Topaz upscale

Give it a clip and a target width; get back the same clip at that width, with the same frames and the same faces. Topaz's Proteus model enhances what the render already has rather than generating new detail, so it is safe on people. It is a cost saver, not a quality route: Seedance 2.5 renders at 720p in about 6 minutes against 6–30 minutes at 1080p, so 720p plus a 1.5× upscale is cheaper, but it softens fine detail. Use it only on shots where quality matters less; when the whole frame ships at 1080p quality, render at 1080p natively, and when only part of the frame is used, render that part's size and don't upscale (guide-prepare-renders-for-delivery). It won't add what the render never had: a soft or smeared take stays one.

Inputs

The untrimmed render, straight from the model. Upscale before trimming (ffmpeg-frame-exact-trim), never after.
Its width (ffprobe-inspect-media). Skip any clip already at the target width.
The target width: 1920 for a 16:9 delivery.

Run it

Call describe("fal/fal-ai/topaz/upscale/video") before the first run in a job and check the fields below against it; they are the ones proven on real jobs.

run(
  model="fal/fal-ai/topaz/upscale/video",
  input={
    "video_url": "<job>/video/raw/shot09_seedancei2v_j.mp4",
    "upscale_factor": 1.5,
    "model": "Proteus",
    "H264_output": true
  },
  into="<job>/video/upscaled",
  unit="shot09_seedancei2v_j"
)

video_url takes the local path; the gateway uploads it.
upscale_factor is target width / source width, rounded to 4 decimals (1920 / 1280 = 1.5).
model: Proteus. Not SeedVR, which generates detail and changes faces.
H264_output: true returns H.264, which every later ffmpeg step reads.

The output's video is downloaded into into.

What you get

The clip at the target size. Put it where the take's name points (shot09_seedancei2v_j.mp4) and keep the original beside it as raw/shot09_seedancei2v_j_source.mp4, so a later round can go back to it. Record the upscale (model, factor) in the take's notes.

Tips and traps

Order: render, upscale, trim, grain. Trimming re-encodes; upscale the model's own output, then trim the upscaled file, then grain at delivery size if a texture note needs it.
Rendering natively at 1080p skips this step, at the cost of the wait: on the first job, a wide shot's first round was 720p plus Topaz, came back soft with its logo illegible, and its second round rendered at 1080p directly.

Check it

Size: exec ffprobe -v error -select_streams v:0 -show_entries stream=width,height,r_frame_rate -of csv=p=0 <out> gives the target size and the source's frame rate.
Frames: the frame count equals the source's (exec ffmpeg -v error -i <out> -map 0:v -f framemd5 -, counting lines that don't start with #).
Faces: crop a face at 100% from the same frame of the source and the upscale, and look. Compare python-measure-detail on that face between the two.

Hypernatural