Black Forest Labs launches FLUX Video Upscale, a regeneration-based 4K video upscaler

Black Forest Labs, the German AI lab behind the FLUX image and video model family, launched FLUX Video Upscale on August 20, 2026 — a standalone API endpoint that upscales video to 1080p, 2K, or native 4K by regenerating frames rather than interpolating pixels.

What's new

In its own words, Black Forest Labs explains the core difference from conventional upscaling: "Naive scaling stretches the pixels you already have, which is why it looks soft or blocky. FLUX Video Upscale regenerates the frame, reconstructing detail at the new resolution instead of interpolating between existing pixels."

The endpoint, built on FLUX 3's super-resolution capabilities, ships with two distinct modes:

• Precise mode (creativity: 0) — preserves the source exactly and sharpens it; intended for footage where identity matters, such as faces, products, brand assets, and real people
• Creative mode (creativity: 1) — restores and invents fine detail more aggressively, aimed at generated footage, textures, crowds, and scenery

Technical specifics: users can set an upscale factor between 1.5x and 3x (2x by default), submit source video from 480p and up, and receive output capped around 14.4 megapixels per frame at 24fps, with the source audio track carried through unchanged. Input is capped at 20 seconds and 50MB, submitted via HTTP(S) URL or base64 and retrieved by polling for a signed MP4. Billing is per megapixel-second of delivered output, and the endpoint lives at /v1/flux-tools/video-upscale-v1.

Context

FLUX Video Upscale extends the FLUX 3 platform Black Forest Labs launched in July 2026, which introduced a unified multimodal model spanning image, video, audio, and — via a partnership with Swiss robotics startup Mimic Robotics — physical action prediction. Since then, BFL has been shipping FLUX 3 as a family of discrete tools rather than a single monolithic release: FLUX 3 Video generation went generally available earlier in August, and the company has continued adding task-specific endpoints (upscaling, virtual try-on, content erasure) built on the same underlying backbone, following a pattern the company also used with the earlier FLUX.2 generation.

Why it matters

Video upscaling has typically been the domain of specialized restoration tools rather than generative model vendors, and BFL's regenerate-don't-interpolate approach is a direct bet that a model trained jointly across image, video, and audio can reconstruct plausible fine detail better than classical super-resolution. The two-mode design — a conservative mode for real people and branded content, a looser mode for generated footage — is a practical acknowledgment that regenerative upscaling carries fidelity risk when the input is real rather than synthetic. For studios and creators working with FLUX-generated video, having a same-vendor upscale path to 4K removes a step that previously required routing output through a separate third-party tool, and it further establishes BFL's push to be a full production pipeline rather than just a base model provider.