works: Apply same license to modified versions 4. Attribution: Credit original Stable Diffusion creators and Zhare-AI adaptation Zhare-AI is focused on democratizing AI technology by making powerful models accessible directly in web browsers. Our mission is to enable privacy-preserving AI applications that put users in control of their data and creative processes. - Website: zhare.ai - Focus: Distributed AI computing and browser-based AI applications - Philosophy: Privacy-first, user-controlled AI experiences - Vision: Making AI accessible, private, and distributed We believe AI should be: - Accessible to everyone, regardless of infrastructure - Private with complete user data control - Distributed across devices rather than centralized servers - Transparent with open-source implementations - Issues: Report technical problems via the repository issues - Discussions: Join the community discussion for tips and examples - Documentation: Comprehensive guides available in the repository We welcome contributions to improve browser compatibility, performance, and user experience: - Performance optimizations for different hardware - Browser compatibility improvements - Documentation enhancements - Example applications and tutorials 🚀 Ready to create amazing images directly in your browser? This model brings the power of Stable Diffusion to web applications while keeping your data completely private and secure. Developed with ❤️ by Zhare-AI for the open-source community 🌐 Visit Zhare.ai | 📧 Contact Us | 💬 Join Discussion