Faithful vs generative upscale: which to pick

Choose a faithful upscaler when the image must stay true to the source, and a generative one when you accept invented detail for a sharper result. fuser.studio frames it as a single question: may the upscaler invent detail? Sume exposes image_upscale_create and video_upscale_create as hosted MCP tools.

The ranges reported

fuser.studio's guide lists the following scale ranges for the two engines it names.

Faithful versus generative

The same guide compares Topaz Wonder 3 with ESRGAN on a ceramic figurine: ESRGAN kept the shapes but smoothed texture, while Wonder 3 rebuilt a speckle pattern that was not recoverable from the source. Faithful models tend to sharpen what is there; generative ones fill in plausible texture, which can alter faces, text and logos. For product shots, documents and brand marks, invented detail is a defect. For old photos and textures, it may be the goal.

• Faithful: product packshots, text, logos, evidence.
• Generative: heavily compressed or tiny sources where some invention is acceptable.
• Always compare against the original at 100 percent zoom.

Testing through Sume

Hosted MCP lists image_upscale_create and video_upscale_create among its paid tools, each needing an idempotency_key. Call tools_schema for each name to see the live options, such as scale factors, instead of assuming them. Sume's docs that I read do not state which upscale engine sits behind these tools or whether it is faithful or generative, so judge by output.

Run your hardest input, a small compressed JPEG with text, and then a clean photo. Use dry_run=true for a cost preview.

Keep both versions and the settings. Decide by looking at faces, edges and any text rather than by sharpness alone.