1.14.0

• added an optional vae_decode_tiled (with vae_decode_tile_size) to the Anima LLLite Tiled ControlNet Sampler for multidiffusion mode. That mode's final VAE decode is a single full-image pass (independent of rows/columns), which can spike VRAM on large images — or, on Windows, crawl by spilling into shared system memory instead of raising a clean out-of-memory error. Enabling it decodes the latent in bounded tiles. Default off, and the two fields only show when sampling_mode is multidiffusion (they have no effect in per_tile mode, where each tile is already decoded on its own).
• added a new model-agnostic MultiDiffusion Tiled Hires Fix node in the vsLinx/sampling group. It's the multidiffusion behaviour of the Anima LLLite Tiled ControlNet Sampler with the LLLite parts stripped out, so it works on any model (SD/SDXL/Flux/etc.): one sampling pass over the whole image, splitting the latent into overlapping tiles each denoising step and averaging the overlaps in latent space — no tile seams, and per-step VRAM stays tile-sized. Upscale the image first, then refine it here with a low denoise. Includes the optional vae_decode_tiled/vae_decode_tile_size and search aliases like hires fix / tiled diffusion / tiled upscale so it's findable without knowing the term "multidiffusion".