Enlarge one image by 1x, 2x, 4x, or 8x with an instant browser resize, free Swin2SR super-resolution on this device, or paid Topaz/Recraft processing.
Where: Flowgen -> Add -> Image & Vector -> Edit -> Image Upscale
Before you start
- One image connected to image_in.
- An open supported browser for Plain resize and On this device.
- Enough memory and patience for on-device CPU inference.
- Sparks for AI upscale.
Key ideas
- Three methods with different promises: Plain resize is free and instant but only interpolates existing pixels. On this device is free Swin2SR super-resolution that creates detail but downloads a model and freezes the tab while CPU inference runs. AI upscale uses a provider, costs Sparks, and is the only method that can finish without keeping this browser tab open.
- Factor is not exact dimensions: Image Upscale targets a multiplication factor. Use Image Resize when downstream work requires exact width and height.
- Method and model affect valid controls: Plain and Topaz allow 1x, 2x, 4x, and 8x. Local omits 1x because it would run a model only to downsample again. Recraft Crisp accepts no factor at all, so the factor control disappears.
- Input changes invalidate the result: The prior result is hidden when the source image changes. Run again so downstream nodes cannot mistake an upscale of the old source for the current image.
Steps
- Connect one image.
- Choose Plain resize, On this device, or AI upscale based on whether speed, local detail, or provider quality matters.
- Choose a factor where the selected method/model supports one.
- For On this device, choose the Swin2SR model and read the first-download and actual pass/interpolation note.
- For AI upscale, choose Topaz Image Upscale or Recraft Crisp; review the quoted Sparks and Topaz NSFW check setting where shown.
- Run the node, compare the result at 100 percent, then download or connect Image Output onward.
Sockets and execution
- Image / image_in: exactly one source.
- Image Output / image_out: the enlarged result.
- No prompt or text socket.
- Explicit Run is required.
- Plain and local methods run in the tab; AI runs through the provider path.
Plain resize
- Free and instant.
- Browser canvas interpolation.
- 1x, 2x, 4x, or 8x.
- Adds pixels but does not invent detail.
- Good for satisfying an exact scale requirement quickly; use Image Resize for exact dimensions.
On this device: Swin2SR
- Free local super-resolution.
- Swin2SR 2x (classical): best all-round clean-source default, about 19.0 MB CPU weights.
- Swin2SR 2x (lightweight): smallest/fastest download, about 5.6 MB CPU weights.
- Swin2SR 4x (real-world): soft/noisy photos, about 19.0 MB.
- Swin2SR 4x (JPEG artefacts): blocking/compressed sources, about 19.2 MB.
- Runs on CPU and freezes this tab during inference.
- Some requested factors use one or two native model passes followed by plain interpolation; the card reports that honestly.
AI upscale and current Sparks
- Topaz Image Upscale: 1x or 2x = 7 Sparks, 4x = 14, 8x = 28.
- Topaz can run its NSFW checker; the default is on.
- Recraft Crisp: 1 Spark, no factor input.
- Provider availability and the live quote remain authoritative.
Tips
- Use Plain resize for layout/proxy work, local Swin2SR for free detail, and paid AI only when the quality or background execution justifies Sparks.
- Inspect edges, faces, text, and compression blocks at 100 percent.
- Pick the local model whose native scale and source training match the input; avoid repeated passes unless the larger factor is necessary.
- Do not read interpolated final pixels as model-created detail when the run note says the last step was a plain resample.
Limitations and important notes
- Local inference runs on the CPU and can freeze the tab.
- First-use model weights must download and cache in the browser.
- Browser memory can limit very large sources or repeated 4x passes.
- Recraft Crisp does not expose a factor.
- Upscaling cannot reliably reconstruct truly missing text, faces, or fine structures.
Troubleshooting
On this device is unavailable
Use a browser with WebAssembly and createImageBitmap support, or choose Plain resize/AI. Headless runs cannot use the browser-only methods.
The tab appears frozen
The local method intentionally runs CPU inference on the main thread. Wait for completion or use Plain resize/AI on the next run.
The factor control disappeared
Recraft Crisp takes only the image; choose Topaz, Plain, or Local when a specific factor is required.
The output looks only larger, not sharper
Check the method. Plain resize only interpolates. Choose On this device or AI for super-resolution/detail synthesis.
FAQ
Is local upscale free?
Yes. It costs no Sparks, but it downloads model weights and uses this device’s CPU.
Why is local 1x missing?
The fixed-scale model would still enlarge and then downsample, which is slow and does not match the 1x promise.
Does 8x mean every pixel came from AI?
Not always. The local run note states whether native model passes were followed by plain interpolation.