Skip to content

ArtTic-LAB v1.5.0: Next-Gen Models, Seamless Performance

Choose a tag to compare

View all tags
@rootLocalGhost rootLocalGhost released this 30 Jul 08:46
· 237 commits to main since this release
1.5.0
f7178e1

We are thrilled to announce the release of ArtTic-LAB v1.5.0: The Power & Efficiency Update!

This version represents a major leap forward, transforming ArtTic-LAB from a powerful tool into a truly professional-grade creative suite. We've gone under the hood to completely re-engineer the pipeline system, enabling support for the latest next-generation models while simultaneously introducing critical features for memory stability and workflow efficiency. This update is all about pushing the boundaries of what's possible on Intel ARC hardware.

🎉 Highlights of v1.5.0

  • 🚀 Next-Gen SD3 Support: Be among the first to experiment with Stable Diffusion 3! ArtTic-LAB now intelligently handles single-file SD3 models by automatically fetching the required base components, giving you cutting-edge power with zero hassle.
  • ✨ Full SD2.x & SDXL Compatibility: The model detection engine is now a master of its craft, flawlessly identifying and configuring pipelines for SD1.5, SD2.0, SD2.1, SDXL, and SD3 models.
  • 🧠 Memory-Safe VAE Tiling: Say goodbye to "Out of Memory" errors on high-resolution generations! VAE Tiling is now integrated and enabled by default, ensuring smooth and stable performance even when pushing the limits.
  • ⚡ Instant Model Unloading & Hot-Swapping: A new "Unload Model" button gives you direct control over your VRAM. Plus, the app is now smart enough to automatically unload the old model before loading a new one, making switching between models seamless and crash-free.
  • ⚙️ Smarter, Faster Pipeline System: A complete architectural overhaul of the backend makes the entire application more robust, faster to load, and easier to extend for the future.

✅ What's New in v1.5.0: The Full Changelog

✨ Features

  • Stable Diffusion 3 Support: The application can now load single-file SD3 models. It automatically downloads the official base components (like the text encoders and VAE) from Hugging Face on the first run and injects the local model's weights, giving you a seamless experience.
  • Stable Diffusion 2.0/2.1 Support: Full support has been added for SD2.x models. The backend automatically detects them and uses the appropriate pipeline and 768x768 native resolution.
  • Dynamic Aspect Ratio Presets: The one-click aspect ratio buttons are now even smarter! They check the loaded model type (SD1.5, SD2, SDXL, or SD3) and apply the correct resolution presets, removing all guesswork.

🚀 Performance & Stability

  • VAE Slicing & Tiling: This critical feature is now integrated to dramatically reduce peak VRAM usage during the final image decoding stage. It is enabled by default to prevent out-of-memory errors when generating large images.
  • Explicit Model Unloading: A new "Unload Model" button has been added to the UI, allowing you to instantly clear the active model from your GPU's VRAM.
  • Automatic Model Hot-Swapping: The model loading process has been fortified. When you load a new model, ArtTic-LAB now automatically unloads the previously loaded model first, ensuring a stable transition without memory conflicts.

⚙️ Backend & Architecture

  • Unified Pipeline Architecture: The pipeline system has been completely refactored around a new base_pipeline.py. This centralizes all core logic like IPEX optimization and device placement, making the code cleaner, more efficient, and follows the DRY (Don't Repeat Yourself) principle.
  • Advanced Model "Sniffing": The model detection logic in pipelines/__init__.py has been significantly upgraded. It can now reliably differentiate between all major Stable Diffusion versions by inspecting their unique tensor key structures.
  • Expanded IPEX Optimization: The IPEX optimization step is now more comprehensive. In addition to optimizing the main U-Net or Transformer, it now also optimizes the VAE, providing a broader performance uplift across the entire generation process.

A Note on Our Journey

The evolution to v1.5.0 was driven by a single question: "How can we make this not just work, but work flawlessly?" This update is a testament to the rigorous testing and forward-thinking vision of our lead tester, who pushed for the stability and next-gen features that define this release. We've moved from building a functional car to engineering a high-performance vehicle.

What's Next? 🔮

With a rock-solid and future-proof foundation, we're positioned to bring even more incredible features to ArtTic-LAB. Here's a look at what's on the horizon:

  • Low VRAM Mode: Implementing "CPU Offloading" to allow users with less VRAM to run massive models like SDXL and SD3.
  • LoRA Support: A full-featured system for loading and managing LoRA files to customize your creations with incredible flexibility.
  • Image-to-Image (Img2Img): The highly requested feature to guide generations using an input image.

Thank you for your continued support. The creative ceiling has been raised.

Now, go make something amazing. ✨

Full Changelog: 1.0.0...1.5.0

Assets 2
Loading