Q: Your thoughts on SD3 Medium?
A: Yeah, it's decent, nothing really special. Uses 3 freaking Text Encoder with T5XXL fp16 takes up 9.5GB of storage, Clip G takes up 1.3GB, and Clip L takes up 240MB, excluding the actual model, which takes up 4.2GB. Totalling around 15.5GB.
Q: Are you interested in developing a SD3 model?
A: I'm interested, but not anytime soon. Currently no training tools are really available for SD3. And even my GPU (RTX 2060 12GB VRAM) struggle to generate an image. 1024x1024 took 13.5 minutes to generate. Really bad for model development.
Q: What will happen with your SDXL-based model?
A: For now, nothing. You will receive usual update of my SDXL-based models. If SDXL getting irrelevant, and I still do not have the resources to do SD3 training, probably will stop my activity completely.
Q: What are your hopes on developing a SD3-based model?
A: God, I wish someone buys me a 4090. or even a 50xx series when it comes out.