ANIMA_BOOSTER v1.2.0 — Stability & Performance Update

This major update focuses on robust stability, codebase refactoring, and fixing scaling bugs for stochastic/SDE samplers. We have eliminated unstable components and made the suite bulletproof for everyday generation.

🆕 What's New in v1.2.0:

1. 🐛 TeaCache Fixed for SDE/Stochastic Samplers (e.g., er_sde, sde gpu)

• The Issue: Stochastic samplers working on a sigma scale previously confused TeaCache's fixed threshold. This triggered aggressive caching on the very first step, resulting in fast generations but heavily distorted images covered in artifacts.
• The Solution: Implemented dynamic timestep scale auto-detection (st.max_t). TeaCache now mathematically adapts to any sampler and scheduler (sigmas, 1000..0, or 1..0). Early structural steps are fully protected, while late-stage detailing is safely cached. Enjoy perfect image quality with SDE samplers!

2. 💎 Safe One-Click JIT Compilation (torch.compile)

• Unstable AnimaTorchCompile node removed: The complex external compilation node was prone to PyTorch crashes (CUDA Graphs tensor overwrite errors).
• Integrated JIT Toggle: We integrated a safe, one-click torch_compile toggle directly into Anima Booster Loader and Checkpoint Loader. It runs on the stable inductor backend (default mode) without CUDA Graphs. Enjoy the same +20% to +40% speed boost with 100% stability!

3. 🗑️ Codebase Cleanup & Optimization

• Removed AnimaSparseAttention: Local sparse attention on blocks trained on Full Attention destroyed global image geometry and caused structural artifacts.
• Removed AnimaTorchCompile: Replaced by the native, JIT toggle in the model loaders.
• The package is now cleaner, lighter, and completely safe.

4. 📦 Graceful Degradation & Portable Windows Support

• All high-performance modules (like SageAttention) are now fully optional. If not installed, the loader will seamlessly fall back to PyTorch's native SDPA without throwing import errors.
• Windows/Portable Tip: Refer to the installation instructions in the README to download and install precompiled Triton and SageAttention binary wheels for Windows portable environments.

---

🎛️ Recommended Settings for Maximum Speed & Quality:

• Anima Booster Loader: Set sage_attention to auto and enable torch_compile. (Note: The first 2-3 generations will have a warm-up phase while PyTorch compiles the blocks).
• Anima TeaCache: Set threshold to 0.15 and keep adaptive ON.
• For SDE Samplers (like er_sde): Now fully compatible and artifact-free! If you want to push the speed further while maintaining great quality, try raising the TeaCache threshold to 0.22 - 0.25.