DLSS 5 Video Depth Anything (Temporal)

Per-frame depth estimation has a known disease: it flickers. Each still is fine, but run it over a clip and the scale pumps and edges waver, which is poison for a temporal upscaler. This pack's first answer was Depth Anything V2 plus a stabilization node bolted on afterward. This node is the better answer - it runs Video Depth Anything Small (VDA-S), the temporal sibling from the same Depth Anything team, which reasons over 32-frame windows and produces depth that's consistent because the model itself is temporal, not because something smoothed it afterward.

The pack's own docs make the call easy: for new video projects, this is the recommended depth path (workflow 04_video_vda_small_temporal_2x.json), replacing the framewise node plus stabilizer chain entirely. RAFT stays around for the separate motion-vector guide; depth is simply this node's job now.

How it works

On first use the node clones the official VDA source at a pinned commit into ComfyUI/models/dlss5/video_depth_anything/ and downloads the official Small weights - that's the Apache-2.0 source and official vits checkpoint, and yes, the first run can sit there looking dead while Git and Hugging Face do their thing. After that it evaluates the clip in 32-frame temporal windows with upstream overlap alignment, which is the mechanism that keeps the depth from drifting as the window slides. On post-Ada GPU architectures it automatically switches to PyTorch's scaled-dot-product attention.

The two inputs

• input_size - 518 (best), 392 (fast), or 280 (compatible). 518 is the quality setting when your GPU supports it. 392 is the sensible preview/regular default. 280 exists for the edge case the README calls out: a brand-new GPU whose xFormers kernel the model doesn't know about yet.
• precision - FP16 (recommended) or FP32. FP16 is the daily driver - half the memory, near-identical quality - and the default expectation. FP32 is the fallback when FP16 misbehaves on your hardware.

That's the whole surface. The single output, temporally_consistent_depth, plugs into the depth input of the Full Pipeline node next to a RAFT motion guide.

Where it stands and where it stops

Two practical notes. First, its resolution ceiling: 518 on the short side is the biggest window, which is fine for typical 1080p-ish work but not for genuine 2K+ footage - that's what the optional FlashDepth node in this pack is for. Second, licensing is a quiet win here: unlike Depth Anything V2's Base and Large weights (non-commercial), the VDA-S source and weights this node pulls are Apache-2.0, so it's the commercially clean depth choice in the pack.

Everything around it is the pack's shared ritual: Windows + NVIDIA, the runtime installed through DLSS Runtime Setup (One Click) with a legally obtained nvngx_dlssnr.dll, health-checked via DLSS 5 Runtime Status, and the pack itself installed from https://github.com/HECer/ComfyUI-DLSS5 through ComfyUI Manager. And the usual honesty clause applies - this is pixel-estimated temporal depth, excellent for guiding an upscaler, not the engine's real depth buffer, so don't expect game-perfect geometry. Start with a short clip, confirm the frame count and dimensions, and let the first full run download its weights on your schedule, not mid-render.