Add Bria 3.2 text-to-image generation node

[georgi]

Summary

This PR adds support for Bria 3.2, a commercial-ready text-to-image generation model, as a new HuggingFace pipeline node. Bria 3.2 is trained entirely on licensed data and offers high-quality image generation with accurate text rendering capabilities.

Key Changes

• New Bria node class (src/nodetool/nodes/huggingface/text_to_image.py):

  • Implements Bria class extending HuggingFacePipelineNode
  • Configurable parameters: prompt, negative prompt, guidance scale, inference steps, dimensions, seed, sequence length, and memory optimization flags
  • Sensible defaults tuned for Bria 3.2 (guidance_scale=5.0, steps=30, max_sequence_length=128)

• Precision handling (_apply_precision_fixes method):

  • Casts T5 text encoder to bfloat16 for efficiency while keeping the final dense layer in float32 (Bria-specific requirement)
  • Keeps VAE in float32 when shift_factor is zero (VAE doesn't support mixed precision)
  • Includes error handling with informative logging

• Memory optimization (move_to_device method):

  • Supports CPU offloading to reduce VRAM usage
  • Handles device movement with proper error messages for out-of-memory scenarios
  • Enables attention slicing on GPU to further reduce memory footprint
  • Disables CPU offload when moving to CPU for better performance

• Comprehensive test suite (tests/test_text_to_image_bria.py):

  • Tests node metadata and configuration
  • Validates precision fixes for both zero and non-zero VAE shift factors
  • Tests device movement behavior with and without CPU offload
  • Includes mock pipeline classes for isolated testing

• Metadata registration (src/nodetool/package_metadata/nodetool-huggingface.json):

  • Registers Bria node with UI metadata
  • Specifies basic fields (prompt, height, width, seed) for simplified UI
  • Includes recommended model configuration pointing to briaai/BRIA-3.2

Notable Implementation Details

• Uses async/await pattern with asyncio.to_thread() for non-blocking pipeline execution
• Implements reproducible generation via optional seed parameter (-1 for random)
• Includes progress callbacks during inference for UI feedback
• Properly handles generator creation on CPU for reproducibility
• Follows existing HuggingFace node patterns for consistency with other text-to-image models

https://claude.ai/code/session_017UVzej67SDCFVHVfe6VQ2n