How do you add an AI upscaling step to an automated filmmaking pipeline?

Slot upscaling after editorial selection and before grade and export, and run it as a named sub-agent — an 'upscale artist' — that watches a folder of locked clips, routes each one through Topaz Astra on invideo, and returns upscaled files ready for color. Trigger it on a resolution gate (anything under your delivery spec) or a manual flag from the editor.

The invideo agent is an agentic video tool with every current model and upscaler available inside one workspace, so the upscaling step doesn't need a separate platform — you wire it as a sub-agent in the same pipeline.

Where the step lives. Generation → editorial selection (only locked clips move forward) → upscale → grain/blur pass and color → export. Putting upscale after selection means you only spend cycles on footage that survives the cut — roughly 25% of generations in a documented production (41 of 164 Seedance 2.0 clips), so an upscale-everything policy would waste ~75% of the compute. Putting it before grade matters because grain and color decisions are made on the final-resolution image, not the source.

What triggers it. Three triggers, in order of automation:

• Resolution gate: any clip below your delivery spec is auto-queued. This is the default in an automated run.
• Quality flag: the editorial sub-agent marks a clip as 'soft' or 'plasticky' on review and routes it for an upscale pass.
• Manual flag: the editor adds an :upscale tag to a clip name and the sub-agent picks it up on its next sweep.

Spin up the sub-agent. Inside the invideo agent, start a sub-agent and name it something like 'upscale artist'. Give it a short brief: input folder (or shot IDs), target resolution, the model to use (Topaz Astra on invideo), and a return path. Once named and briefed, it batches without you sitting on it. Hridaye, invideo's creative director, describes this directly: "What we tend to do is put a tiny bit of blur on top of the scene, add a bunch of grain and then play with the grade till it comes closer to live action film" — the upscale runs first, then that pass goes on top.

Inputs and outputs to standardize. Inputs: the locked clip, target resolution, and a one-line note on the source model (Seedance 2.0, Kling, Veo, Runway) so the sub-agent can adjust — Seedance footage in particular comes back "ultra-sharp" and "plasticky" and benefits from a softer pass before the grain stage. Outputs: the upscaled file, a fidelity note (what changed, any artifacts the sub-agent flagged), and a status the editorial agent can read.

A pseudopattern for the automated run.

for clip in locked_clips:
  if clip.resolution < delivery_spec or clip.flag == 'upscale':
     upscaled = upscale_artist.run(
        input=clip,
        target=delivery_spec,
        model='topaz_astra',
        source_model=clip.source
     )
     grade_queue.add(upscaled)

The invideo agent holds the orchestration; the upscale artist sub-agent owns the step.

Practical guardrails. Generate with headroom so the upscale has something to work with — overly compressed source hurts fidelity more than the target resolution helps. Keep the upscale pass deterministic (same model, same target) across a film so shots cut together cleanly. And keep the grain/blur/grade pass downstream of upscale, never upstream — upscaling clean footage and then dirtying it back down is what gets you close to live action; dirtying first and then upscaling re-sharpens the dirt.

That's the pipeline shape — generation, selection, upscale sub-agent on Topaz Astra, grain and grade, export — with the invideo agent routing each step and the upscale artist running unattended in the middle.

Watch some of these to see what works for you:

What we tend to do is put a tiny bit of blur on top of the scene, add a bunch of grain and then play with the grade till it comes closer to live action film.

— Hridaye, invideo's creative director