sketch to mockup

Making a pretty image from a sketch is easy. Making the generator keep the designer's choices is the whole job.

Image generators "beautify". Left alone they straighten an asymmetric seam, drop a contrast panel, add pockets nobody drew, swap a bell-bottom for a straight leg, shrink a volume sleeve to a normal one, copy the reference photo pasted in the corner, and print the designer's red handwriting onto the garment. This skill is a fidelity workflow that works with any generator:

1. 1Read the sketch the way a pattern-maker would.
2. 2Choose the render style and the generator.
3. 3Lock the design in plain words.
4. 4Write one master prompt.
5. 5Translate it for the person's generator.
6. 6Check the result against the lock and fix only what failed.

The sketch does the structural work wherever the generator accepts an image; the words defend the details the generator would otherwise "improve".

1. Read the sketch

A design page usually carries five layers drawn on top of each other. Separate them before writing anything, because mixing them up is the most common cause of a wrong render.

For each design element, note:

• What it is, in construction terms (raglan seam, welt pocket, stand collar, funnel neck, patch pocket, box pleat, raw edge, piping, gusset).
• Where it is. Say "viewer's left" or "wearer's left" explicitly; generators mirror things.
• How many, and whether it is symmetric.
• How big, relative to the whole object.

Describe proportions against body or object landmarks, not adjectives. "Hem ends at the high hip" beats "cropped"; "sleeve is twice the width of the torso at the elbow, gathered into a narrow cuff" beats "puffy sleeves"; "legs widen from the knee to a hem that covers the shoe" beats "wide-leg".

Several designs on one page. Two figures side by side may be front and back of one garment, or two different looks. Check (same closures and seams mirrored = one garment). Render separate looks separately unless the person wants a line-up.

Ambiguity. If a line could be a seam or a fold, or an inner shape could be lining or print, take the reading that agrees with the annotation. If it's still unclear, either ask one short question or render the likelier reading and say which one you took.

2. Choose the render style and the generator

Render style. If the person shared reference images of the finish they want, read the style off them (mannequin type, background, lighting, colour grade, crop) and use that. Otherwise offer this menu and default to A for clothing and E for objects:

• A. 3D garment render. CLO3D / Marvelous Designer look: faceless matte-grey or chrome mannequin, soft grey cyclorama, visible stitching, true fabric weight and drape. The best way to check a design.
• B. Studio lookbook. A real model with the face hidden (balaclava, cap pulled low, or cropped), seamless white or light-grey backdrop, soft editorial light, often black and white.
• C. Editorial / campaign. Dramatic studio portrait or on-location shot with mood.
• D. E-commerce packshot. Ghost mannequin or flat lay, even light, pure white background.
• E. Product render. For non-garment objects: KeyShot-style studio render, three-quarter view, soft contact shadow.
• F. Detail sheet. Close-ups of construction (seams, hardware, material texture) beside the full view.

Generator. Ask which tool they use, if they haven't said. If you, the AI running this skill, can generate images yourself, offer to render it directly with the sketch attached. If the person has no preference, write the universal image-edit version (section 5, family 1). It works in the most tools.

Ask at most one question at a time. If neither is given, go ahead with the defaults and say what you assumed.

3. Write the design lock

The design lock is a numbered list of the non-negotiables: 6 to 15 concrete, visible facts, most important first. Silhouette and proportion come first, then construction lines, then colour and contrast, then small details. Each line must be something you could point at in the finished image.

Then add a short materials and colour spec. Use what the notes say. Where the sketch says nothing, choose sensible defaults and mark them as assumptions, so the person can change them:

DESIGN LOCK
1. Silhouette: ...
2. Length / proportion: ...
3. ...
MATERIALS & COLOUR
- Shell: ... (from note: "...")
- Lining / contrast: ...
- Hardware: ...
- Assumed (not in sketch): ...

Show the lock to the person before or with the prompts. It is the part they will correct ("no, that's a seam, not a pocket"), and correcting it is far cheaper than re-rendering.

4. Write the master prompt

One tool-neutral prompt, in this order. The order matters: most generators weight early words more heavily.

