Restore and Upscale

Make a photo bigger and cleaner without changing what it shows. Upscale and repair are separate steps with separate models: repair first, then upscale the repaired result.

Inputs to collect

• Source image: a public https URL or a TokenLab result URL.
• Goal: larger only, larger and sharper, or repair (scratches, tears, fading, noise, blur) and then larger.
• Target size: the pixel size or print size the user needs. This decides the upscale factor.
• Faithfulness: must faces, text, and fine patterns stay exactly as they are, or may the model add plausible detail?

Models

<!-- tokenlab:models:start --> <!-- Generated from the live catalog by scripts/sync_models.py. Do not edit; change preferredModels in catalog.json. -->

Operation	Model	Pick	Use
image-edit	stability-upscale-fast	default	Upscale
image-edit	topaz-labs-wonder-3-5	alternative	Upscale, alternative (not live-tested)
image-edit	bria-image-increase-resolution	alternative	Upscale, alternative (not live-tested)
image-edit	stability-upscale-conservative	alternative	Upscale with a guiding prompt (not live-tested)
image-edit	p-image-upscale	alternative	Low-cost upscale
image-edit	nano-banana-2	alternative	Repair damage, fading, or blur before the upscale
image-edit	p-image-edit	alternative	Low-cost repair

This list refreshes daily from the live catalog. Preferred models come first; otherwise models are ranked by their 24-hour success rate. Before a call, confirm the model with GET /v1/models/{model} and read its price from GET /v1/models/{model}/pricing.

<!-- tokenlab:models:end -->

How to choose, by stage:

• Repair: a Nano Banana edit model, or a P-Image edit model for a low-cost repair.
• Faithful upscale (no prompt): the Stability fast upscaler, Topaz, Bria increase-resolution, or the P-Image upscaler for low cost.
• Upscale that rebuilds detail: the Stability conservative upscaler. It takes a guiding prompt.

Observed on 2026-10-12 with a 1024 × 1024 input: the Stability fast upscaler returned 4096 × 4096, and the P-Image upscaler returned 2048 × 2048. Neither contract has a scale field. Check the output size of each new model with one test call before you promise a size. Exact fields come from each model's contract.

Workflow

Run each call with tokenlab-run (contract → estimate → call → poll → result). Commands below use its helper, ../tokenlab-run/scripts/tokenlab_run.py, a path relative to this skill's folder. If you installed this skill without tokenlab-run, send the same requests with curl. --wait stops after 100 s; a longer task keeps running, so continue with the resume command it prints.

1. Read the contract for each model. Use request_endpoint_by_operation["image-edit"] for the endpoint. Send only the listed fields. The upscale models have no prompt, except the Stability conservative upscaler.
2. Estimate the cost from GET /v1/models/{model}/pricing. Upscalers here are priced per image. Add the repair step if you run one.
3. Repair (only if needed). Send the source to the edit model with one repair instruction (see Technique). Check that faces and text did not change.
4. Upscale. Send the repaired result URL (or the source) to the upscaler. Take MODEL_ID from the Models list above for the upscale operation:
   • /v1/images/generations models (for example the P-Image upscaler): {"model": "MODEL_ID", "operation": "image-edit", "image_url": "SRC"}.
   • /v1/images/edits models (for example the Stability upscalers): {"model": "MODEL_ID", "image_url": "SRC"} (JSON image_url or multipart image).
5. Read the response. A sync response has data[].url. A response with status: "pending" is a task: poll GET /v1/tasks/{id} every 5–10 s until completed or failed.
6. Download the result and check its pixel size and detail at 100% zoom. If the size is still too small, upscale the result once more, then check again.
7. Return the result URL, the final pixel size, the model per stage, and the task IDs.

Technique

• Repair before you upscale. An upscaler makes scratches, noise, and JPEG blocks larger too.
• One repair instruction, with a keep list. "Remove the scratches, dust, and the tear on the left. Keep the faces, the clothing, and the composition exactly the same. Do not add color." Ask for color only when the user wants colorization.
• Faithful versus creative. The Stability fast upscaler and the P-Image upscaler take no prompt, so they cannot be steered toward new content. Use the Stability conservative upscaler with a short prompt that describes the photo when the input is soft and you want the model to rebuild detail. Read its contract for the creativity field before you set it.
• Do not chain many passes. Each pass can add artificial texture. Two passes is a practical upper limit. Compare against the source after each one.
• Keep the source. Give the user the original URL with the result so they can compare.

Quality bar

• The result shows the same people, objects, and text as the source. No new features, no changed faces, no rewritten letters.
• The pixel size meets the user's target.
• Edges are sharp, with no halos, plastic skin, or repeated texture.
• Repairs removed the damage and left the rest of the photo as it was.

Related skills

• tokenlab-run: the call, poll, and result mechanics.
• tokenlab-edit-image: change, add, or remove content.
• tokenlab-photoreal-stills: a new realistic photo from a brief.
• tokenlab-product-photography: product shots on new backgrounds.