AGENT SKILL

studio-image

Create and refine still images with Open Imagine's Image API. Direct composition, camera angle, framing, lens and lighting effects across photography, historical art, fantasy, anime, and illustration. Also covers images that carry text or layout (posters, covers, logos, app screens, infographics, packaging, signs, memes), characters and game assets, and reference edits such as restoration, composites and try-on. Use for shoots, coherent image sets, visual exploration, and reference-based edits with visual review.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 0
Forks
1

01 前言 Front Matter

name
studio-image
description
Create and refine still images with Open Imagine's Image API. Direct composition, camera angle, framing, lens and lighting effects across photography, historical art, fantasy, anime, and illustration. Also covers images that carry text or layout (posters, covers, logos, app screens, infographics, packaging, signs, memes), characters and game assets, and reference edits such as restoration, composites and try-on. Use for shoots, coherent image sets, visual exploration, and reference-based edits with visual review.

02 资源文件 15

03 SKILL.md 正文

Studio Image

Turn a brief into still images that work together, then inspect and refine the actual images. Operate the user's Open Imagine server through its public API. Keep this workflow usable with different Image models; discover their capabilities rather than assuming a particular checkpoint. Requires a reachable server and tools for HTTP, saving images, and visual inspection.

Choose the path

If the request is short or vague ("a nice photo of my dog"), or the user says they do not know where to start, follow the brief-to-final flow. It asks at most three plain questions, shows four concepts from the style library, and makes the final from the concept the user picks. A user with a full brief (subject, use, style) continues below.

Establish the shoot

Use the user's server URL; otherwise try http://127.0.0.1:8765. Read the API operating notes before submitting work. If the server is unavailable, report the connection problem; starting or installing a server is a separate task unless already authorized. Once the backend is known, read model notes for measured behavior on that checkpoint. An unlisted model is untested; use its live capabilities and its own prompt wording.

Extract the subject, intended use, deliverable count, aspect ratios, and visual direction from the brief. Identify what must remain exact: a face, product silhouette, label, material, garment, or supplied composition. Use sensible defaults for missing creative details; ask only when an unknown changes the requested result materially. Respect the user's generation/time budget. A request for one finished image does not imply permission for an extensive concept round (the brief-to-final flow's four small concepts are its normal start, not an extra batch).

Give each reference a role: subject identity, product geometry, wardrobe, lighting, pose, or layout. Refer to images in their submitted order using the selected model's supported wording. Distinguish a reference's required content from incidental background details. If a supplied product or person must remain recognizable, prefer reference-based Edit over reinventing it from text. Do not claim exact brand or identity fidelity without inspecting the result.

Before you write the prompt, check the request against boundaries if it involves a real person, a document, someone else's work, a brand or a mark, health, or news. Check it before Rewrite, because Rewrite carries the request into a longer prompt.

Ask for the end use (screen, print, sticker, platform, vector, cutout) when the brief does not imply it, and otherwise state the end use you assume, because it sets the background, the margins and the size; delivery has the recipes. Split the job: the model makes the picture, and a layout tool or script adds what must be exact (long copy, exact colors, QR codes, grids, counts, cut lines, the final pixel size, transparency).

Place each request on one accuracy level, and tell the user which:

  • Mood only. Mood, style and composition. The model is fine.
  • Roughly right. History, botany, products, architecture: it looks right and needs a human check before anyone relies on it.
  • Must be exact. Data charts, maps with real geography, anatomy for teaching, floor plans, math and chemistry figures, sewing patterns. Draw these with code or CAD, not this model.

Direct and generate

Find the brief's kind of image in use cases. It names the file to read, a default ratio, and whether Rewrite helped on the tested model. Before you write a custom prompt or use Rewrite, read prompt craft. For an image that carries words or a layout, read design and text. For stickers, icons, characters, comics, sprites or other drawn assets, read illustration and assets. For any Edit, read editing.

For composition, viewpoint, framing, focus, camera effects, lighting, or subject staging, read shot design. Treat these as independent decisions from style. Choose the visible outcome first, then use compatible camera shorthand to support it. Camera numbers in prompt text are creative cues, not calibrated API controls or proof of an actual capture setup. Keep them separate from discovered engine settings.

For a real-camera brief, direct a credible setting, material condition, motivated light and degree of processing. Inspect skin, hands, object joins, reflections and fine texture at full size. Photographic appearance and obeying the requested shot are separate checks; a convincing portrait can still miss its crop or lighting pattern.

For photography beyond portraits, a named camera brand, or an in-camera processing look, read photography genres and camera cues. Use the genre's actual subject and capture constraints to direct the image; a camera name alone does not establish its optics, sensor response, or proprietary color processing.

To pick a tested style (an art style, an animation school, a film stock, a lens, a director's way of shooting, a light), use the style library: each row gives the prompt sentence that worked and what to watch for. For style exploration or a nonphotographic brief, read style directions. Choose a specific visual language and describe its marks, space, palette, and material. Do not append photographic realism to painting, anime, or abstraction unless the user wants a hybrid. Separate a movement, a medium, a genre, and a production technique rather than treating their names as synonyms.

Choose an offered template when it fits; fill its actual fields with the brief. Choose Looks deliberately rather than filling every row. Avoid contradictory camera or lighting instructions. For a custom prompt, describe subject and action, framing, setting, light direction and quality, surface detail, and placement of any empty space or exact lettering. State preservation and change instructions explicitly for edits. Use the selected model's official guidance linked from the live API documentation when its prompting behavior matters.

