Image Quality and Inference Optimization
I've been testing SD3 Medium for production image generation and wanted to share some observations and questions:
Image Quality vs Speed Trade-offs: What inference steps count provides the best balance? I've found 28 steps works well, but curious about community experiences.
Memory Requirements: The 2B parameter MMDiT architecture - what's the minimum VRAM needed for batch processing? Any optimization techniques beyond standard quantization?
Typography Performance: One of SD3's key improvements is text rendering. Has anyone fine-tuned this for specific use cases (logos, technical diagrams)?
Commercial Licensing: For those using this in production, how are you handling the Stability AI Community License requirements?
Would love to hear real-world deployment experiences from the community.