FireRed flowchart best_v7 optimized assets

This repository stores the large runtime assets for deploying the flowchart erasing model. The code, service entrypoints, and experiment scripts are maintained in the Git repository.

Contents

• models/flowchart_v7_lora_best/: best_v7 LoRA adapter trained from the Lightning-initialized run.
• runtime/optimized/: recommended optimized inference dependencies and local wheels.
• service/default_config.json: default inference/service parameters.
• metadata/: validation and optimized benchmark metadata.
• eval/full_best_new_benchmark_optimized/: optimized benchmark run metadata.

Expected Base Model

The remote server is expected to already have the base model at pretrained_models/FireRed-Image-Edit-1.1.

If your server stores the base model elsewhere, set MODEL_PATH or edit service/default_config.json.

Default Inference Parameters

• Prompt: Remove all the bright green overlays
• Seed: 43
• Steps: 8
• True CFG scale: 1.0
• Guidance scale: 1.0
• Pipeline mode: optimized
• Optimized max side: 1664
• Max sequence length: 512

Notes

The full Python virtual environment is intentionally not uploaded because it is not portable across servers. Use runtime/optimized/requirements_optimized.txt and the wheels under runtime/optimized/wheels/ to rebuild the environment.