FL RTX VSR Restore

Why you'd reach for this one first

FL RTX VSR Restore runs NVIDIA's RTX denoise and deblur effects on an image or IMAGE frame batch without changing the resolution. If that sounds like the least glamorous node in the pack, you're missing how video workflows actually fall apart. Generated video is often slightly soft or slightly noisy, and if you upscale that mush, you've just made bigger mush - every defect gets magnified along with the pixels. The fix the community converged on is a two-pass recipe: deblur or denoise at native resolution first, then upscale the cleaned frames. In one of the most-upvoted write-ups of the NVIDIA RTX enhancement stack, the author reported the deblur-then-upscale order "produced much better results" and dropped their Topaz subscription over it. This is the deblur/denoise half of that recipe, and the pack's own README says the same thing: restoration "can be run before upscaling when restoration is needed."

It's also genuinely fast. Same NVIDIA VFX machinery as the upscale node, same DLPack GPU-side frame exchange, no CPU round trip - restoration at 4 GB VRAM and seconds per clip is very much the point of this hardware path.

How it works

Restore feeds your frames through the same VideoSuperRes effect the Upscale node uses, but instead of asking for a scale it asks for a same-resolution quality level built from your two dropdowns: an operation (denoise or deblur) and a strength. The node keeps input dimensions untouched - it pads internally for the runtime's row-alignment requirement and then crops back to exactly what you fed in, so the output is pixel-for-pixel the same size as the input.

The inputs that matter

• runtime - from FL RTX VSR Runtime Loader, same as the other node.
• images - one image or a decoded video frame batch.
• operation (default Deblur) - Denoise or Deblur. Deblur is the sharpening workhorse for soft or slightly out-of-focus AI video; Denoise cleans up noisy or grainy footage.
• strength (default Medium) - Low, Medium, High, or Ultra.

Output: a single images IMAGE, unchanged resolution, ready to feed straight into FL RTX VSR Upscale or on to your video-combine node.

Two honest cautions from people actually running this stack. Denoise is worth applying selectively on AI-generated video - it can scrub away the very grain that makes footage feel like footage, and over-aggressive denoise plus upscale can read as "plastic." And deblur at Ultra is where you'll most often see the softening side effect show up on already-clean sources; Medium to High is a sensible first pass. Strength is a per-clip judgment, not a "max it out" knob.

Wiring and install

The natural graph is Runtime Loader → Restore → Upscale → combine:

FL RTX VSR Runtime Loader ──runtime──> FL RTX VSR Restore ──images──> FL RTX VSR Upscale ──> combine/save

Install the pack the normal way - ComfyUI Manager, search ComfyUI-FL-RTXVSR, or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-RTXVSR.git

Restart ComfyUI. Same platform floor as the rest of the pack: 64-bit Windows 10+, an RTX 20-series-or-newer GPU, a current driver, Python 3.12+, and about 1.5 GiB free for the NVIDIA runtime the loader downloads and verifies on its first run. Don't expect this node to add resolution or rebuild detail - that's not the job. Its job is making the source worth upscaling, and on blurry or noisy video it's the difference between an upscale you keep and one you quietly delete.