Video upscale

Take a low-resolution or soft video and return a sharper, higher-resolution one, up to 4K, with detail recovered and compression noise reduced. The lever is a dedicated video-to-video model that upscales every frame while holding temporal consistency, so the result stays stable across frames instead of shimmering. This is a thin capability: only three models cover it today, so routing is simple but options are limited.

Inputs to collect

• The source video. A URL or upload. (Ask only if none provided.)
• Target resolution or factor. A target like 1080p / 2K / 4K, or a 2x / 4x factor. What each model accepts differs, so confirm against the schema before promising one.
• What needs fixing: just too small/soft, or also noisy, compression-artifacted, or aliased. This nudges the model choice.
• Frame rate, if it should change (Topaz can resample fps). Otherwise keep the input's rate.

Models

• Default, highest quality: Topaz Starlight Precise 2.5 (topazlabs:starlight-precise@2.5). Diffusion-based enhancement that upscales, denoises, de-aliases, and sharpens with full temporal consistency and natural faces/skin. Best when detail recovery and a clean, photoreal result matter. Most expensive of the three.
• Enterprise 4K restoration: ByteDance Video Upscaler (bytedance:50@1). Boosts to 1080p, 2K, or 4K with denoising and motion enhancement, restores color, and cuts compression artifacts. Strong general pick for legacy films, UGC, and short clips. Cheapest per second.
• 2x / 4x with detail preservation: Bria Video Increase Resolution (bria:50@1). Upscales 2x or 4x while preserving detail and transparency. Good for product footage and design workflows that need clean output without rework.
• All three are video-to-video (op:upscale), not still-image upscalers. For images, use restore-and-upscale.
• Confirm the live model + its schema via the runware-models and runware-run skills before calling. Never hardcode a stale choice.