AI video upscaling to 4K locally: SeedVR2, RTX Video Super Resolution and Topaz on RTX PRO

Upscaling AI clips and old footage to 4K on your own machine

To upscale video to 4K on your own computer, a company has three routes: SeedVR2, an open model from ByteDance Seed under Apache 2.0 that runs natively in ComfyUI; NVIDIA RTX Video Super Resolution, an upscaler NVIDIA describes as much faster than other local tools, available as a ComfyUI node; and Topaz Video, a commercial desktop app with local rendering. A workstation with one NVIDIA RTX PRO 5000 Blackwell, with 48 or 72 GB of GPU memory, runs the larger SeedVR2 model at full precision and leaves memory for longer frame batches at 4K; a 24 or 32 GB card handles the same jobs with smaller batches, more tiling or a smaller file.

Upscaling means turning a low-resolution video into one with more pixels, with a model that adds plausible detail instead of stretching the image. A 4K frame (3840 × 2160) has about 8.3 million pixels, four times a Full HD frame and nine times a 720p frame, and a 30-second clip at 24 frames per second has 720 frames, by our arithmetic. This article is for the marketing or video lead who wants that work on the company's own workstation, and for the IT colleague who sets it up. It covers two jobs: AI clips that come out of a generator at 480p or 720p, and existing footage such as older product videos and event recordings. Generating the clips is covered in our local alternative to Runway and Kling, our guide to video generation GPU requirements and our comparison of the AI video workstation for marketing tiers.

Why generators leave the step to 4K for later

Most open video models generate at 480p or 720p: Wan 2.2 lists its text-to-video and image-to-video models as supporting "480P & 720P", and the presenter models InfiniteTalk and LongCat-Video-Avatar-1.5 do the same. Memory and time grow steeply with the number of pixels, so teams draft small and upscale the clips they keep. NVIDIA's blog of 10 March 2026 says that "most prefer to generate smaller, faster previews first, and then upscale them".

LTX-2.5 is the exception among the generators: its card gives settings for 3840 × 2176 output (--spatial-upscalings 2 --width 3840 --height 2176) inside its own pipeline. Lightricks also publishes an upscaler for its model, the LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler, which "creatively upscales a low-resolution video, synthesizing fine detail rather than simply interpolating", in a 2× variant. Its card adds that it is "not intended for upscaling live-action footage where factual fidelity to the source matters". For anything that is not an LTX clip, the routes below apply.

Three local AI video upscalers compared

Comfy's SeedVR2 guide, the ByteDance-Seed/SeedVR repository, NVIDIA's blog of 10 March 2026, the nvidia-vfx package page, Topaz Labs' product and system-requirements pages and Lightricks' model cards, read 30 October 2026.

SeedVR2 trades time for detail, RTX Video Super Resolution is built for speed by NVIDIA's account, and Topaz Video suits an editor who prefers a finished application. All three run locally; the differences are in licence, operating system and how much the result may differ from the source.

SeedVR2: the open model in ComfyUI

Comfy's guide describes SeedVR2 as "a one-step diffusion-based video restoration model" from ByteDance Seed, and ByteDance's repository states: "SeedVR and SeedVR2 are licensed under the Apache 2.0." One-step means the model restores the frames in a single denoising pass instead of the many passes a video generator runs. ComfyUI has supported it natively since a merged pull request (#14424) that added SeedVR2 attention, VAE tiling and temporal chunking, and its templates include "SeedVR2 3B Int8: Upscale Video". ComfyUI itself is under GPL-3.0 and actively developed, with several releases in October 2026.

File list of Comfy-Org/SeedVR2 on Hugging Face, read 30 October 2026; the shared VAE adds 501 MB, and a 7B "sharp" checkpoint comes in the same sizes.

The guide gives no absolute memory figures: the 3B model has "lower VRAM requirements, suitable for users with limited hardware", while the 7B "delivers higher quality results at the cost of increased VRAM and processing time". The maintainer of the community node ComfyUI-SeedVR2_VideoUpscaler (Apache 2.0; its last release came out in December 2025) gives three levels: 8 GB or less with heavily quantised GGUF files, block swapping and VAE tiling; 12 to 16 GB with FP8 files; and "24GB+: Use FP16 models for best quality and speed without memory optimizations". ByteDance's own inference code states that "1 H100-80G can handle videos with 100x720x1280" and that four such GPUs "further support 1080p and 2K videos". That card belongs in a data centre and is outside our range; the figure shows how fast the need grows with resolution.

