🎬 Flux 3 Video Upscale (API) *DRE

The sibling to Flux3VideoDRE that answers "okay, but I want it sharper." It POSTs your clip to BFL's video upscale endpoint (/v1/flux-tools/video-upscale-v1), waits, and returns a bigger VIDEO plus metadata. No local upscaler, no VRAM hit, no model files - your GPU isn't touched for the heavy lifting. Everything about it is a hosted API call, and that's both the point and the limitation.

How it picks a target size

You pick target_resolution from 1080p, 2K, or 4K, and the node does the math: it measures the source's short side and derives upscale_factor = target_short / min(w, h), clamped between 1.5x and 3.0x, keeping your aspect ratio. So a 720p clip going to "4K" isn't a literal 4x blowup - it's whatever factor lands the short side at roughly 2160, capped at 3x. That's a sensible guard against asking a generative upscaler to invent more than it can.

Then there's the creativity switch, and this is the one to think about before you hit run:

• precise - preserves identity and stays sharp. Use it for faces, products, real people.
• creative (default) - invents detail. Great for generated footage, landscapes, textures; faces and products can drift.

This maps directly onto the upscaling knowledge base's core distinction: adding pixels vs. inventing detail. "Creative" is generative refinement and it will rewrite things. If your shot has a person's face in it and you pick creative, expect the face to come back different.

Inputs that matter

• video - the clip to upscale, mp4, max 20s and 50MB. Or input_video_url instead.
• input_video_url - an HTTP(S) URL to the source, and the clever bit: it takes precedence over the video input, and the node only streams the moov atom of the URL to measure dimensions - no full download, no giant base64 blob. If your source is already hosted somewhere, this is the cheaper path.
• prompt - optional description of the clip that steers the enhanced detail, most useful in creative mode.
• safety_tolerance - 0 (strictest) to 4, default 2. Moderates the prompt and delivered frames.
• timeout_minutes - default 45. An upscale takes about as long as generating a video, and the same timeout warning applies: when it fires, only the node stops; the job keeps running server-side and still costs credits.

Outputs: video (VIDEO) and metadata (STRING).

Setup and the usual API caveats

Same as its sibling: BFL_API_KEY=bfl_... in .env (or the api_key input), pack installed via ComfyUI Manager or git clone https://github.com/DenRakEiw/DenRakEiw_Nodes + pip install -r requirements.txt + restart.

The costs and caveats are the cloud-video standard: metered, slow, your footage leaves the machine. One practical note from the upscaling playbook that applies regardless: a generative upscale of generated footage is usually safe, but don't point "creative" at archival or real-world video you need to stay faithful - that's what precise is for. And if all you need is more pixels on something already sharp, a free local ESRGAN-class upscaler is the cheaper answer; this node earns its keep when you want BFL's model to also add detail while it climbs.