Fast-BAGEL (LGCC): Fast and Detail-Enhanced Image Editing

1. Slow Inference Speed: Existing models require many steps to produce high-quality results.
2. Detail Degradation: Fine details are often lost or blurred during the editing process.
3. Content Inconsistency: Edits can unintentionally alter parts of the image that should remain unchanged.

Model Details

• Base Model: Finetuned from ByteDance-Seed/BAGEL-7B-MoT.
• Architecture: Flow matching-based model for text-guided image editing.
• Key Innovations:
  • Local Gaussian Noise Coupling (LGNC): A novel noise scheduling technique that preserves fine-grained spatial details by coupling noise to the local characteristics of the latent space. This is the primary driver of improved detail and faster convergence.
  • Content Consistency Loss (CCL): A new loss function applied during training that ensures semantic alignment between the text prompt and the image, preventing unintended alterations to background content.
• Training: The model was finetuned on a private dataset using a curriculum learning strategy to efficiently integrate our improvements into the pre-trained ByteDance BAGEL checkpoint.

How to Get Started

1. Clone the Repository

2. Download Model Files

3. Run Inference

• --model_dir: The Hugging Face repository ID (zanchat-ai/fast-bagel).
• --num_timesteps: Number of inference steps. 50 steps are recommended for high-quality results (vs. 100 for the original BAGEL). Use fewer steps (e.g., 20-30) for faster previews.
• --time_shift: A parameter for the solver schedule. 4 is a good default.

Performance and Benchmarks

Inference Speed-Up

Quality Improvement (on I2EBench)

Fig. 1: LGCC (Fast-BAGEL) achieves comparable or superior results in just 20-30 steps, while the original BAGEL requires 50 steps.

Limitations and Bias

• Inherited Limitations: As a finetuned version of ByteDance's BAGEL, this model may inherit any limitations or biases present in the original model and its training data.
• Content Domain: The model's performance is strongest on general, real-world images. It may struggle with highly abstract, technical, or niche visual concepts.
• Bias in Training Data: The model was trained on large-scale datasets which may contain societal, cultural, and gender biases. Generated or edited images may reflect these biases. We recommend users be mindful of this when using the model for downstream applications.
• Complex Edits: While significantly improved, the model may still face challenges with edits that require major topological changes, complex logical reasoning, or understanding of nuanced, abstract concepts.

Citation

Contact