Black Forest Labs' FLUX Video Upscale is now live on fal and Replicate, offering 1080p, 2K and 4K output with two distinct processing modes

Black Forest Labs has quietly shipped something that every AI filmmaker working at sub-HD resolutions will want to test immediately. FLUX Video Upscale is now live on both fal and Replicate, turning lower-resolution AI video output into 1080p, 2K or 4K deliverables through an API call. The launch is clean and functional: two modes, optional prompt steering, and source audio preserved throughout. No waiting list, no private beta, no vague roadmap date.

The detail that separates this from a standard upscaler is the modal split. Precise mode is built to hold faces and product surfaces exact, which matters enormously when a brand brief requires consistent talent likeness or a product shot where any reconstruction artefact reads as a flaw. Creative mode takes the opposite approach, actively rebuilding fine detail rather than preserving it, with an optional text prompt to steer what gets added. That prompt-guided detail pass is the genuinely new thing here: it pushes the tool closer to an AI-assisted restoration than a traditional super-resolution pipeline.

What this changes in a real workflow

Most AI video generation right now tops out at 720p or 1080p natively, and the path to broadcast or large-screen delivery has involved third-party upscalers, frame interpolators, or manual post passes in Topaz. FLUX Video Upscale slots into that gap as an API-native step. Because it sits on fal and Replicate, it can be chained directly after a generation call in ComfyUI, a custom Python pipeline, or any tool that already integrates either platform. The source audio carryover means you are not losing a sound mix you assembled before the upscale step.

For practitioners using Seedance 2.5, Wan 3.0 or any model that outputs at 720p, this is a credible route to 4K delivery without touching the generation stack. The Precise mode is the obvious first test for anything character-driven. The Creative mode with a prompt deserves a separate test pass, because the results will depend heavily on how the diffusion model interprets the steering text against the source footage, and that behaviour is not yet documented in the wild.

What has not been said

Pricing is not stated in the launch posts. Both fal and Replicate operate on pay-per-second or pay-per-run models, so costs will vary by resolution tier and clip length. There is no published benchmark comparing output quality against Topaz Video AI or the LTX-2.5 LoRA upscale approach that has been circulating independently. The Creative mode's prompt interaction has no documented examples from BFL at this stage, so the degree of control it actually offers is still an open question. No maximum clip length or file size limits are mentioned.

What to test first

Start with Precise mode on a 720p character clip that has consistent facial identity across cuts. Check whether the upscale holds the likeness across motion blur frames, which is typically where super-resolution models introduce drift. Then run the same clip through Creative mode with a neutral prompt and compare the texture on non-face surfaces. If the Creative mode adds plausible grain or surface detail without hallucinating geometry, it becomes a useful tool for footage that needs atmospheric texture, not just resolution. The prompt steering is worth experimenting with: try a short descriptor of the surface material (rough stone, worn leather, soft linen) and observe whether it biases the reconstruction sensibly.

Both the fal and Replicate endpoints are live now. No date or pricing announcement has been made by Black Forest Labs directly.

---

Sources: fal announcement · fal try-it link · Replicate announcement