Claude
61a7a3bc8c
Improve canvas responsiveness and document text encoder architectures
Canvas Improvements:
- Added forced redraw when width/height inputs change
- Canvas now updates immediately (setDirtyCanvas) instead of waiting for next render cycle
- Better visual feedback for users adjusting dimensions
Documentation (CONTEXT.md):
- Comprehensive text encoder architecture table for all major models
- SD 1.5/2.x: 1 encoder
- SDXL: 2 encoders (CLIP-L + OpenCLIP-G)
- Flux: 2 encoders (T5-XXL + CLIP-L) - ONLY model using encode_from_tokens_scheduled
- Chroma: 1 encoder (T5-XXL only) - Flux-based architecture but single encoder!
- SD3/SD3.5: 3 encoders (CLIP-L + CLIP-G + T5-XXL)
- HiDream: 4 encoders in ComfyUI
- WAN 2.1/2.2: 1 encoder (UMT5-XXL)
- Qwen-Image: 1 encoder (architecture unknown)
Key Insight: Don't assume encoder count from model architecture
- Chroma is Flux-based but uses only 1 encoder (not 2 like Flux)
- Current detection strategy (hasattr 'encode_from_tokens_scheduled') correctly identifies Flux
- All other models use standard encode_from_tokens (ComfyUI handles multi-encoder internally)
2025-11-22 02:48:55 +00:00
..
2025-11-22 01:10:55 +00:00
2025-11-22 01:10:55 +00:00
2025-11-22 01:10:55 +00:00
2025-11-22 02:48:55 +00:00
2025-11-22 01:10:55 +00:00