Topaz Video Enhance

Topaz's Video AI is the industry-standard tool for making old, soft, low-res footage look like it was shot this decade. This node is that engine - upscaling and frame interpolation - running as a ComfyUI node over Topaz's API. You feed it a video, it comes back sharper, higher-resolution, and optionally smoother, without you ever touching the Topaz desktop app.

This is the second version of the node, and the "V2" is the part worth caring about: instead of a pile of separate toggles for every model, the inputs are dynamic combos. Pick Astra 2 and the node grows a creativity slider, a prompt box, and sharpness and realism dials. Pick Starlight (Astra) Creative and you get a low/middle/high creativity dropdown. The interface only shows you what the current model can actually use, which kills a whole class of "why is this setting being ignored" confusion. The older TopazVideoEnhance node still exists in the menu, marked legacy - this one replaces it.

How it works

It's a cloud job, not a local render. ComfyUI uploads your MP4 to Topaz's pipeline, creates a task, uploads the file, then polls until the enhanced video is ready and pulls it back down. That's why it needs a Comfy account with credits and an internet connection - but also why it'll happily upscale a 4K render on a machine that couldn't hold the footage in VRAM. The node validates that your input is actually MP4, so a .mov or .webm will fail fast with a clear error.

The two knobs that matter most:

• upscaler_model - this is the quality-vs-cost decision. Astra 2 is the current flagship (creative control, an optional descriptive scene prompt, sharpness/realism sliders), Starlight (Astra) Fast is the cheap daily driver, Starlight Precise 2.5 sits between them for fidelity, and Creative is the heavier re-imagining pass. "Disabled" turns off upscaling entirely, which is only useful if you just want interpolation.
• interpolation_model - currently apo-8 or Disabled. Apo-8 synthesizes in-between frames, which is how you go from a 24fps clip to 60fps, or get slow-motion without the judder. It exposes an output frame rate (15–240) and a slow-mo factor. Frames are cheap relative to an upscale, so this is where people stretch short clips.

The third input, dynamic_compression_level (Low/Mid/High, default Low), is the CQP encoding quality for the output - leave it at Low unless you're fighting banding.

Gotchas

The two big ones, straight from the node's own checks: Astra 2 caps at 9000 input frames (about five minutes at 30fps), and if you give it a scene prompt, the cap drops to 450 frames (~15s). Long clips plus a prompt will fail with a specific error - split the clip. And if you set both upscaler and interpolation to Disabled, it refuses to run, because there's literally nothing left to do.

Cost-wise, every frame is billed, so long 4K clips add up fast - the node shows a credits-per-frame estimate so there are no surprises. It's not a free local upscaler, and it shouldn't be: this is the "more pixels over time" job from the upscaling playbook, where temporal consistency is the whole game. If your source is already decent and you just want more pixels, the community has compared this against local NVIDIA RTX video super-resolution and some people dropped their Topaz subscriptions entirely. But for genuinely degraded footage, or anything you need at true 4K without a monster GPU, this is the reliable answer.