FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space
Abstract
- Sequence concatenation approach
- KontextBench, comprehensive benchmark with 1026 image-prompt pairs → validate effectiveness
Introduction
Local editing
- LaMa, Latent diffusion inpainting
- Repaint
- Stable diffusion inpatinting
Generative editing
- Extraction of a visual concept
- IP-adapter
Adversarial diffusion distillation
Evaluations & applications
Kontextbench, novel benchmark featuring real-world image editing challenges
Existing benchmarks
- GEdit-bench
- IntelligentBench
- DrawBench
Crowd sourced real-world use cases
1026 image-prompt pairs derived from base images
Contribution
- Character consistency
- Interactive speed
- Iterative application
Experiments
FLIX.1 is built from a mix of double stream and single stream blocks
T2I, I2I latency is surprisingly improved.
Discussion
Flow matching model which combines in-context image generation and editing in a single model.
