Closed i-amgeek closed 1 year ago
Hi i-amgeek,
That's an excellent question. The primary distinction lies in the model's size and configuration. Our UniControl operates as a single, comprehensive model (1.4B parameters, 5.78GB checkpoint, https://console.cloud.google.com/storage/browser/_details/sfr-unicontrol-data-research/unicontrol.ckpt) that can handle all tasks concurrently. In contrast, the Multi-ControlNet requires the coordination of multiple task-specific ControlNets (SD+N ControlNets, https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main) to function properly, each taking up 361M parameters and 1.45GB storage, which results in a heavier, more challenging model to deploy (SD also takes 1.1B parameters).
We believe that a compact, unified model such as UniControl represents a promising direction for the field. Moreover, we anticipate that our unified approach will provide additional benefits, exceeding the performance of a single-task ControlNet. Our user studies substantiate these expectations with positive results. We eagerly await further user feedback to continue enhancing our system.
Thank you!
Can You guys extract the model so it can be used with our own custom models ? i guess asking for extension for auto11 webui is too much , maybe there will be someone who can adapt it
Yes. Thank you for your suggestion. We will extract model to let developers to easily deploy UniControl on their own dataset. I will likely update it next week.
Great news
Architecture might be different but I am having trouble understanding if this is different from multi-controlnet for application purpose? Does it enable any new functionality or does it improve resultant image quality as compared to multi-controlnet?