Update CLAUDE.md with remote text encoder documentation

- Document HF official remote text encoder endpoint
- Explain why local text encoders can't be used (storage limit)
- Update MAX_IMAGE_SIZE to 2048

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

CLAUDE.md

@@ -33,7 +33,7 @@ The app runs on Hugging Face Spaces with ZeroGPU infrastructure. Requires `HF_TO
 
 ### Key Pipeline Details
 
-1. **Text Encoding**: Offloaded to remote Gradio client (`multimodalart/mistral-text-encoder`) - runs on CPU, network-bound
+1. **Text Encoding**: Uses Hugging Face's official remote text encoder service at `https://remote-text-encoder-flux-2.huggingface.co/predict`. This is required because loading local text encoders (T5-XXL ~10GB + CLIP ~1.5GB) exceeds ZeroGPU's storage limit. Requires `HF_TOKEN` for authentication.
 2. **Prompt Upsampling**: Uses ERNIE-4.5-VL via Hugging Face Inference API - two modes:
    - Text-only: Enhances prompts with visual details
    - Image+text: Converts editing requests into concise instructions
@@ -41,6 +41,6 @@ The app runs on Hugging Face Spaces with ZeroGPU infrastructure. Requires `HF_TO
 
 ### Configuration Constants
 
-- `MAX_IMAGE_SIZE`: 1024
+- `MAX_IMAGE_SIZE`: 2048
 - `dtype`: torch.bfloat16
-- Dimensions auto-adjust to uploaded image aspect ratio (multiples of 8, min 256, max 1024)
+- Default aspect ratio: 1:1 (2048x2048)