In short: bring a drawing - photographed, scanned or dropped onto a reference plane in the scene - and the assistant builds the model to match it, guessing depth from the outline and any hints you give. The first pass is a shape, not a finished asset: compare it against the drawing, correct proportions by name, then detail and clean up. Where it stops: drawings carry far less information than photos, so the less the page shows, the more the result interprets - and interpretation still needs an artist's eye at the end.
What a sketch gives the generator - and what it does not
- Silhouette and proportion - the drawing is a contract for the outline; that is what converts faithfully.
- Not depth - the profile is inferred. A side view alongside the front view is the cheapest accuracy upgrade available.
- Not material - pencil tone is ambiguous. Say in the brief whether shading is texture, lighting or nothing.
- Not hidden geometry - whatever the drawing occludes is invented; expect the back of the model to be a sensible guess, not a design.
Sketch hygiene: what to fix before you feed it in
- One object per sketch - three designs on one page ask for three objects stitched into one.
- Strong outer contour - the outline is the data. Faint, sketchy edges convert into uncertain surfaces.
- Flat capture, high contrast - photograph the page straight-on, in even light; shadows across the paper are noise.
- Annotate in the brief, not on the page - say "mirror the left arm" in the prompt; arrows and notes drawn on the sketch confuse the read.
Describe a conversion so it lands
The brief adds the dimension the drawing lacks. Compare:
| Vague prompt | Prompt that lands |
|---|---|
| make this drawing | A chair from this side-view line drawing: keep the silhouette exact, infer a sensible seat depth, flat colours |
| my sketch | This front-view robot sketch: match the proportions, treat the pencil shading as lighting only, not texture |
| concept art | This painted concept of a hut: match the silhouette and palette, simplify details into a game-ready prop |
| two views | These front and side views of the same character: use the side view for depth, keep both views consistent |
From first pass to finished model
01 - Generate to match the drawing. Judge silhouette first: overlay or eyeball the render against the sketch before touching detail.
02 - Correct by name. "Add 20 percent to the leg length", "widen the shoulder pads", "flatten the backpack" - each pass edits the same object.
03 - Add the absent dimension. The model is the moment to add depth cues the drawing could not show: rounded volumes, bevels, thickness. If you have a side view, feed it in before this step.
04 - Clean up. Run a retopology pass if the target is a game or animation asset; keep it dense only if it is a display piece.
05 - Finish and export. Textures, scale check, export - the same mesh-finishing workflow as any other generated model.
Sketch, photo, or prompt?
The input changes the result contract. A text prompt invents everything and follows style; a drawing fixes the silhouette and invents depth; a photo fixes proportion, depth cues and material at once. Pick the cheapest input that carries the information you actually care about - for "the shape in my head", that is rarely a prompt and usually a five-minute sketch.
Where hand work still wins
- Design intent - the differences between the sketch and the render are decisions; an artist decides which ones are improvements.
- Text and logos - lettering and precise graphic shapes survive drawing better than conversion.
- Mechanical precision - engineering parts with tolerances belong in parametric modelling, not generation.
- Multi-view consistency - complex characters with four orthographic views still need an artist to reconcile what each view contradicts.