Restore and upscale

Take a degraded or low-resolution image and return a cleaner, sharper, larger one: deblur, denoise, dehaze, recover detail, and enlarge 2x to 4x. The lever is matching the kind of damage to the right model, not running everything through one upscaler. For video, a dedicated temporal model exists. For worked end-to-end recipes (straight upscale, old-photo restore, video upscale), see references/examples.md.

Inputs to collect

• The source image (or video). A URL or upload. (Ask only if none provided.)
• What's wrong with it: soft/blurry, noisy/grainy, hazy, compressed/JPEG artifacts, just too small, or an old/damaged photo. This routes the model.
• Target size or factor: 2x or 4x, or a target resolution. The supported factors vary by model, so confirm against the schema before promising one.
• Fidelity vs invention: "stay faithful to the original" (transformer/GAN upscalers) or "add believable detail" (diffusion upscalers). They behave differently and this is the most important routing question.
• Content type: photo, document/text, artwork, or screenshot. Documents and text demand a faithful (non-diffusion) path so characters don't get reinvented.

Models