Denoise
Reduce sensor noise and grain with SCUNet after its verified model and weights are installed. The dialog checks readiness first and keeps the source available for comparison.
Denoise, upscale, or combine both in one focused workflow. Varve keeps restoration local, bounded, and explicit about what each operation can do.
One workflow
Choose a reviewed starting point or the task you actually need. Presets cover Auto, CPU and AI photo work, illustration, pixel art, denoise, deblur, and combined restoration; changing an individual processing setting switches to Custom without changing your output behavior. Denoise uses a task-specific local model; Deblur reduces motion and defocus blur with a validated model; Upscale uses the existing CPU or bundled/downloaded AI path; Restore + Upscale runs only those selected stages, in order.

Available today
Reduce sensor noise and grain with SCUNet after its verified model and weights are installed. The dialog checks readiness first and keeps the source available for comparison.
Reduce motion and defocus blur with a validated local model — preview a bounded crop before committing to full-resolution work.
Choose edge-retentive Quality CPU resampling by default, smoother Balanced scaling, pixel-art algorithms, or bundled Real-ESRGAN x4 for supported image work. AI runs locally; if its selected provider cannot complete, the workflow reports the CPU fallback instead of hiding it.
Compose only the stages you selected. Restoration runs before enlargement so noise is not simply made larger.
Start with a task-specific recipe, then change mode, scale, quality, denoise, or deblur strength. The action label and visible controls update with the selection.
A quick local analysis suggests what your image actually needs, in plain language, with a qualitative signal label you can override. It is a heuristic, not a probability.
For illustration and linework
AI enhancement is an estimate, not a lossless resize. Compare the result with the untouched source at 100%, especially thin lines, silhouettes, palette shifts, grain, and transparent edges. Classical CPU and pixel-art scaling remain available when predictable geometry or exact source colours matter. A model's verified file hash establishes which bytes ran; it does not by itself establish visual quality, conversion provenance, licensing, or memory suitability for every image. Keep the source until the result has been reviewed.
Optional models, on your terms
Model-based tasks start only when you choose them. The download dialog names the model source, transfer size, and a separate estimate for peak working memory.
Model bytes are written to local staging storage and checked against their expected size and digest before the model becomes available. A split model stays unavailable until its graph and companion weights are both verified.
Apply the generated pixels as a new layer, then save the document. Reopening that document does not need the model or another inference run.
Download space and inference memory are different costs. A verified model can still be too large for a particular image or device to run; core editing remains available when an optional model cannot be installed or admitted.
Read the model and memory guide· Lower-memory and Chromebook guidance
Source pixels stay explicit
Object → Resize Image changes the embedded raster dimensions while keeping the placed image's bounds and crop contract intact. The dialog keeps encoded sRGB as the compatible default and offers an explicit linear-light path for translucent or antialiased photo edges. It does not secretly change the page, object scale, or the saved source when you cancel.
Desktop acceleration
On the installed desktop app, conventional resampling may use a native Vulkan, Metal, or D3D12 GPU path after a real device self-test. AI upscaling and other local models use the bundled native runtime and a verified WebGPU provider when that optional component, driver, model, and graph all pass their checks. The result can still use CPU work when that is faster for a small or cold request.
The browser demo has its own WebGPU/WASM capability path; it cannot borrow the desktop app's device or runtime. Varve does not infer NPU use from a product specification, and the current release does not ship a vendor NPU runtime. Every unsupported or lost accelerator falls back to a correct CPU workflow without changing the source layer.
See the capability stages and fallback guide →Review first
The preview gives the result room to breathe instead of shrinking it into a tiny thumbnail. Compare the untouched source crop on the left with the actual generated result on the right, then apply it as a new layer when it looks right.
A 3×3 focus picker moves the review crop to corners, edges, or the center where artifacts often hide.
Fit view uses the available preview area; 100% view shows the output at pixel size for text, linework, and transparency.
CPU previews are generated from the selected settings. AI previews are explicit and start only when you ask for them.