I fine-tuned a set of weights for a recovered NVIDIA DLSS neural renderer that's been ported to Mac. The upstream project is MLX-DLSS by WaveCut. The stock weights do a decent job on sharp stills but fall apart on soft AI video, phone footage, or upscaled frames. This version makes it useful for that too.

It's a detailer, not a super resolution miracle. It synthesizes fine texture. Maybe 30% of what it adds lines up with real detail. The rest is plausible filler. On shots that are already sharp it won't do much. It really only earns its keep on soft sources or anything that's been through a 2x upscale.

Using it

Grab the "Neural Re-Detailer - Image Upscale" or "Neural Re-Detailer - Video Upscale" nodes from the ComfyUI-MLX-DLSS pack. Example workflows are in there. Weights auto-download, you don't need to grab anything manually.

For the CLI: mlxdlss-video convert in.mp4 out.mp4 --scale 1.5 --detail-strength 2 --temporal

Settings quick reference

• Scale: 1.5 to 2.

• - Detail strength: 2 is the sweet spot. 1 is subtle, 3 gets grainy.

• - For video, turn on temporal and set hp-history to 0.5.

• - Auto-mask skin when you're working on faces.

Best results on soft or 2x-upscaled footage. ComfyUI gets RAM-heavy with long clips, so use the CLI for those.

Credit to WaveCut (github.com/iamwavecut/MLX-DLSS). I'm not affiliated with NVIDIA. The weights carry the vendor's original terms, so don't redistribute them.