SeedVR2, LTX 2.5, Topaz + 5 more: video upscalers compared

Eleven ways to double the resolution of a video clip, run on the same five clips, with every output at exactly 2x the source width and height, the same frame rate and the same frame count. No frame interpolation, and denoise, sharpening and face enhancement held off wherever they can be turned off. The running order is by what the method needs rather than by what it costs: an ffmpeg resize that needs no GPU at all, then the upscaler inside the NVIDIA driver, then Real-ESRGAN, then the methods that want 12 GB, then the ones that want an 80 GB datacentre card, and the two paid services last. The curve that produces is not a straight line, which is most of why the comparison was worth running. Everything local ran on an RTX 3060 with 12GB of VRAM, an i5-13400F and 32GB of system RAM, in ComfyUI 0.34.0. The 80 GB tier ran on a rented A100, which cost $6.67 in total. Local and rented timings are not comparable with each other and the video says so where it matters. If you have a bigger card you should get better results. The test clips are generated rather than filmed, so what you are watching is how these upscalers behave on synthetic footage. Ten of the eleven methods are general-purpose video upscalers; the MiniMax H3 latent upscaler is the exception, because it round-trips through H3's own latent space and only works on H3 output. ── THE COMPARISON PAGE ────────────────────── Every clip in this video, playable side by side at up to 4x zoom, with the workflows, the models and the measurements behind each one. https://loopforge.cc/projects/video-u... ── THE MODELS ────────────────────── SeedVR2, which ComfyUI supports natively. https://docs.comfy.org/tutorials/util... LTX 2.5 and the Pixel Spatial Upscaler IC-LoRA. https://huggingface.co/Lightricks/LTX... Real-ESRGAN. https://github.com/xinntao/Real-ESRGAN ── THE RENTED GPU ────────────────────── The 80 GB tier ran on a rented A100 rather than on my own hardware. Affiliate link: Loop Forge gets credit if you sign up through it, at no extra cost to you. https://runpod.io?ref=5qkyhesp ── WHAT EACH ONE NEEDED ────────────────────── Lanczos resize no GPU 2 to 22 s RTX Video SR 1.0 GB 15 to 25 s Real-ESRGAN x2plus 3.9 GB 85 s to 2 min 35 s MiniMax H3 latent 12.0 GB 4 to 8 min SeedVR2 3B 12.1 GB 17 to 60 min SeedVR2 7B fp16 66.0 GB 2 to 7 min LTX 2.5 BF16 79.3 GB 4 to 9 min Peak VRAM is total GPU memory in use during the run, sampled from nvidia-smi, so it is an upper bound rather than the method's own footprint. The 12 GB rows ran on the 3060 and the 80 GB rows on the A100, so the times in the last column are only comparable within a tier. ── CHAPTERS ────────────────────── 0:00 The same shot twice, and the running order 0:15 Lanczos resize, no GPU 0:45 NVIDIA RTX Video Super Resolution 1:16 Real-ESRGAN x2plus 1:46 MiniMax H3 latent upscaler 2:16 SeedVR2 3B, four frames 2:45 SeedVR2 3B, twenty-one frames 3:13 SeedVR2 3B, whole clip 3:43 SeedVR2 7B fp16 4:15 LTX 2.5 BF16 4:55 Topaz Starlight 5:26 Flux 3.0 5:54 What to use 6:47 Close ── METHOD AND LIMITS ────────────────────── Five clips per method is a small sample and the clips were chosen to be different from each other rather than to be representative. The two paid services take a target resolution rather than a scale factor, so neither landed on exactly 2x and both were resampled to the comparison size; Topaz was downscaled on every clip, which is a supersampling advantage over the methods that rendered natively at 2x, and Flux was upscaled on two of them. Neither can be reproduced from the page, because there is no workflow to publish and no checkpoint to pin. FlashVSR was planned and never ran, so nothing here says anything about it. Quality is called by eye. The only numbers on screen are wall-clock time and peak VRAM. If you upscale generated video, which clip in this set is closest to the footage you actually work with? The five here are a face, fine detail, particles, flat graphics and fast motion, and I suspect the answer changes which method I would recommend. TOOLS · ComfyUI · SeedVR2 · LTX 2.5 · Real-ESRGAN · NVIDIA RTX Video Super Resolution · MiniMax H3 · Topaz Starlight · Flux 3.0 · RunPod · ffmpeg · Claude Code #videoupscaling #seedvr2 #ltx25 #comfyui #aivideo #upscaler #realesrgan #topaz #localai #openweights #rtx3060 #lowvram #a100 #minimaxh3 #claudecode