Flex.2-preview

Hi everyone! Sharing Flex.2-preview from Ostris, a new open-source (Apache 2.0) text-to-image diffusion model that really leans into its name – the focus is all on flexibility and integrating multiple capabilities into one model. Instead of needing separate models or complex workflows for common tasks, Flex.2 (an 8B parameter model) builds these directly into the base: 🎛️ Universal Control: A single control input mechanism trained to understand guidance from pose maps, line art, or depth maps. 🎨 Built-in Inpainting: Handle image edits and object replacement directly. 🔧 Fine-tunable: Designed for LoRA fine-tuning, with tools like AI-Toolkit already supporting training with control/inpainting inputs. Flex.1 LoRAs often work too. Keep in mind this is an early access preview. Ostris is actively developing it and notes known areas for improvement (like anatomy and text rendering). It's great seeing this kind of powerful tool being developed openly with community feedback and support!