Nodes/ComfyUI-TJ_NODE/RTX Deblur (TJ)
ComfyUI Node

RTX Deblur (TJ)

RTX Deblur Won't Make Anything Bigger, and That's the Whole Point

By designloves2·Created 4 months ago·Updated a day ago· 13
RTX Deblur (TJ)
  • images
  • images
strengthMEDIUM
get_name(none)
setnode_nameRTX_Deblur

Upscaling gets talked about as one job, but it's three. Adding pixels to an image that's already sharp is interpolation - free, instant, and physically incapable of inventing anything. Adding detail that was never there is generation, and it will rewrite faces. Keeping either consistent across video frames is a third problem entirely, and it's the one that actually hurts.

RTX Deblur is firmly in the first camp, with a twist: it's a repair tool from the same hardware family as NVIDIA's super-resolution path, and it never changes your resolution. Input 1280x720, output 1280x720. The tooltip says it in as many words: "Deblur intensity. Resolution is never changed - this sharpens, it does not upscale." If you came here expecting an upscaler, this is the wrong node, and the pack would rather you knew that before you queued a 200-frame video.

What you'd actually reach for it for

Soft, smeared, or motion-blurred footage you want to keep the size of. Old camera clips, phone video with digital stabilisation mush in it, archival material you're de-noising anyway, or a generative video whose frames are slightly soft and where a real upscale would hallucinate texture your shot doesn't have.

The mechanism is worth knowing because it explains the limits: NVIDIA's SDK doesn't expose deblur as a separate effect. It's a quality level of the same VideoSuperRes filter the RTX upscaler uses - DEBLUR_LOW through DEBLUR_ULTRA. So there's no exotic model being loaded; it's the same hardware path, just configured to leave the output dimensions alone. The node processes your batch in chunks (it budgets roughly 16 megapixels per pass) and runs frames through it one at a time, which keeps VRAM flat instead of proportional to your frame count.

What it can't do is add detail that isn't there. The community's read on the NVIDIA restoration path is roughly this: it cleans up an image and fixes pixelation remarkably well, and it doesn't give you pores or fabric weave. That's a feature when your source is fine, and a disappointment when it isn't. For genuine detail recovery you want a generative restorer, not this.

Inputs and output

  • images - an IMAGE batch; video frames land here straight from a loader like VHS_LoadVideo.
  • strength - LOW / MEDIUM / HIGH / ULTRA, default MEDIUM. Start here. If the source is genuinely smeared, climb; if it's already decent, ULTRA is mostly going to make compression noise look confident.
  • get_name - the wireless receive slot. Populated by this pack's provider list, so IMAGE data can arrive without a wire across the graph.
  • setnode_name - defaults to RTX_Deblur; that's the name the output gets published under for other nodes to pick up.

One output: images, same resolution, same batch length, ready for a save or video encode node.

Install

The pack itself is the usual:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE.git
# restart ComfyUI

But this node is a thin wrapper and the runtime is the real dependency. You need an RTX GPU and NVIDIA's VFX SDK libraries, which arrive with the NVIDIA RTX video nodes pack - the README calls it comfyui_nvidia_rtx_nodes. Without it, the node fails immediately with a message telling you exactly that, plus the underlying import error.

Worth knowing why these three filters live here rather than as a patch to the NVIDIA pack: that pack is a git checkout, and modifying it in place leaves the tree dirty, at which point the ONE STUDIO updater skips it and freezes that pack at its current version forever. Duplicating three filters into TJ_NODE is the honest fix.

None of this pack's own pip requirements (llama-cpp-python, openpyxl, imageio, imageio-ffmpeg) matter for this node. Manager will install them anyway.

When it goes wrong

RTX Deblur (TJ): NVIDIA VFX SDK not available - either the NVIDIA RTX nodes pack isn't installed, or you're on a non-RTX card. This is a hardware path; GTX and AMD get nothing.

this NVIDIA VFX SDK build has no DEBLUR_... level - the node asks the SDK for a named quality level and tells you the SDK version it found if the lookup fails. An old NVIDIA runtime is the usual cause. Update that pack rather than this one.

Nothing looks different. That's plausible on clean footage - and it's the correct outcome. This node fixes damage; it doesn't apply a look.

Category ✨ TJ_Node/Image

Inputs (4)

NameTypeDefaultDescription
imagesIMAGE
strengthCOMBOMEDIUMDeblur intensity. Resolution is never changed — this sharpens, it does not upscale.
get_nameoptCOMBO(none)1 options: (none)
setnode_nameoptSTRINGRTX_Deblur

Outputs (1)

NameTypeDescription
imagesIMAGE