RTX Video upscaling in ComfyUI: make the pipeline inspectable before making claims

ComfyUI pipeline note: Open workflows where a node graph, model file, and image pair can be inspected are useful only when the image workflow is shown separately. This article gives a controlled answer with a site-owned pair and a manifest that another reader can inspect.

The core work in DLSS 5 Studio is a style-guided redraw into a DLSS 5-style visual reference, with the desktop app as the handover. This pipeline article covers the limited entry point of inspecting poor-quality existing pictures; follow the before-and-after cases for the style-conversion story rather than treating upscaling as the product.

TL;DR

• Open pipelines need inspectable nodes and files.
• Keep the source image beside the graph output.
• Record graph, model, and export details.
• Do not infer hardware performance from one graph.

What the linked coverage does and does not say

ComfyUI pipeline note: The search phrase RTX Video upscaling ComfyUI points to a practical question: what can a person check after a feed, export, or tool has softened a reference? Our answer starts with the file, not with a vendor promise.

ComfyUI pipeline note: The public landing page is /image-upscaler, where the browser path is labeled as an independent visual workflow. The page is not a driver download and it does not claim to be a product from the company named in RTX Video and ComfyUI coverage.

ComfyUI pipeline note: The first source is ComfyUI repository (https://github.com/comfyanonymous/ComfyUI). It provides context for rtx video and comfyui coverage, while our sample files provide the evidence for the image-handling steps.

ComfyUI pipeline note: The second source is NVIDIA RTX Video page (https://blog.comfy.org/p/nvidia-vfx-nodes-for-video-upscaling). It is linked so a reader can check the surrounding report or product material instead of relying on a paraphrase in this article.

The input is the evidence

ComfyUI pipeline note: A current headline can explain why rtx video and comfyui coverage is being searched, but the headline is not an image license. We use site-owned files and retain their original filenames so the visual pair has a clear provenance.

ComfyUI pipeline note: For this angle, the most revealing crop is not a dramatic skyline or a logo. It is an ordinary edge: a building line, a product contour, or a small region where compression turns a smooth boundary into blocks.

ComfyUI pipeline note: Open the low-resolution input first. Read a few characters, look at the repeated texture, and note where your eyes stop trusting the picture. That observation is part of the record before any enhancement step is run.

ComfyUI pipeline note: Run the browser sample once and keep the resulting file beside the input. Do not replace the input halfway through the comparison, because a different source makes RTX Video upscaling ComfyUI impossible to reproduce as a check.

A reproducible run

ComfyUI pipeline note: Use the same displayed width for both files. A larger result can look more impressive simply because it occupies more pixels, so the comparison needs a stable viewing frame and a caption that names the two files.

ComfyUI pipeline note: Record the operation, input dimensions, output dimensions, and date. We do not fill an absent field with an estimate. A missing setting is a limitation to report, not a reason to invent a number.

ComfyUI pipeline note: If a face, sign, or label changes meaning, keep the original as the decision reference. The visual-enhancer step can make an edge easier to inspect, yet it cannot recover ground truth that was not present in open workflows where a node graph, model file, and image pair can be inspected.

ComfyUI pipeline note: For a developer, the useful artifact is a request manifest. It shows which endpoint the operator controls, which files were read, and where the output was written; it does not hide a provider key in a sample command.

Failure cases worth keeping

ComfyUI pipeline note: For a designer, the useful artifact is a side-by-side crop with a caption. The crop should reveal a concrete question such as whether a diagonal line is continuous or whether a repeated pattern has been softened.

ComfyUI pipeline note: For a researcher, the useful artifact is a repeatable condition. Keep the seed, prompt, tile size, export format, or node graph fixed when those fields exist. Change one condition only when the goal is to learn what caused a visual difference.

ComfyUI pipeline note: This is where RTX Video upscaling ComfyUI becomes a workflow term rather than a slogan. The phrase names a task people can try, while the article records what happened in one controlled sample and what remained unknown.

The accompanying captions say what is visible, so the image pair is not asked to carry a silent performance claim or a hidden comparison.

A practical handoff

ComfyUI pipeline note: This page does not provide a command-line tool. The example command sends only to an endpoint the operator controls. It does not require our hosted credits and it does not pretend to reproduce a vendor benchmark.

ComfyUI pipeline note: If you want to test a personal image, use only a file you have permission to process. Keep client material local when the client has not agreed to an upload, and preserve the untouched source for later review.

ComfyUI pipeline note: The durable lesson from rtx video and comfyui coverage is therefore modest: make the transformation visible, make the source traceable, and write down the failure case. That is more useful than claiming a universal result from one attractive crop.

ComfyUI pipeline note: Readers can follow the numbered procedure below, then open the image-upscaler page for a cached example. The online path and the separate desktop request path have different availability and should not be merged in the copy.

Reproduction command

python run_workflow.py --input architecture-low.jpg --output architecture-output.jpg --record metadata.json

Sources and disclosure

ComfyUI pipeline note: This independent project is not affiliated with NVIDIA, Adobe, Tesla, Topaz Labs, or ComfyUI. The names are used to identify the subject of the linked sources; they do not imply endorsement, sponsorship, or shared ownership.

ComfyUI pipeline note: The boundary is explicit: this article does not publish frame rates, does not promise a fixed quality change, and does not turn a visual pair into a hardware test. It documents an inspectable image operation.

A node graph is part of the evidence

ComfyUI pipeline note: Open ComfyUI workflows can be copied, forked, and changed. That flexibility is useful only if the graph, model reference, input file, and export step are named together. A screenshot of a graph without those pieces is a visual hint, not a reproduction.

ComfyUI pipeline note: Start with a tiny graph and one site-owned image. Export the graph, record the model identifier, and keep the output beside the source. If a node is unavailable or a model license does not permit your use, stop and replace it with a permitted option instead of silently substituting a different system.

ComfyUI pipeline note: The RTX Video discussion adds an important boundary: a pipeline label does not prove that every card, driver, or runtime supports the same operation. This article therefore reports the graph and the image pair, not frame rate or cross-card performance.

ComfyUI pipeline note: Readers who want to compare two graphs should change one node at a time and keep the export settings fixed. That makes a soft edge or a changed texture a question about a specific node rather than a vague argument about AI enhancement.

Reproducible steps

ComfyUI pipeline note: 1. Open the site-owned input named in the caption. ComfyUI pipeline note: 2. Run the topic-specific workflow once and save the output with the date. ComfyUI pipeline note: 3. Compare both files at the same display size. ComfyUI pipeline note: 4. Record dimensions, operation, source, and one limitation.

Reproduction command

python run_workflow.py --input architecture-low.jpg --output architecture-output.jpg --record metadata.json

Sources and disclosure

• ComfyUI repository: https://github.com/comfyanonymous/ComfyUI (checked 2026-10-16).
• NVIDIA RTX Video page: https://blog.comfy.org/p/nvidia-vfx-nodes-for-video-upscaling (checked 2026-10-16).
• Landing page: https://www.dlss5nvidia.com/image-upscaler

ComfyUI pipeline note: Independent project, not affiliated with NVIDIA or the other named companies.

Related reading

Read the documented style-conversion case · Open the existing-image tool