Take soft generator frames — or camera plates — to the size your delivery needs. Choose FAST kernels for speed and temporal certainty, or AI RESIZE when SPAN’s 2× cascade should recover sharper detail.
In the app this is a single horizontal toolbar. Here it is unfolded so every control stays readable.
Resize sits late in the pipeline: after ACEScg grade and effects, before encode. Target width and height can be typed or loaded from a built-in preset (or a delivery-profile preset). Fill & Crop fills the frame and crops overflow; Fit letterboxes inside the target. Pixel aspect ratio covers square, anamorphic, and scope plates.
The mode choice is the big fork. FAST uses classical resize kernels — deterministic, temporally coherent, no GPU required. AI RESIZE runs Transduce’s SPAN 2× super-resolution cascade for higher-ratio detail recovery, then finishes with a precise mathematical resize to the exact target. Generative / diffusion “upscalers” that boil between frames are deliberately out of scope.
Same Resize Bar, two engines. Blue columns make the difference obvious.
| Mode | Engine | When to use | Notes |
|---|---|---|---|
| FAST | Mathematical kernel | Preview speed, modest scale changes, GPU-light machines | Filter combo chooses Lanczos / Mitchell / Bicubic / Gaussian / Linear |
| AI RESIZE | SPAN 2× cascade | Larger upscales where you want sharper recovered detail | Quality Best or Medium; GPU recommended; requires SPAN model installed |
SPAN is Transduce’s 2× super-resolution module (Apache-2.0 ONNX model, ~8MB class) — not a diffusion image generator. It was chosen because motion-picture delivery needs temporal coherence: generative upscalers invent new detail each frame and “boil.” SPAN applies a trained 2× pass that recovers detail without that generative flicker.
Colour contract: SPAN is trained on display-referred sRGB. Transduce converts ACEScg → sRGB for inference, runs one or more 2× SPAN passes (cascade length depends on source-to-target ratio), then converts back to ACEScg and finishes with a mathematical resize to the exact pixel target. Quality Best / Medium trades speed vs thoroughness. Model file: span_2x.onnx, managed through the Model Manager / Model Hub path — AI RESIZE enables when the model is present.
Built-in targets on the Resize Bar and in Project Settings. Custom width × height is always available; delivery profiles can add project-specific presets.
| Preset | Width | Height | Typical use |
|---|---|---|---|
| 720p HD | 1280 | 720 | Fast review / proxy size |
| 1080p HD | 1920 | 1080 | Default finishing HD |
| 2K DCI | 2048 | 1080 | Digital cinema 2K |
| 4K UHD | 3840 | 2160 | UHD delivery |
| 4K DCI | 4096 | 2160 | Digital cinema 4K |
| Vertical HD | 1080 | 1920 | Portrait / social HD |
| Vertical 4K | 2160 | 3840 | Portrait / social 4K |
| Custom | Any | Any | Typed W×H (up to engine limits) |
Everything else on the bar — aspect handling, pixel aspect, and the FAST filter set.
| Control | Options | What it does |
|---|---|---|
| Aspect handling | Fill & Crop · Fit | Fill covers the target and crops; Fit contains the image with letterbox/pillarbox |
| Pixel aspect (PAR) | Square 1.0 · Anamorphic 1.333 · Scope 2.0 | Non-square pixel delivery for anamorphic / scope pipelines |
| FAST — Lanczos | Kernel | High-quality sinc-windowed filter — strong default for most ups/downs |
| FAST — Mitchell | Kernel | Balanced cubic family — smooth without extreme ringing |
| FAST — Bicubic | Kernel | Classic cubic interpolation — familiar, general-purpose |
| FAST — Gaussian | Kernel | Soft anti-aliased path — useful when downscaling needs gentler pre-filter |
| FAST — Linear | Kernel | Bilinear — fastest, softest; preview or when sharpness is not the priority |
| AI RESIZE — Quality | Best · Medium | Shown when AI RESIZE is selected — trades SPAN cascade thoroughness vs speed |
Transduce bridges the gap between AI and cinema. Your vision. Your pipeline. Perfectly matched.