Diffusion models have a native resolution where they’re smartest. Fight it and you get mutants; respect it and finish with an upscaler, and you get clean 4K from a 1-megapixel render. This machine is about the finishing line.

The basic pattern

Generate near the model’s native size → Upscale Image (using Model) node with a dedicated upscaler → optionally a light second diffusion pass at low denoise to re-sharpen details the upscaler smoothed. Upscalers are small, fast models — stack one at the end of any workflow.

The three traps

  1. The waxy-skin trap. Aggressive upscaling + smoothing = plastic people. Counter-punch with grain: a film-grain or detail-enhancer pass after upscale reads as photography instead of render. Our footage ships with a grain pass for exactly this reason.
  2. The zoom-then-upscale trap. Cropping deep into a small source and then upscaling blows a few hundred pixels into a soft mess. We once shipped weeks of video with a 1.8× zoom into a small render — faces were 300 pixels pretending to be 1080. Fix the framing at generation time; zoom gently (≤1.25×).
  3. The face-size rule (video). Any face that will be lip-synced or scrutinized needs to be large in the source frame — post upscaling cannot invent a face that was never rendered. When a wide shot fails QC on face detail, re-render the plate at higher resolution or tighter framing; don’t upscale harder.

When to upscale what

  • Stills for thumbnails/cards: always — render 1MP, deliver 4K.
  • Video: upscale the assembled result once at the end (portrait 1080×1920 or 1920×1080), not each clip repeatedly. Every extra generation-decode cycle costs texture.
  • Faces: prefer a face-aware detail pass over global sharpening; global sharpening finds every artifact you hoped nobody would see.

QC like you mean it

Automated pipelines need a gate: sample frames, read them with a vision model, fail on melted features/smeared mouths/mangled hands (we gave our AI eyes partly for this). And calibrate the gate to your finish — a rubric tuned for one engine’s failure modes will false-fail another’s intentional look.

Engine Room series: From Zero · Edit & combine · Video · 3D · Character pipeline