Image Enhance
Clean up and enlarge a selected image without leaving the canvas.
Choose the task
Open the image enhancement command for a selected image and choose a preset or the operation that matches the source problem. Presets cover Auto, CPU and AI photo work, illustration, pixel art, denoise, deblur, and combined restoration. Changing an individual processing setting switches the dialog to Custom; output behavior remains unchanged. Auto / Recommended runs a quick local analysis (noise, blur, compression, resolution) and suggests a task with a qualitative signal label — you can always override it. The label is a heuristic, not a calibrated probability.
- Denoise reduces sensor noise and grain using the task-specific SCUNet model. The dialog verifies that the graph and external weights are installed before enabling Apply; choose None for an explicit no-op.
- Deblur reduces motion and defocus blur with a validated NAFNet checkpoint. Restoration is bounded: the model runs on a cropped preview first, and full-resolution work tiles adaptively so memory stays within budget.
- Upscale enlarges with CPU, pixel-art, or the available local AI super-resolution path.
- Restore + Upscale and Deblur + Upscale compose restoration before enlargement so noise or blur is not simply made larger. The apply action names the selected operation instead of always saying “Apply recommended.”
Review before applying
The dialog previews a bounded representative crop and provides a draggable before/after comparison. The left side is the untouched source crop; the right side is the actual generated result for the selected operation and settings. This avoids comparing the output with a second resampled copy of the source.
- Inspect uses a 3×3 picker to move the crop to the center, edges, or corners of the source.
- Fit fills the available review area, while 100% shows output pixels at their native size. Use 100% for text, logos, thin lines, transparency, and edge transitions.
- The comparison slider supports dragging and the left/right arrow keys. Auto resolves its recommendation before generating a CPU preview; AI modes wait for Generate AI preview so model work is explicit.
A preview is a bounded review crop, not a full-resolution final result. Changing the operation, mode, scale, focus, or quality policy discards the old preview before the new one is generated.
The default CPU mode is Quality (Lanczos-3) so edges stay inspectable in the preview. Balanced (bicubic) is still available when a smoother photographic result is preferable.
Output behavior
New layer is the default and preserves the source. Replace source commits the derived pixels atomically and remains undoable. Use the output options only when their semantics suit the document; enhancement results are materialized pixels and do not silently become re-editable model operations. Once saved, those pixels reopen and export with the document without the model being installed or inference running again.
Models and memory
General Real-ESRGAN x4 is bundled for local AI upscaling. SCUNet denoising and NAFNet deblurring are optional, explicit downloads; split models are not considered ready until both the ONNX graph and external weights pass their recorded size and SHA-256 checks and are published together. Downloads stream into local staging storage. Interrupted or corrupt bytes do not become an installed model; a partial transfer is reusable only when the server's range and ETag still match. Missing models keep Apply disabled and offer Download model.
The output panel reports a conservative peak-memory estimate that includes decoded frames, model runtime peaks, AI's fixed 4× intermediate, and any final resize staging. It is an admission estimate rather than a promise about exact allocator usage.
Download storage and inference working memory are separate limits. A model may fit on disk but still be refused for a particular image when the shared working-set budget is full or too small. The dialog reports whether it is connecting, downloading, verifying, finishing installation, or recovering from an error. Cancellation remains in progress until the reader has stopped and any resumable partial has been handled.
Which accelerator ran?
The installed desktop app keeps presentation, GPU compute, and model inference as separate capabilities. CPU resampling remains the baseline; supported non-AI resampling can use a native GPU only after device creation and a bounded self-test succeed. Native AI sessions prefer the optional ONNX Runtime WebGPU provider in Automatic mode when the staged plugin and model are compatible, and otherwise use native CPU inference. A GPU provider is not an NPU provider.
The current release does not bundle a vendor NPU runtime, so the desktop settings panel reports NPU providers as unavailable rather than presenting a misleading enabled switch. Browser WebGPU and WebNN are independent browser capabilities and do not certify a Tauri WebView or a desktop NPU. Provider status is evidence about the completed workload, not a promise based on a device list.
Honest capabilities
JPEG artifact removal is not offered as an operation: no model has passed Varve's design-content quality corpus for that task (denoising models smear 1px lines and UI text, and the JPEG-trained candidates were rejected on provenance or shape safety). Varve does not relabel a denoiser as a compression-restoration model, and it does not substitute an unrelated fallback when a requested task cannot run. Model quality and conversion evidence are published in the benchmark report.