Most patent drawings don't start from a blank page. They start from whatever material already exists: product photos from a store listing, a pencil sketch drawn for the factory, or screenshots of an engineer's CAD model. All three can go straight into PatentFig AI's generate page and come back as line art — but each one fails in a completely different way. Photos drag in backgrounds, reflections, and lens distortion, and they can't show anything internal. Sketches lack detail and their views rarely agree on proportions. CAD screenshots carry too much truth: render shading and thread geometry turn into visual noise.
Picking the wrong input won't stop you from getting figures. It will cost you extra rounds of rework. This guide gives you a decision table for the three inputs, preparation tips and fixes for each, and three copy-paste prompts.
The same invention can enter the pipeline as a photo, a sketch, or a CAD capture — each with a different failure mode.
Quick answer: which input fits which invention
| Input | Best for | Strength | Typical rework |
|---|---|---|---|
| Product photo | Housings, enclosures, exterior styling; improvements visible from outside | Zero preparation; exterior contours captured in one pass | Remove background and reflections, force orthographic views; internals are invisible |
| Hand sketch | Mechanism-centric inventions; early-stage filings with no finished product | Every line is deliberate — the clearest carrier of structural intent | Add missing detail, reconcile proportions across views |
| CAD screenshot | Teams that already have engineering files; dimension-critical assemblies | Accurate geometry, hole positions, clearances | Strip render shading, simplify threads, hide non-claimed internals |
Short version: photos for appearance, sketches for mechanisms, CAD for geometric accuracy. Wherever the inventive point lives, start with the input that shows it best.
Photos: how to shoot, and what a photo can never recover
Four habits at capture time save half the cleanup later: shoot the product square-on, step back and avoid wide-angle lenses, use a plain background, and turn off mood lighting. Styled marketing shots — the 45-degree hero angle with props on a desk — are the worst raw material: prop outlines get absorbed into the product's linework.
Common failures and fixes:
- Background bleed. One chat instruction — "remove the background, keep only the product" — then regenerate just that figure. Treat this as a standard step on the photo route.
- Reflections become white voids. Specular highlights on glossy plastic or chrome read as blank patches or ghost contours. Ask for reflective parts to be treated as plain, matte surfaces.
- Lens perspective. A phone photo's "side view" is really a three-quarter view. Request true orthographic front and side views explicitly.
- Occlusion. A hand in a demo photo covering the key structure cannot be reconstructed. Swap the source material; no prompt fixes what the camera never saw.
And one structural limit: photos cannot produce section views. Cameras don't see inside housings. Use a sketch or an engineering diagram for internal structure, and let photos handle exterior views only.
Sketches: ugly is fine, unclear is not
A sketch's value is pre-filtering. Photos and CAD dump everything indiscriminately; every line in a sketch exists because a human decided it mattered. If you drew the cam surface of a locking collar, that's because you want it in the figure. For mechanism inventions, that control outweighs neat draftsmanship by a wide margin.
When drawing, prioritize three things: get the structural relationships of the inventive point right (how parts engage, hook shapes, the section you care about); label parts or reserve reference numerals next to them; keep multiple views at roughly the same scale.
The weaknesses are equally predictable: anything you didn't draw won't appear, and if two sketched views disagree on proportions, the generated views will disagree too. The fix is to treat the sketch output as a structural skeleton — confirm the mechanism reads correctly first, then add detail incrementally through chat edits, saving a version at each step so a bad edit is just a rollback.
Wobbly source lines are recoverable; a mechanism drawn wrong is not.
CAD screenshots: subtract first, add later
Preparation before capturing: use orthographic views, switch to wireframe or line display if your CAD package offers it, and hide irrelevant internal parts before taking the screenshot.
Even then, the first conversion is usually noisy. Rendered gradients become dense hatching; every screw thread becomes a stack of concentric circles. USPTO practice under 37 CFR 1.84 calls for durable black solid lines, with shading only where it genuinely aids understanding of shape — a raw CAD render is close to the opposite.
