Low-Compute Ultra HD Upscale · Open Skill 公開技能
Low-Compute Ultra HD Upscale
Low-Compute Ultra HD Upscale 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。
Manhua
0 1 1
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/22y4scxp.jpg
# Low-Compute Ultra HD Upscale Turns one user-uploaded reference image into a faithful, sharp, high-definition 2K image. This is an image-to-image upscaling job, not a new picture: the source subject, structure, colors and framing must survive unchanged while blur, noise and compression artifacts are removed and micro-detail is reconstructed. There is no structured project state. One run produces exactly one image, which lands on the canvas. ## When to use - The user attaches or points at an image and asks to upscale / enlarge / make it HD / 2K / clearer / sharper. - The user wants to re-run a previous upscale or try another output size. - The user wants small fixes on an image that was already upscaled by this skill. Not for: generating a brand-new image from a text idea with no reference image, or editing the image's content (adding/removing subjects, restyling into a different art style). ## Input and output - **Input — reference image:** exactly one image. It arrives as a conversation attachment (asset id / `s3://` uri), a real https link, or a preset asset key. Use the reference the user actually attached. - If no image is attached: ask the user to upload one, and do not call the generation tool. - If several images are attached: ask which one to upscale (or, if the user clearly wants all of them, upscale them one at a time, one call each). - **Output — image:** one upscaled image, default 2K, shown on the canvas. Every input image is used the same way: it goes into `input_urls` as the single reference for the image-to-image call. Never re-upload it, never invent a url, never substitute a stock image. ## Steps 1. Pick up the reference image the user attached. Read back what it is in one short clause (e.g. "an old photo of a street at dusk") so the user can confirm you are looking at the right file. 2. Build the English prompt from the template below. Fill the subject slot from what is actually visible in the reference (and from anything the user described), and keep everything else verbatim. If the user explicitly asked for one extra change on top of upscaling, append it as a single explicit clause; add nothing else. 3. Call `generate_image` once, with the parameters in the next section. 4. Reply with one or two natural sentences: what was upscaled, to which size, and that the result is on the canvas. Do not paste urls. ## Prompt template (fill in, keep in English) ``` Faithful 2K ultra-high-definition upscale of the provided reference image. Subject / content: keep exactly the subject of the reference image — <one short English clause naming the main subject, e.g. "a woman in a red coat standing on a rainy street"> — with no added, removed, moved or replaced elements. Style: keep the original color palette, contrast and white balance, the original materials and textures (skin, hair, fabric, metal, glass, wood, foliage, water, sky) and the original light direction and light quality exactly as in the source; no restyling, no color grading shift, no conversion to illustration / anime / painting, no added filters, no glow or bloom. Layout: keep the source composition, framing, camera angle, perspective, subject position, background and aspect ratio; no crop, no reframing, no stretching, no rotation, no added borders, captions, logos or watermarks. Detail reconstruction: recover crisp micro-detail and clean edges — natural skin texture, individual hair strands, fabric weave and stitching, leaf and grass definition, architectural lines, crisp legible edges on any text or signage without changing its wording; remove blur, motion blur, noise, grain, banding, jpeg artifacts and jagged edges. Output: a single clean, sharp, artifact-free 2K image, natural-looking and true to the source. ``` ## Parameters (call convention) - `input_urls=[<the reference image>]` — required; this is what makes it image-to-image. - `model='i2i.hd'` — image-to-image channel with the higher-quality binding, matching the "HD upscale" int