RTX Deblur (TJ)

First, kill the name. "RTX Deblur" sounds like it belongs to the upscaler family, and it's a trap - the node's own tooltip spells it out: "Resolution is never changed - this sharpens, it does not upscale." You feed it a soft or motion-blurred image, you get the same dimensions back, just crisper. That honesty is the whole point, and it's why this little node is worth a slot in your graph when your renders come out of the sampler looking like they were smeared with vaseline.

AI video and anime renders are chronically soft - motion blur baked in at generation, diffusion averaging everything toward mush. The community answer used to be "throw SeedVR2 at it," but SeedVR2 is a generative restorer: it invents detail, and in doing so it rewrites identity (the modidex upscaling notes are blunt about this - it can change a face). Deblur is the opposite school of thought. It runs on NVIDIA's RTX VFX silicon, treats the frame as signal, and sharpens what's actually there without hallucinating pores and lashes that were never in the source. For a blurry-but-decent frame, that's not a compromise, it's the correct tool.

What it actually does under the hood

The node wraps the same NVIDIA VFX SDK (nvvfx) that powers the RTX video super resolution upscalers. The SDK doesn't expose deblur as a separate effect - it's a quality level of the VideoSuperRes filter. Levels 12–15 map to DEBLUR_LOW through DEBLUR_ULTRA, and NVIDIA documents them as "same-resolution processing." The implementation sets output_width/output_height to the input's and never touches a scale factor.

It also batches: a long clip isn't staged on the GPU all at once, it's processed in ~16MP chunks, which keeps VRAM sane on a mid-range card. The same author uses this SDK for the MiniMax H3 gallery's deblur stage in the same pack - that one runs over stitched video, this node runs over whatever IMAGE you hand it.

The two inputs that matter

That's the whole interface - this is a deliberately tiny node:

• images - any IMAGE batch, a still or a sequence of frames. Feed it a batch and every frame gets sharpened.
• strength - LOW / MEDIUM / HIGH / ULTRA, default MEDIUM. Start at MEDIUM. ULTRA is for genuinely wrecked footage and it can ring edges on clean input.

Output is a single images socket, same resolution, which you wire straight into whatever you were about to save. There's nothing else to configure, and that's the feature.

Installing it (the part people get wrong)

You install the pack like any other - ComfyUI Manager, search "ComfyUI-TJ_NODE_STUDIO_ONE", or:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE.git

But the pack itself is the easy half. The hard requirement is the RTX VFX backend: this node is a thin wrapper around the NVIDIA VFX SDK, which arrives via the Comfy-Org/Nvidia_RTX_Nodes_ComfyUI node pack, and it only runs on an NVIDIA RTX GPU with a reasonably current driver. No RTX card, no AMD, no Mac, no iGPU - the import nvvfx will fail before anything else. And Manager does not auto-install dependency node packs, so after installing this pack run the included one-time script from the package folder and then restart ComfyUI:

bash install_requirements.sh   # Windows: double-click install_requirements.bat

That VFX backend is genuinely the fiddly bit - the nvidia-vfx pip package declares NVIDIA's wheel-stub as its build backend, which pip doesn't always fetch, and pip's isolated build env doesn't inherit your global build tools. The install script handles it with a wheel-stub preinstall plus a --no-build-isolation retry, which is why the author kept shouting "run the script once" in the README. If you skip it, the startup console prints a banner listing exactly which packs are missing.

Common issues

• "RTX Deblur initialization failed" - nvvfx isn't installed. You skipped the install script, or the RTX nodes pack isn't present. Run it, restart.
• "No DEBLUR_ULTRA level" - your SDK build is older and doesn't expose that quality tier. Drop to HIGH; the error tells you the SDK version.
• "Why didn't it get bigger?" - because it sharpens, it doesn't upscale. That's the design. If you want pixels and sharpness, chain this into a real upscaler - the community's cheap rung is DeBlur-then-VSR on RTX, roughly 4GB VRAM and seconds per clip.

It's not going to rescue a genuinely ruined image, and it won't add detail that was never captured. For the far more common case - a soft render you want to look clean and stay looking like the person you generated - it's the cheapest honest fix on an RTX card.