When a prompt has two or more people, in any style, describe each person, not the group: an age, a face shape, hair, a build and one mark, placed "from left to right". Otherwise the faces come back as near-clones. A sentence asking for variety does not help. For a family, describe each parent or grandparent the same way, then give each child the traits they take from a named relative ("a boy with his father's square jaw and wavy black hair"). Do not write "they look related": it dropped a person. See groups of people.

For a coherent set, hold the subject or location, palette, light logic, and treatment steady while varying the requested shot: hero, detail, contextual scene, or alternate crop. A repeated seed does not guarantee the same identity. Once there is a good hero, use it as an Edit reference for dependent variations. Generate those only after reviewing that hero. For architecture, preserve the defining materials, openings and layout; a reference image does not establish the geometry hidden behind its visible surfaces.

Tell the user before you start when the brief needs something the tested model did not do. On Qwen-Image-2.1-Turbo (on another backend, treat this list and every Findings table as hypotheses to check with a concept):

  • No transparent background. Generate on pure white and cut it out afterwards.
  • No guaranteed seamless repeat. Keep motifs inside the frame, then tile 2 x 2 and check.
  • No usable animation frames. A sprite sheet is concept art.
  • Edit cannot zoom out or extend the canvas.
  • Rare or technical Chinese, Japanese or Korean characters come out malformed at any size.
  • Technical diagrams look right but are not accurate. Do not use them as reference material.
  • Chart bars are not proportional to their values. Draw a data chart with code.
  • Grid counts drift. A sheet of 8 icons can come back with 9, and tiles do not connect.
  • Quoted names can lose their letter order ("Sheryl" came back "Sherly"). Check names letter by letter.
  • Small Japanese, Korean or Arabic lines come out with wrong characters, even when the headline above them is exact. Set small non-Latin copy in a layout tool.

Use a small concept when the brief and budget allow it. Compare one meaningful variable at a time when testing a style. Keep known settings and seed fixed for a controlled comparison. A smaller concept helps composition decisions but cannot prove fine text or final-resolution detail. If the user declines a concept, tell them which styles in the brief the style library marks partial or does not list, and what to watch for in the final. Changing dimensions can change composition; do not describe a new render as a pixel upscale.

Batch independent shots with one backend when useful. Do not put an edit that needs an earlier output into the same batch: the reference bytes do not exist yet. Save requests and operation IDs before continuing, and download completed images promptly, including partial successes.

Do not switch to the full model only to raise the quality. Six finished Turbo prompts, made again on full Qwen-Image-2.1 at its own 40 steps, took about five times as long. Two came back richer (finer stained glass, more dragons), three about the same, and one worse (the jellyfish turned into flat saucers). Offer the full model as another try at a select, and compare both.

Inspect, select, and refine

Inspect saved images, not just API status. A contact sheet helps choose a direction; inspect selects individually at full size for the details that matter. Check against the brief:

  • Subject identity, product shape, proportions, label spelling, and material finish.
  • Hands and contact points, plausible shadows/reflections, duplicated or merged objects.
  • For complex scenes, exact inventories, containment and object-to-object relationships; separate physical objects from their reflections and occluded parts.
  • Framing, edge cutoffs, intended aspect ratio, useful copy space, and readable exact text.
  • Requested camera height and viewpoint, relative object scale, focus target, blur direction, and the light/shadow arrangement; do not infer adherence from a lens name in metadata.
  • Consistency across the set without accidentally repeating every composition.

Adapt these checks to the intended language: flattened space or distorted anatomy may be deliberate. Judge whether the image follows the chosen direction before calling those defects. For style comparisons, assess the rendering method separately from subject retention.

Keep the chosen original. Make a targeted edit with a concrete change and explicit preservation instructions; compare it with its source. A language instruction to change a region is not a mask. Where the mode offers region marking, a mark places a change precisely (see editing), but it does not protect the rest of the frame either. Do not promise that pixels outside any edit stay unchanged.

Stop when the requested set meets the brief or its budget is exhausted. Report remaining defects instead of silently launching more attempts. Preserve successful items if a later item fails.

Deliver

Save selects with descriptive filenames and report their paths, dimensions, and useful selection notes. For a multi-image shoot, keep a compact record linking each select to its request, operation/item ID, seed, reference files, and edits; retain the original references. A simple single-image task only needs enough provenance to reproduce or revise it. Separate requested finals from concepts and rejected attempts. Mention unresolved fidelity or text issues. Use local links or the user's requested destination; publishing is a separate action. Run the finishing steps the end use needs (cutout, trace, resize, upscale, alt text) from delivery, and say which ones you ran and which the user still has to.

Words used here

  • Brief: what the user wants, in their words: subject, use, style.
  • Concept: a small, fast image that shows one possible direction before the final.
  • Select: an image the user chose to keep.
  • Hero image: the main image of a page, banner or set.
  • Style library: the table of tested styles, one prompt sentence each.
  • Looks and Rewrite: the studio's own features. Looks adds preset style text to the prompt; Rewrite expands a short request into a full prompt.
  • Accuracy level: how true the image must be: mood only, roughly right, or must be exact.
  • Safe area: the part of a banner or post that every device shows uncropped.
  • Dieline: the flat cutting template of a package, drawn by a designer, not by the model.

04 同仓库更多 Skills