AetherScale v0.9.2 — Native DLSS 5, Frame Generation, Low-RAM Video Pipeline & Production Video Tools

AetherScale v0.9.2 is a major update over v0.5.5, expanding the project from an NVIDIA post-processing experiment into a broader GPU-native video enhancement and finishing pipeline for ComfyUI.

Major changes since v0.5.5

• Added experimental native DLSS Frame Generation support through the new MFG pipeline.
• Added AetherScale • MFG and direct-stream MFG Video workflows for frame interpolation and high-FPS output.
• Added a new native in-process DLSS 5 Neural Rendering backend with CUDA/D3D12 interop, while retaining carrier and legacy paths as fallbacks.
• Added chunk-aware Neural Rendering designed to reduce peak VRAM and RAM usage on long sequences.
• Added AetherScale • Video Combine as a high-performance alternative to traditional ComfyUI video saver workflows.
• Added NVENC encoding, CPU codecs, MOV output and ProRes profiles.
• Added audio muxing, selectable audio bitrate, silent-copy control and embedded/sidecar metadata.
• Added generation metadata support for model, seed, sampler, scheduler, steps, CFG and denoise.
• Added an integrated video preview directly inside the Video Combine node.
• Fixed the Video Combine preview layout so the node no longer grows every time the UI refreshes or the workflow is edited.
• Added automatic NVENC preflight, GPU routing diagnostics and codec fallback for resolutions unsupported by the selected H.264 NVENC path.
• Added AetherScale • Video Loader, a low-RAM FFmpeg-based loader intended for long or high-resolution video workflows.
• Added bounded mmap / disk-spill storage for outputs that cannot safely fit in RAM or Windows commit space.
• Reworked temporary cache handling and crash recovery to avoid persistent multi-gigabyte cache files.
• Added automatic stale-cache cleanup and configurable cache location through AETHERSCALE_CACHE_DIR.
• Added progress reporting across long-running AetherScale nodes.
• Added proper ComfyUI interruption support throughout long frame-processing, encoding and streaming operations.
• Reworked Super Resolution staging to reduce unnecessary CPU / disk serialization and improve GPU utilization.
• Added HDR-style enhancement fallback for systems where the NVIDIA VFX runtime exposes no compatible HDR effect.
• Improved Motion Analysis and MFG motion handling for emissive details, flicker, thin structures and unstable optical flow.
• Added safer MFG output-size checks to prevent accidental 80–160+ GiB IMAGE materialization.
• Added direct streaming paths for workflows where full intermediate frame batches would be impractical.
• Expanded diagnostics, runtime detection, GPU selection and Windows resource reporting.
• Updated README, changelog, Registry metadata and third-party notices across the project.

Recommended pipeline

For large video jobs, the preferred workflow is now:

Video Loader → Restoration → Neural Rendering → Super Resolution → HDR → MFG Video

This avoids materializing extremely large interpolated IMAGE batches and keeps frame generation close to the final encoder.

Notes

DLSS 5 Neural Rendering and Frame Generation integrations are experimental and rely on NVIDIA runtime behavior that is not part of a stable public ComfyUI API. Runtime compatibility may vary by GPU generation, NVIDIA driver and runtime build.

AetherScale remains Windows / NVIDIA focused and is currently developed primarily around RTX 40 / RTX 50 class workflows.

Fix clean_cache to eliminate persistent mmap files

Fixes clean_cache=true still leaving large .mmap files while ComfyUI retains completed workflow outputs. Clean-cache outputs now use anonymous/pagefile-backed mappings, so no cache file is created on disk. clean_cache=false remains available as the explicit persistent disk-backed mmap mode.

Add automatic mmap cache cleanup

Adds a clean_cache switch to mmap-backed AetherScale nodes and fixes multi-gigabyte cache files accumulating between runs. Cache mappings are now removed automatically when ComfyUI releases their tensor storage, with PID-aware cleanup for orphaned files from terminated processes.

Fix AetherScale HDR on current NVIDIA VFX runtimes

AetherScaleHDR no longer fails when the installed NVIDIA VFX runtime exposes no HDR effect. The node now uses AetherScale’s CUDA HDR-style enhancement path as the working default, while preserving the same strength, saturation, contrast, highlight preservation, and profile controls. Native NVIDIA VFX HDR probing remains available for future SDK support.

GPU-native NVIDIA video enhancement, restoration, temporal analysis, and experimental DLSS 5 Neural Rendering nodes for ComfyUI.

v0.5.2

Create publish_action.yml