VideoGigaGAN
Adobe's generative video super-resolution: GigaGAN's detail-rich image upsampling inflated to video, with the flicker engineered out.
About
VideoGigaGAN (arXiv 2404.12388, CVPR 2025) is Adobe Research's generative video super-resolution model, built with the University of Maryland. The idea is straightforward and the execution is not: take GigaGAN, a large-scale generative image upsampler, and inflate it into a video model. Doing that naively produces severe temporal flicker, so the team added temporal attention in the decoder, a flow-guided propagation module for consistency, anti-aliasing blocks against AI artifacts, and high-frequency skip connections to compensate for detail lost in downsampling.
The result is up to 8x upscaling (the project page demos 128x128 to 1024x1024) with fine-grained detail — eyebrow hairs, skin pores, eyelashes — that conventional VSR blurs away, and the authors report beating prior VSR models on public datasets (their numbers). The catch, and it matters: as of the research announcement Adobe had released no code, no weights, and no product — independent coverage at the time described it as a demonstration, not a pending release. A community unofficial implementation exists but deviates from the paper in several components.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (CVPR 2025): outpaces prior VSR models including BasicVSR++ on public datasets; 8x upscaling showcased on project page
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: paper arXiv 2404.12388 (Xu et al., Adobe Research/UMD), CVPR 2025; project page videogigagan.github.io; independent coverage (TechXplore, April 2024) notes it was announced as a demo, not a pending release. Strengths: top-tier lab, impressive 8x generative detail, well-documented method. Docked for: (1) no official code or weights released — the claims cannot be independently tested; (2) all numbers are author-reported; (3) no product, API, or timeline for availability. License: proprietary research; no public weights.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 2.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · arXiv 2404.12388, CVPR 2025 (Xu et al., Adobe Research + UMD)
- · Generative VSR from GigaGAN + temporal attention, flow-guided propagation, anti-aliasing; up to 8x demoed
- · Announced as a research demo, not a pending release (independent coverage, April 2024)
- · No official code/weights; only unofficial community reimplementation; all claims author-reported