High-Res Upscale · skill pública Open Skill

High-Res Upscale

High-Res Upscale é uma Open Skill pública no Digen. Abra para ver o que este fluxo cria.

@JackRhino

Product

2 0 3

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284865/2026-10-21/nb2nk4yj.jpg

# High-Res Upscale

Turns an uploaded image into a clean, high-resolution (around 2K) version of the **same** picture: same subject, same crop, same colors and text — only sharper, with fine detail and texture restored.

There is no structured project state. Inputs are the uploaded reference image(s) plus the user's text instruction; the output is a single generated image per source image.

## When to use

- The user uploads a photo, product shot, poster, scan or screenshot and asks to enlarge, upscale, sharpen, enhance, denoise, or "make it high resolution / 2K / print quality".
- The user wants the blurry, grainy or compressed look removed while the picture stays identical.
- The user asks for a before/after comparison of an upscaled image.
- Do **not** use it to redesign, restyle, re-compose, or generate a brand-new picture from a description — that is ordinary image generation, not upscaling.

## Inputs and outputs

| Input | How it is used |
|---|---|
| Reference image (required, 1 or more) | The source picture. Its content defines the subject; its pixels define identity, framing, palette and text. Each uploaded image gets its own upscale call. |
| Text instruction (optional) | Read it as *constraints on the upscale*, not as a new scene: which details matter (face, skin, fabric, label, fine print), target sharpness/size, "keep it looking analog", "clean it up completely", plus any explicit must-keep / must-avoid item. Fold those into the prompt; never let the text replace the source content. |
| Output | One upscaled image per source image, added to the canvas. |

## Steps

1. Look at the source image(s) first. If the picture is dense (faces, packaging, small print, patterned fabric), call `recognize_image` on it to list the details that must survive the upscale and the artifacts to remove (blur, noise, JPEG blocking, banding, oversharpening halos). Keep this list; it becomes the must-keep / must-avoid part of the prompt.
2. Call `read_skill_file` for `references/style-and-prompt.md` and follow its Style / Subject / Layout / Must keep / Must avoid rules. Everything visual comes from that file, not from improvisation.
3. Call `read_skill_file` for `references/workflow-stages.md` and run the stages in order (input audit → intent parse → prompt composition → generate → verify → respond).
4. Call `read_skill_file` for `references/prompt-templates.md`, pick the template that matches the subject type (person / product / text-heavy or packaging / landscape or architecture), and fill it with the concrete details from steps 1–2. The final prompt must be specific to *this* image — never ship a generic "high quality, 4K, masterpiece" string as the whole prompt.
5. Call `generate_image`:
   - `prompt`: the filled template from step 4, in English.
   - `input_urls`: the uploaded source image (s3:// uri, asset id, or preset key — never a made-up URL). One source image per call.
   - `model='i2i.hd.lite'` (fixed channel for this skill).
   - `orientation`: match the source image — `portrait` for tall, `landscape` for wide, `square` for 1:1. Do not reframe.
   - `resolution='2K'` unless the user explicitly asks for something smaller or larger.
   - `name`: short canvas title, e.g. `Upscale - portrait`, `Upscale - product label`.
   - `entity_type='upscaled_images'`.
6. Verify the result against the source before replying: same crop and subject placement, identity and expression unchanged, garment print and product/label blocks intact, text still reads correctly, no new or missing objects, no plastic skin or HDR halos. If something drifted, retry once or twice using the narrowing rules in `references/workflow-stages.md`; if it still drifts, keep the best version and say plainly what could not be restored.
7. Reply in one or two natural sentences: what was restored (detail, texture, text) and that composition/identity were preserved. Do not paste long URLs. Offer a next action with `set_guidance` (e.g. stronger cleanup, comparison poster,