The number of frames processed together matters for quality. The community node asks for batches of 4n+1 frames (5, 9, 13 and so on), with a "Minimum 5 for temporal consistency", and recommends matching the shot length where possible. A card with more memory holds longer batches at 4K and needs fewer tiles. Speed depends on the model size and precision, the output resolution, the batch, the tiling and the card; the node's notes say "The VAE is slow", the step that turns the model's compressed frames into pixels. We found no published SeedVR2 timing on an RTX PRO card.

ByteDance lists the limits itself. The models can fail on heavy degradation and very large motion, and they "tend to overly generate details on inputs with very light degradations, e.g., 720p AIGC videos", which can leave "oversharpened results occasionally (especially on small resolutions, e.g., 480p)". Comfy's guide adds that attention backends "can silently return black or corrupted output on quantized checkpoints"; turning FlashAttention off fixed all-black INT8 output on an RTX 3090, and switching to the FP16 or FP8 file of the same size is the other remedy. Test the 3B and 7B files on one of your own clips before you run a batch.

You can order the workstation alone, the setup of ComfyUI and SeedVR2 alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

NVIDIA RTX Video Super Resolution in ComfyUI

NVIDIA's blog of 10 March 2026 announced that "users can quickly upscale generated video to 4K with NVIDIA RTX Video Super Resolution, available as a node for ComfyUI", "30x faster than alternative popular local upscalers" and "at a fraction of the VRAM cost". The blog names neither the GPU nor the tools behind that comparison. The node's repository, Comfy-Org/Nvidia_RTX_Nodes_ComfyUI under Apache 2.0, says it "Only works on Nvidia RTX GPUs" and installs through the ComfyUI manager.

The node calls NVIDIA's nvidia-vfx package, "Python bindings for NVIDIA Video Effects SDK", version 0.2.0.0 of 30 September 2026, marked "3 - Alpha". It upscales "SDR video up to 4x while reducing compression artifacts" in 2×, 3× and 4× steps, with denoise, deblur and high-bitrate modes. It runs on 64-bit Windows 10 or later and on Ubuntu 20.04 to 24.04, Debian 12 and RHEL 8 and 9, with Linux drivers 570.190, 580.82 or 590.44 and later, on GPUs with Tensor Cores from Turing to Blackwell and Hopper. The package carries a proprietary NVIDIA licence, and its page has no licence text and no statement on commercial use, so read the terms delivered with the package before client work. For the same reason it is not part of the software we preinstall; your team adds it from the ComfyUI manager after reading NVIDIA's terms. The package lists GPU architectures, not cards, and the node says RTX GPUs; for data-centre cards such as the L4 and L40S, test before you plan a server around them.

Topaz Video as a commercial desktop app

Topaz Video is a desktop application with "Over 20 AI enhancement models" for upscaling, frame interpolation, stabilisation and denoising, and its product page says: "Your files never leave your computer with local rendering." The app still needs internet access "for activation, model downloads, updates, and cloud rendering", and part of its Starlight family is offered as cloud rendering, which uses monthly cloud credits.

Its system requirements, read on 30 October 2026, state that "Linux Operating Systems are not supported"; it runs on Windows 10 or 11 and macOS 13 or later. NVIDIA cards need at least 6 GB of GPU memory, with 8 GB or more recommended; the local Starlight models ask for 10 to 16 GB at minimum, and Topaz recommends 32 GB for Starlight Fast 2. System memory starts at 16 GB, with 32 GB or more recommended. The page names no professional cards; an RTX PRO Blackwell card exceeds the compute-capability levels it lists. Topaz is sold as a subscription, and the plan matters for a business: plans marked "Limited commercial use", the Topaz Video desktop plan among them, are defined as "Personal and commercial use for orgs under $1M USD annual revenue", while the Pro plans list "Full commercial use". Topaz licences are bought from Topaz Labs; we supply the workstation it runs on.

AI clips or old brand footage

AI clips at 480p or 720p. These are the inputs ByteDance warns about: very light degradation, where SeedVR2 can add too much detail. Run a short clip through the 3B and the 7B model and through RTX Video Super Resolution, and compare the three by eye on your own material; for previewing a whole campaign, the NVIDIA node is the faster tool by NVIDIA's account. Clips made with LTX-2.5 can stay inside LTX's own 4K pipeline.

