Added tiled VAE support to FaceDetailer and tiled DiT support to SeedVR2 for lower-VRAM usage

Developer marres has submitted two significant pull requests (PRs) targeting the ComfyUI ecosystem, specifically designed to tackle the platform's most notorious VRAM bottlenecks. The first PR (#1195) for the Impact Pack adds proper tiled VAE (Variational Autoencoder) encode and decode support to the FaceDetailer node, complete with explicit tile controls. The second PR (#551) for a SeedVR2 fork introduces optional tiled DiT (Diffusion Transformer) support for the diffusion upscaling phase, using overlapping latent-space tiles with blending. This dual approach directly attacks two of the worst points for memory spikes in complex workflows.

The technical impact is substantial for professionals running heavy AI image generation pipelines. By removing these VRAM spikes, users can now execute high-resolution, multi-model, and multi-stage workflows—such as chaining SDXL, FLUX.2, and SeedVR2 stages—without the previous necessity of inserting 'Clean VRAM Used' nodes between every expensive model boundary. This prevents the constant unloading and reloading of models, keeping them resident in VRAM for significantly faster processing. The PRs also include crucial hardening: FaceDetailer gets adaptive tiling and post_detail_shrink features, while SeedVR2 receives shared cache lifecycle management and failure-path fixes that resolve sporadic freezes, notably on Windows Subsystem for Linux (WSL).

This advancement democratizes access to more sophisticated AI art creation. It lowers the hardware barrier, allowing users with mid-range GPUs to experiment with workflows that were previously reserved for systems with excessive VRAM. The optimization represents a meaningful step towards more efficient and stable large-scale generative AI operations within the popular ComfyUI framework.