v0.3.1 - Classifier-Free Guidance (CFG) Support

🔄 Version Note

This is v0.3.1 - a coordinated release across the FluxFlow ecosystem.
Note: v0.3.0 does not exist on PyPI due to release coordination. This release contains all features originally planned for v0.3.0.

---

🚀 Major Features

Classifier-Free Guidance (CFG) Support

Enhanced inference control for better prompt following and quality:

New Parameters:

• use_cfg - Toggle CFG on/off (boolean)
• guidance_scale - Control conditioning strength (1.0-15.0)
  • 1.0 = no guidance (standard sampling)
  • 7.0-10.0 = recommended for most cases
  • 15.0+ = very strong conditioning
• negative_prompt - Specify unwanted features (string)

Implementation:

• Dual-pass sampling for CFG inference
• Efficient batching for conditional and unconditional passes
• Compatible with models trained with cfg_dropout_prob > 0

Infrastructure:

• CFG-aware model loading and validation
• Enhanced configuration validation for CFG parameters
• Backward compatible - works with non-CFG models

---

🎨 Bézier Activations (v0.1.1+)

TrainableBezier for per-channel learnable transformations:

• Optimized implementation with torch.addcmul (1.41× faster)
• Used in VAE latent bottleneck (mu/logvar) and RGB output
• Total 1,036 learnable parameters: 1,024 (latent) + 12 (RGB)

Performance:

• Benchmark: 90sec/step → 8sec/step after SlidingBezier removal
• Fused multiply-add operations for efficiency
• Cached intermediate values (t², t³, t_inv², t_inv³)

---

✨ VAE Architecture Improvements

Enhanced Attention:

• Increased attention layers from 2 to 4
• Better global context modeling
• Richer latent representations

Improved RGB Output:

• Wider channels: 128→96→48→3
• GroupNorm + SiLU activation
• No squashing for full color range
• Fixes color tinting issues

Input Validation:

• Channel dimension validation in FluxCompressor.forward()
• Early detection of shape mismatches

---

🔧 Technical Improvements

SPADE Context Handling (v0.2.1):

• Use None instead of torch.zeros_like() when disabled
• More efficient - avoids unnecessary tensor allocation
• Semantically correct - explicitly means "no context"

Removed SlidingBezier:

• 25× slower than SiLU
• 5× memory overhead from unfold()
• Performance: 90sec/step → 8sec/step

---

📦 Installation

pip install fluxflow==0.3.1

---

🔗 Links

• PyPI: https://pypi.org/project/fluxflow/0.3.1/
• Documentation: https://github.com/danny-mio/fluxflow-core/blob/v0.3.1/CHANGELOG.md
• Training Tools: https://github.com/danny-mio/fluxflow-training/releases/tag/v0.3.1

---

📋 What's Changed

Full Changelog: v0.2.1...v0.3.1

---

🧪 CI Status

• ✅ All tests passing (Python 3.10, 3.11, 3.12)
• ✅ Code quality: flake8 clean, black formatted
• ✅ Type checking: mypy clean