feat: Add tiled multi-diffusion upscaling support for FLUX and Z-Image models by Pfannkuchensack · Pull Request #9002 · invoke-ai/InvokeAI
…e models Extends the upscaling tab to support FLUX.1 and Z-Image models in addition to the existing SD1.5/SDXL support. Each model architecture gets its own tiled denoise backend node and frontend graph builder that handles the architecture-specific denoising pipeline (transformer vs UNet), text conditioning, VAE, ControlNet, and scheduler. Backend: - Add TiledFluxDenoiseLatents node with FLUX transformer tiling, ControlNet cropping (InstantX/XLabs), LoRA and scheduler support - Add TiledZImageDenoiseLatents node with Z-Image transformer tiling, ZImageControlNet cropping, LoRA and scheduler support Frontend: - Add buildFluxMultidiffusionUpscaleGraph and buildZImageMultidiffusionUpscaleGraph graph builders - Update useEnqueueUpscaling to dispatch correct graph per model base - Update readiness validation for FLUX/Z-Image specific requirements - Show model-appropriate scheduler (FLUX/Z-Image/SD) in upscale tab - Show Z-Image VAE/Qwen3 selectors in upscale advanced settings - Fix UpscaleWarning auto-select to also match "union" ControlNet names - Reuse existing i18n keys for validation messages