DLSS 5 FlashDepth (External, Optional)

Every depth node in this pack has a resolution ceiling baked in. Video Depth Anything's biggest setting is 518 pixels on the short side, which is great until your footage is 2K. This node is the pack's answer for that: it runs the official FlashDepth model from Eyeline-Labs, a temporal video-depth network built for higher-resolution footage, and hands you back a temporally consistent depth map ready for the DLSS pipeline.

The word "External" in its display name is doing real work. FlashDepth's official dependencies conflict with current ComfyUI builds - enough that the author decided not to add it to the pack's dependency list at all. So instead of pretending it could live in your ComfyUI environment, this node deliberately executes FlashDepth in its own Python environment and shuttles frames in and out. You set that environment up once by hand, following docs/FLASHDEPTH.md in the pack, and then this node just talks to it.

What you have to provide

Because it's an external backend, the node can't find things on its own. The required inputs include:

• flashdepth_python - the path to the isolated environment's Python interpreter (the one with FlashDepth's deps installed, not ComfyUI's).
• flashdepth_repository - where you cloned the official FlashDepth repo.
• fps (default 24) - the clip's frame rate, which the temporal model needs.

Leave those path fields blank and the node will refuse with a config error. This is not the node you reach for first; it's the node you reach for when you've actually outgrown everything else and you're willing to babysit an environment.

Picking a variant

• FlashDepth-L (low resolution) - what the upstream project recommends below roughly 518 pixels on the short side.
• FlashDepth Full (2K) - the high-resolution path, meant for footage around 2K and up.

If your material isn't genuinely high-resolution, you're paying the FlashDepth setup tax for nothing - the pack's own Video Depth Anything node will give you comparable temporal depth with zero extra environment. Use this one when the short side is well past 518 and you want the temporal consistency without downscaling your source into the floor.

The single output, temporally_consistent_depth, drops straight into the depth input of DLSS SR + Experimental Neural Rendering (Advanced) alongside a RAFT motion guide, which is exactly the graph in the optional high-res workflow 05_video_flashdepth_highres_2x.json. First real run downloads the FlashDepth checkpoint, so budget for a slow start and some disk.

Everything else about this node is the pack's shared story: Windows + NVIDIA only, runtime set up via DLSS Runtime Setup (One Click), install through ComfyUI Manager (https://github.com/HECer/ComfyUI-DLSS5). And remember the standard caveat - depth here is estimated from pixels, not read from an engine buffer, so treat the output as a very good guess rather than ground truth.