This content originally appeared on DEV Community and was authored by Takara Taniguchi
Zhenxiong Tan, NUS のXinchao wangのグループ
Omnicontrol, novel approach addressing how image conditions are integrated into DiT
Subjects200k, a large-scale dataset of image pairs.
Effective image control
Contribution
- They propose omnicontrol, cotrol framework for DiT models
- Two key technical innovations
- unified sequence processing
- adaptive position encoding
- flexible attention bias mechanism →conditioning
- Subjects200k, large scale dataset
Related works
Diffusion models
- DiT
- FLUX
Controllable generation
- Controlnet
- T2I adapter: lightweight adapter
- Unicontrol
- mixture of experts
Methods
Preliminary
Diffusion transformer model is used for FLUX.1, SD3, PixArt.
Minimal architecture
Unified sequence processing
Experiment
Outperformed controlnet
Conclusion
Omniccontrol offers parameter-efficient image conditional control for DiT.
Required parameters are only 0.1 % compared to the traditional one.
感想
ControlnetでやってることをDiTでやってる感じですかね
This content originally appeared on DEV Community and was authored by Takara Taniguchi