1. 1Task and fidelity. "Render the attached sketch as a realistic [render style]. Reproduce the design exactly as drawn."
2. 2Subject and silhouette. What it is and its overall shape and proportions.
3. 3Design details. The design lock, as flowing sentences or a tight list.
4. 4Materials and colour.
5. 5Presentation. Mannequin/model/object setup, pose, crop, view (front, three-quarter, back).
6. 6Camera and light. e.g. "85mm, eye level, soft key light from upper left, gentle fill, subtle contact shadow".
7. 7Background.
8. 8Exclusions. "No text, no handwriting, no arrows, no sketch lines, no logos, no extra pockets, seams or trims that are not in the sketch; do not copy the small reference photos."

Template:

Render the attached sketch as a [STYLE]. Reproduce the design exactly as drawn: same silhouette, proportions, seam lines, panels and details. Do not simplify, add or "improve" anything.

[SUBJECT]: [one sentence on what it is and its overall shape].
Details: [design lock items 1–N as short sentences].
Materials and colour: [spec].
Presentation: [mannequin/model/object, pose, view, crop].
Camera and light: [lens, angle, lighting].
Background: [backdrop].
Exclude: handwritten notes, arrows, guide lines and the croquis body; text and logos; any pocket, seam, trim or hardware not in the sketch; the reference photos pasted on the page.

5. Translate for the generator

Generators fall into four families according to how they take the sketch. Find the person's tool, then use that family's version.

Family 1: instruction-based image editors (best fidelity)

ChatGPT image generation, Gemini image generation, Flux Kontext, Qwen-Image-Edit, Seedream, and the edit/reference modes of Ideogram, Leonardo, Krea and others. Adobe Firefly belongs here when used with Structure reference at high strength.

• Upload the sketch and paste the master prompt as written. It is already in instruction form.
• Start with "Use the attached sketch as the exact design blueprint" so the tool edits rather than reimagines.
• If the tool lets you attach a second image as a style reference (one of the person's mood images, for example), say which image is which: "Image 1 is the design. Image 2 is only for lighting, background and render style. Do not take any garment details from image 2."
• Iterate inside the same chat or session, so the tool keeps the context.

Family 2: Midjourney

Midjourney reads short, comma-phrased prompts and forgets long lists, and an image prompt alone gives only loose structure.

• Best route: the web Editor's Retexture. Upload the sketch, choose Retexture, and paste the prompt. It keeps the sketch's shapes and replaces the surface, which is exactly this job.
• Otherwise, use an image prompt. Put the sketch URL first, then the text, with a high image weight (--iw toward the top of the range your version allows).
• Shrink the design lock to its 5 or 6 most visible features, front-loaded. Turn exclusions into --no.
• Add --style raw (less Midjourney "flavour"), and set --ar to match the sketch (2:3 or 9:16 for a full figure).
• Parameters change between versions, so check the current docs if one is rejected.

[sketch URL] photoreal 3D garment render, [subject], [lock feature 1], [2], [3], [4], [5], [material], faceless matte grey mannequin, soft grey studio cyclorama, soft key light --ar 2:3 --style raw --iw [high] --no text, handwriting, arrows, sketch lines, logos, extra pockets

Family 3: Stable Diffusion, SDXL and Flux with ControlNet

For ComfyUI, Automatic1111/Forge, Fooocus, InvokeAI and similar.

• Feed the sketch into ControlNet lineart (or scribble for very loose sketches, canny for clean flats). Starting points: control strength 0.8 to 1.0, ending around 75 to 85% of the steps. Lower the strength if the output still looks like a coloured drawing; raise it if details drift.
• Without ControlNet, use img2img on the sketch with denoise around 0.6 to 0.75.
• Positive prompt: the master prompt condensed into descriptive phrases (these models don't follow instructions). Negative prompt: the exclusions plus sketch, drawing, illustration, lineart, flat colours, cartoon, deformed, extra pockets, text, watermark.
• Generate a batch, keep the most faithful, and fix local errors with inpainting.

Family 4: text-only generators (no image input)

Use the full master prompt, replace "the attached sketch" with a complete written description, and expect drift. Spend two or three rounds on the check-and-fix loop.

Preparing the sketch (all families)

• Upload a clean version. Crop away reference insets and handwritten notes where you can; the notes have already become words in the prompt. If they can't be cropped, the exclusions line covers them.
• One design per image. Split a multi-look page.
• Raise the contrast. Dark lines on a white background, straightened, no paper shadow. For ControlNet, a clean sketch matters more than any setting.
• Back views. If the sketch has a back view, render it separately with the same lock, plus "back view".

