Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, Applications and Opportunities
Paper
•
2408.07666
•
Published
•
1
Reading/Expts
A collection of papers that I'm excited to experiment on. Once the expt is done, I'll probably move them to other collections.
Note Looking to apply it to DiT's. Sayak already did a simple weight merge for Flux dev and schnell. Can we extend and see what results can be derived from Flux ? For Auraflow, v0.2 has better prompt adherence than v0.3 but lower aesthetic quality. Can we extend that by figuring out a way to retain prompt adherence while also improving quality ? > Doesn't work, v0.2 and v0.3 have different pos embed dims.
CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models
Paper • 2407.15886 • Published • 1Note The proposed unet architecture is pretty small and lightweight for a condition diffusion model. Can we extend this to use for other models like Canny conditioning ? Will have to sacrifice text condition or can we keep it ?
RB-Modulation: Training-Free Personalization of Diffusion Models using Stochastic Optimal Control
Paper • 2405.17401 • Published • 5Note Based on two ideas: 1. SOC: Stochastic Optimal Control which is very similar to moving the point on the manifold closer as in inverse problems. But uses a style descriptor. 2. AFA: Averaging Forward Attention (iirc?)