Black Forest Labs has released FLUX Upscale — a standalone FLUX Tool and API endpoint that reconstructs video from 480p sources up to native 4K resolution. The tool natively understands FLUX 3 output and fixes what generic upscalers miss: blurred faces and mesh-like artifacts on textures, such as water and grass. Two modes allow choosing between preserving character identity and maximizing detail, while a ten-second clip costs $6–8.

image
image

What happened

The release includes two parts: a FLUX Tool in the Black Forest Labs ecosystem and a separate API endpoint. FLUX Upscale does not simply enlarge frames; it reconstructs (regenerates) video from 480p sources. Two working modes are provided: Precise — 4 steps, $0.07 per megapixel-second, better preserves character identity and reference matching; Creative — 8 steps, $0.10 per megapixel-second, restores details more strongly but may alter or replace appearance. The upscale_factor parameter supports 1.5x, 2x, and 3x multipliers: from an HD input, you get approximately 1080p, 2K, and 4K. A ten-second clip costs $6–8. The tool is available now via the Black Forest Labs API and the playground at dashboard.bfl.ai.

Context

The problem FLUX Upscale addresses is the "last mile" of AI video. Video models, including FLUX 3, generate content in HD resolution, while delivery formats — broadcast, advertising campaigns, and the big screen — require higher-quality masters. Generic upscalers, built for natural video material, do not recognize specific artifacts of generative content: blurred faces, mesh-like patterns on water and grass, while existing high-quality super-resolution solutions work slowly. Against this backdrop, a tool tailored to its own model's output and operating on a transparent price per megapixel-second is an applied production tool, not a generic super-resolution utility.

Why this matters for the industry

For the industry, the release turns upscaling into a routine, automatable production step: a single API endpoint with predictable pricing allows embedding the "last mile" of generated video — from HD render to 4K master — directly into the pipeline without local GPUs. For studios and content teams, this closes the gap between generation quality and delivery format: upscaling stops being an experiment and becomes a standard stage. The architecture with two modes is a technology-class signal: the Creative mode's ability to change a character's appearance means the model resynthesizes the image, not just scales it. For generic upscalers, this is direct competitive pressure specifically in the AI-generated content segment, where artifacts are specific and quality is decisive.

Why this matters for users

For the reader, this is a ready-made way to get a 4K master from an HD render without your own GPUs: connect to the Black Forest Labs API or try the tool in the playground at dashboard.bfl.ai, choose a 1.5x, 2x, or 3x multiplier, and pay per megapixel-second. Budget $6–8 for a ten-second clip. If character consistency is important — for example, a recurring hero in an advertising series — choose Precise, since Creative mode may alter or replace a face. If strict identity yields to maximum detail, your option is Creative.

What is still unknown / limitations

All sources — the blog article, documentation, and product page — were published by Black Forest Labs itself; there is no independent verification of the model's capabilities. Claims about "native understanding" of FLUX 3 output and that the tool fixes what generic upscalers miss are not supported by quantitative metrics — PSNR, SSIM, LPIPS, human eval, identity preservation metrics — or by comparison with specific competitors on a single test set. The architecture is not disclosed: the price difference between modes (4 steps vs. 8) only indirectly indicates an iterative pipeline. Data on latency and throughput have not been published, so the actual throughput in production is currently unknown. The possible change in a character's appearance in Creative mode is confirmed by the vendor itself — this is a real risk for consistency tasks.

Sources

Author

Look at AI, editorial team