Older footage. Comfy's guide lists "Conservative upscaling" among SeedVR2's capabilities, which "preserves original structure and detail while enhancing clarity". For damaged material, Lightricks publishes the LTX-2.5-22b-IC-LoRA-Restore, "An In-Context LoRA that restores old and damaged footage". It targets "compression damage, tape and sepia casts, flicker, dirt and scratches", colourises black-and-white scans from the scene content and aims at "delivery at HD to 4K class". Its card warns that "the pass is a switch, not a gentle enhancer", it needs progressive frames, so interlaced tape is deinterlaced first, and access on Hugging Face requires accepting its conditions. It falls under the LTX-2.x Community License: free for companies below USD 10 million of annual revenue, with a paid licence from Lightricks at or above that level. LTX-2.5 and its LoRAs are not part of the setup we deliver; a company that meets the licence terms installs them itself.

Keep the original files in every case. A generative upscaler invents detail, which is fine for a social clip and wrong for footage that documents a product, a site or an event as it was.

Which workstation for which upscaling job

GPU memory decides the model size, the precision and how many 4K frames go through at once; more memory means fewer tiles and longer batches. Upscaling occupies a card while it runs, so each card serves one job at a time.

Card specifications from NVIDIA datasheets as listed on our GPU pages; memory levels from the community node's README and our arithmetic for LTX-2.5. SE = Server Edition.

Entry. The NVIDIA RTX PRO 4000 Blackwell has 24 GB at 145 W in a single slot, the RTX PRO 4500 Blackwell 32 GB at 200 W. Both reach the node maintainer's 24 GB line for FP16 and exceed the memory and compute-capability levels Topaz lists. Our sizing rule counts 0.9 × the memory the driver reports, minus 3 GiB, as usable: 18.5 GiB on the 24 GB card and 25.7 GiB on the 32 GB card. The 7B file in FP16 with the VAE takes 17 GB, or 15.8 GiB, which leaves under 3 GiB for frames on the 24 GB card, so the 3B model or the FP8 7B file suits 4K work there better; the 32 GB card leaves about 10 GiB, by our arithmetic. How much of a 4K job fits without tiling is not published; tiling and shorter batches make it fit at the cost of time.

Mid. The RTX PRO 5000 Blackwell has 48 or 72 GB at 300 W and 1,344 GB/s. The 7B file in FP16 and the VAE take 15.8 GiB, which leaves about 24 GiB of the 40.0 GiB usable on the 48 GB card and about 46 GiB of the 61.5 GiB on the 72 GB card for frames, by our arithmetic; the 72 GB card can also hold a generator such as Wan2.2-TI2V-5B, which Wan states at 24 GB, next to the upscaler. Our page on the RTX PRO 5000 72 GB workstation covers the build.

Large workstation. LTX-2.5's restore and upscaler LoRAs run on its 22B model, whose transformer and 12B text encoder take 68.3 GB, or 63.6 GiB, in BF16 by their published files. That fits within the 83.0 GiB our rule counts as usable on a 96 GB RTX PRO 6000 Workstation Edition or Max-Q; see our RTX PRO 6000 Workstation Edition page.

Server. For an archive of thousands of clips, a server runs one job per card. The community node offers a "Multi-GPU standalone CLI" and "Batch Directory Processing", and notes that "Multi-GPU doubles VRAM usage (each GPU loads full models) but roughly halves processing time". A server with four RTX PRO 6000 Server Edition draws about 3.4 kW by our estimate; our AI server for media covers transcripts, subtitles and search for the same archives. The L40S has 48 GB in a passive card; check RTX Video Super Resolution on it first.

Describe a typical clip with its resolution, length and target format, and we suggest the tier that fits your volume and budget.

Drivers, frames and storage for IT

• Operating system. SeedVR2 in ComfyUI and nvidia-vfx run on Linux and Windows; Topaz runs on Windows and macOS only, so a Linux server excludes it.
• Drivers. nvidia-vfx asks for Windows driver 570.65 or later (595 or later for cards in TCC mode) and Linux 570.190, 580.82 or 590.44 or later.
• Frames. Keep SeedVR2 batches at 4n+1 frames; for multi-GPU runs the node blends overlapping frames between chunks.
• Black frames. On quantised files, test without FlashAttention first, as Comfy's guide suggests.
• Storage. Keep originals, upscaled masters and the workflow file with each job, so a result can be reproduced.

FAQ