6. Check and fix

When the person shares the render, or when you generated it yourself, go down the design lock line by line and mark each item kept, changed or missing. Look closely at the usual failures:

Write a fix prompt that names only what failed. In edit-capable tools, run it on the latest render, not the sketch:

Keep everything in this image exactly the same. Change only:
1. [failed item]: [what it should be].
2. ...
Do not change anything else.

Stop when every lock item reads "kept", or when the person is happy.

Other kinds of object

The workflow is the same; only the vocabulary changes.

• Footwear: upper panels, toe box, vamp, eyestay, lace count, collar height, heel counter, midsole thickness and profile, outsole lug pattern. Render style: three-quarter lateral view, product render.
• Bags and accessories: gussets, handle drop, strap attachment, hardware finish (brushed nickel, matte black, gunmetal), edge paint, stitch colour.
• Jewellery and hardware: metal and finish (polished, satin, oxidised), stone setting, scale against a hand or finger. Use macro lens language.
• Furniture: frame material, joinery, upholstery, leg profile, overall dimensions, seat height. Use a room or a seamless set.
• Devices and industrial design: parting lines, radii, button and port positions, CMF (colour, material, finish), screen state.
• Packaging: dieline shape, panel layout, finish (matte, soft-touch, foil). Only render artwork or text the person supplied.

What to hand back

Always return, in this order:

1. 1Design read: two or three lines on what you saw, plus any reading you were unsure of.
2. 2Design lock: the numbered list, plus materials and colour, with assumptions marked.
3. 3Prompt(s): the master prompt and the version for their generator, each in its own code block so it can be copied.
4. 4Sketch prep: one line on what to crop or clean before uploading.
5. 5After you render: one line inviting them to share the result so you can check it against the lock.

If you can generate images yourself and the person wants that, render it, then run section 6 on your own output before showing it.

Worked example

Sketch: a croquis wearing trousers. The solid lines show a mid-rise waistband with a single button, slant front pockets, and a vertical seam down the centre front of each leg, drawn as a double line. The legs run straight to the knee and then flare out widely to a hem that pools over the foot. Notes read "contrast piping" (pointing at the leg seams) and "leather bell bottoms". The inset photo shows dark, shiny, flared leather trousers.

Design lock

1. 1Bell-bottom trousers: fitted through hip and thigh, straight to the knee, then flaring sharply to a wide hem that covers the shoe and breaks on the floor.
2. 2Mid-rise waistband with a single-button front closure and belt loops.
3. 3Slanted front hip pockets, one each side.
4. 4A single vertical seam down the centre front of each leg, from waistband to hem, finished with contrast piping.
5. 5No other seams, pockets or trims.

Materials and colour: black leather with a soft natural sheen (from note); piping in off-white leather (assumed; the sketch doesn't give a colour). The inset contributes only "shiny black leather, flared". Nothing else is taken from it.

Family 1 prompt

Use the attached sketch as the exact design blueprint and render it as a photoreal 3D garment render. Reproduce the design exactly as drawn; do not simplify, add or improve anything.

Subject: black leather bell-bottom trousers on a faceless matte grey mannequin, full length, front view.
Details: fitted through hip and thigh, straight to the knee, then flaring sharply to a wide hem that covers the shoe and breaks on the floor. Mid-rise waistband, single-button closure, belt loops. One slanted hip pocket on each side. A single vertical seam runs down the centre front of each leg from waistband to hem, finished with off-white contrast piping. No other seams, pockets or trims.
Materials: supple black leather with a soft natural sheen and gentle creasing at the knee; off-white leather piping.
Camera and light: 85mm, eye level, soft key light from upper left, gentle fill, subtle floor shadow.
Background: seamless soft grey studio cyclorama.
Exclude: handwritten notes, arrows, guide lines, the croquis body, text, logos, the reference photo, and any detail not in the sketch.

Midjourney (Retexture, or with the sketch as image prompt)

photoreal 3D garment render, black leather bell-bottom trousers, fitted thigh flaring sharply from the knee to a floor-length wide hem, off-white contrast piping down centre front of each leg, single-button waistband, slant hip pockets, faceless matte grey mannequin, soft grey studio cyclorama --ar 2:3 --style raw --no text, handwriting, arrows, sketch lines, logos, extra seams