Work in this order: strip render shading, simplify threads to conventional symbols, hide internals that aren't part of the claim. Once the figure goes quiet, decide where a few shading or section lines actually help show curvature. Per-figure regeneration earns its keep here — typically only one or two views need repeated passes, and re-rolling the whole set wastes the views that were already fine.
Prompt examples: one dumbbell, three inputs
The examples below use an adjustable dumbbell whose inventive point is a locking collar that allows one-handed plate changes. All three prompts are ready to copy and adapt.
Example 1: photo to line art
Convert this product photo of an adjustable dumbbell into a patent line drawing. Remove the background and all props; keep only the product. Do not trace reflections on the chrome plates — treat them as plain cylinders with no highlights. Output a true orthographic side view: black solid lines on a white background, no grayscale shading. Add reference numerals 10 for the handle, 12 for the plates, 14 for the locking collar, with thin lead lines.
Example 2: sketch to a clean multi-view set (bad prompt vs good prompt)
Bad prompt:
Turn this sketch into a patent drawing.
The tool has to guess your intent, and every wrong guess is a rework cycle. A good prompt states the structural intent, the views, and the numbering rules:
This hand sketch shows the locking collar of an adjustable dumbbell. Preserve the cam surface and spring engagement exactly as drawn — that is the invention. Clean up the linework and unify proportions. Output two figures: an orthographic front view and a section view along A-A, with section hatching as thin parallel lines. Number the collar 20, the cam surface 22, the spring 24; use the same numeral for the same part in both figures. Black solid lines, no shading.
Example 3: CAD screenshot to simplified line art
Convert this orthographic CAD screenshot of an adjustable dumbbell into a patent line drawing. Remove all render shading, gradients, and material highlights. Simplify screw threads to conventional symbols. Hide internal components unrelated to the locking collar. Keep the section through the collar mechanism, hatched with thin parallel lines. All lines must be black, solid, and legible at reduced scale; retain shading lines only where needed to show curvature.
Mixing inputs: most products land in the middle
A single input covers a whole filing only at the extremes — a pure-styling product where photos do everything, or a pure-mechanism improvement with complete CAD. Most products sit between, and assigning inputs per figure is cheaper than forcing one source:
- Exterior views from photos, with background removal and orthographic conversion applied across the set;
- Mechanism figures from sketches, keeping structural intent under human control;
- Dimension-critical views from CAD, after the subtraction pass.
Three sources means three chances for part 12 to become part 21 somewhere in the set. Numeral drift is the classic multi-source error; reconcile reference numerals across all figures after assembly. If a professional drafter or foreign associate will edit the figures downstream, export vector versions (SVG or DXF) so they can work on actual line geometry rather than pixels.
Pre-filing checks
- Margins meet the target office's requirements;
- Line weight is even and survives reduction;
- Reference numerals are consistent across all views;
- Every numeral mentioned in the specification appears in a figure;
- No leftover backgrounds, props, or stray text inside any figure.
Run the assembled set through Figure Checker for the formal pass; a printable pre-filing checklist is available at patentfig.ai/resources. If international filing is on the table, hold the master set to PCT Rule 11 conventions from day one — retrofitting a US-formatted set for other offices later is the expensive way around.
FAQ
Are phone photos good enough?
Yes, if you shoot square-on, keep the background plain, and don't cover the inventive structure. Lens distortion can be corrected by requesting orthographic views at generation time, but stepping back and zooming in at capture saves a revision round.
Can I get a section view without CAD?
Yes. A section view needs correct structural relationships, not render-grade precision. Hand-draw a section diagram with parts in the right positions, mark the cutting plane, and ask for output following section-view conventions.
Can figures from different inputs go into one filing?
Yes — that's the normal case. The condition is a unification pass after assembly: consistent line style, consistent view conventions, consistent numerals. No reviewer should be able to tell which figure came from which source.
Will a badly drawn sketch hurt the output?
Wobbly lines don't matter; wrong structure does. If part positions and engagement are drawn correctly, proportions and polish can be fixed during generation and chat edits.
Everything the tool produces is a draft: whether the inventive point is shown correctly, and whether more views are needed, stays a human and attorney decision. The entry point is patentfig.ai/generate — it takes photos, sketches, and CAD screenshots directly.
Top comments (0)