MASSIVE shoutout to the ULTRA powerful @RalFinger - I generated synthetic data using his acidzlime LoRA , and I have plans to repeat this process with other styles. His LoRAs are the best LoRAs, hands down.
These LoRAs are trained for Wan2.1, and will work there, but my primary use case is with autoregressive distillations of Wan2.1 such as Longlive, StreamDiffusionV2, and Krea Real Time.
If you want to try this LoRA in real time, you can do so with the open source tool Daydream Scope. I am one of the developers on the team. All of the demos here were created with Scope - 4090 or 5090 required, though Runpod templates are available, and remote inference under construction. Real-time VACE writeup Scope/Real-Time VACE tutorial:
The datasets for these were chosen with the causal nature of autoregressive models in mind, and that has been where my testing has taken place - you can consider these experimental of course.
