VOSR2, the Open-Weights Restorer
The VOSR2
Image Upscaler,
No Local Rig.
VOSR2 is the one-step diffusion image restorer from CVPR 2026 — Apache-2.0 weights you install on a CUDA rig yourself. No rig, no problem: the same open-weights restoration, photos rebuilt to crisp 4K in the browser, is hosted here on SeedVR2.
Model Options
One restoration job, two hosted engines, no rig
People searching VOSR2 either want its restored stills without the GPU that draws them — or a fair head-to-head. Both live on the same balance here.
SeedVR2
SeedVR2 is the other famous open-weights image restorer — a single-pass diffusion refiner trained on real high-res photos, so it rebuilds skin, fabric, and text instead of just enlarging them. It is hosted here: upload, pick 2K, 4K, or 8K, and the credit price is quoted before anything runs.
Ultimate Image Upscaler
For quick resolution boosts on product shots, scans, and screenshots, the Ultimate Image Upscaler sharpens a still in seconds on the same prepaid balance — no checkpoint, no Python environment, no VRAM ceiling to debug before a deadline.
Why use it on Moosky AI
What the rig quietly charges you for
The weights are free to download; the card, the software, and the hours are where open-weights restoration actually gets paid for.
No GPU shopping
VOSR2’s 1.4B checkpoint wants a roomy CUDA card for the VAE, vision encoder, and diffusion pass. On the hosted lane the render farm is someone else’s problem: any browser laptop pulls restored 4K stills off SeedVR2.
No checkpoint downloads
Running VOSR2 yourself means pulling the Qwen-Image 2D VAE, the SD2.1 base and decoder, a DINOv2 cache, and the weights into an exact folder tree. Here nothing downloads; the engine is already on.
No torchrun rig
The DIY path is a conda environment, pinned requirements, and a torchrun launch on hardware you own — plus the version conflicts that arrive with it. Moosky is an account and a composer; the plumbing never becomes your side project.
Your laptop stays your laptop
A local render pegs the GPU for the whole job — fans up, thermals up, everything else slows down. Hosted restoration streams back while your machine keeps running the actual workday.
Credits without a fuse
A quiet month on a rig still amortizes the card; a quiet month here costs nothing. Credits buy at 100 per dollar, never expire, quote per run before dispatch, and failed generations refund automatically.
Clean, commercial by default
Every Moosky export arrives watermark-free and is cleared for client work, prints, product listings, and monetized channels under commercial-use terms — no license archaeology through a third-party VAE’s fine print.
How It Works
From typed name to delivered still, without the install
Open the studio
Create an account, take the starter credits — no GPU, no weights to download, no Python environment.
Upload the photo
Drop in the still you want restored — a photo, a scan, a screenshot, or one frame pulled from a clip.
Pick the target
Choose 2K, 4K, or 8K; the composer quotes the exact credit cost before anything renders, on the hosted SeedVR2 engine.
Download clean
Collect the restored, watermark-free file and use it straight in print, a deck, or a listing.
Two things people want when they search ‘vosr2’ — this page is honest about both
Some searchers already own the rig: a CUDA card, the conda environment, the checkpoint folder with the Qwen-Image VAE and DINOv2 cache arranged just so. For them VOSR2 is the destination and this page simply says it straight — Moosky does not host VOSR2 or any local one-step diffusion checkpoint, and if you want that exact file on that exact card, the open-source repo is still the place to get it.
The other half typed the name off a paper thread and do not own the rig — or do not want to. That is who this page converts: no GPU, no multi-gigabyte downloads, no torchrun, no machine hogged for a render. The same open-weights restoration — photos rebuilt to crisp 4K, structure held, detail recovered — runs in the browser on hosted SeedVR2, on per-run credits that never expire and export clean for client work.
Credit-based
Simple, honest pricing
No GPU, no subscription, no expiring quota. Credits quote on screen before every run, buy at 100 per dollar, and never expire — one balance covers restoration, upscaling, and the video a restored still can become.
More from Moosky AI
Explore our complete suite of AI tools.
Frequently Asked Questions
What is VOSR2?
VOSR2 is the default one-step (1.4B) checkpoint of VOSR, a CVPR 2026 image super-resolution model. Unlike most recent restorers, it is not a text-to-image generator bent toward restoration — the backbone is trained from scratch on visual data only, with synthesized low-to-high pairs, a vision-encoder guide, and off-the-shelf VAEs. It ships Apache-2.0 weights on Hugging Face as a research release, distributed as files you download and run yourself rather than as a hosted service.
Is VOSR2 a video upscaler?
No. VOSR2 is a still-image restoration model — it has no video mode at all. Looping a single-frame refiner over a moving clip makes each frame’s texture an independent invention, so edges and faces shimmer between frames. If your source is a clip rather than a photo, the video lane (hosted FlashVSR) is the right tool; VOSR2 and the comparable SeedVR2 image engine are both built for single stills.
Can I run VOSR2 in the browser?
Not on Moosky — we do not host VOSR2, so any browser tab advertising ‘VOSR2 online’ is a third party running someone’s rig behind a paywall. Running it yourself means a conda environment, pinned requirements, downloading the Qwen-Image VAE, the SD2.1 base, a DINOv2 cache, and the weights into an exact folder tree, then a torchrun launch on a GPU you own. The honest browser answer to the same intent is the hosted SeedVR2 image restorer on this page.
How much VRAM does running VOSR2 take?
The VOSR2 release is the 1.4B variant, so budget for the backbone plus the Qwen-Image 2D VAE, the SD2.1 base and lightweight decoder, and a DINOv2 encoder all resident at once — a modern CUDA card with generous VRAM, and the conda plumbing to keep their versions compatible. On the hosted SeedVR2 lane there is no VRAM at all: the still is restored server-side and returns in your browser.
What is the difference between VOSR2 and SeedVR2?
Both are open-weights one-step image restorers, but they are different training bets. VOSR2 trains a vision-only backbone from scratch to stay anchored to the input with fewer hallucinations, which makes it the safer pick when inventing detail is a liability — text in a screenshot, a document scan, an unchangeable face. SeedVR2 is an adversarial diffusion refiner that rebuilds believable high-res texture, and it is the one Moosky actually hosts. The comparison article walks both sides and links the primary sources.
Can I use the upscaled photos commercially?
Everything you generate here comes out watermark-free and is cleared for client work, products, ads, and monetized channels under Moosky’s and the underlying provider’s terms. If you run VOSR2 directly instead, what you may do with an output follows that release and its reused components — the SD2.1 and Qwen-Image VAEs and DINOv2 carry their own upstream licenses on top of the Apache-2.0 weights, so read the repo’s license section before shipping.
How does Moosky’s pricing compare to running VOSR2 myself?
A local rig is a fixed cost: the GPU, the electricity, and the hours the render owns your machine. Moosky inverts it to variable — credits buy at 100 per dollar and never expire, the composer quotes the exact cost before every run, and failed generations refund automatically. A quiet month on a rig still amortizes the card; a quiet month here costs nothing, and new accounts get free starter credits to test a photo first.
Disclosure
Moosky AI is an independent platform provider and reseller of AI model access, including access to SeedVR2-powered upscaling. Moosky AI is not officially affiliated with, sponsored by, or endorsed by ByteDance. This page and its examples may include AI-assisted and synthetic content.