Test plan · Generators
Photo vs Concept Art as Image-to-3D Input: How to Run a Controlled Comparison
Does image-to-3D handle a photo, a render, painted concept art or a sketch equally? What vendors report, what they leave out, and how to run the comparison yourself.
This is one of the same-input test protocols. Image-to-3D tools are marketed as accepting “any image,” but a studio photo, a clean render, a painted concept and a pencil sketch give a model very different information. This page covers what the published benchmarks do and do not tell you about input type, and how to run a controlled comparison with your own images. For preparing an image well, see preparing images for image-to-3D.
What the published data covers
Meshy’s geometry benchmark (12 August 2026). Meshy published an automatic benchmark that renders held-out reference 3D models with known cameras, feeds them to each model, and scores overall proportion, spatial distribution and surface detail against the reference. It compares Meshy 7, Meshy 6 Lite, Tripo 3.1, Hunyuan3D 3.1, Rodin 2.5 and Hi3D 2.1 with front, top-quarter and four-view inputs. Its limits for this question: it is run by a vendor that is one of the contestants, the benchmark was not yet released when we read the post (“will be made available separately after the Meshy 7 release”, no date; Meshy’s meshybench repository now publishes the scoring code), and every input is a render, so photos, paintings and sketches are not covered. For the four-view results the post says the gap between leaders is “small relative to the variation across individual test objects”, which argues for more than one object per type.
3D AI Studio’s guide on single versus multiple photos. It gives accuracy percentages for single and multiple photos but no method, sample size or sources, and the publisher sells 3D generation. It does not compare input types such as photo versus sketch.
Meshy’s multi-view tutorial. It gives input specs (photos of at least 1040 by 1040 pixels, plain background, even lighting, up to 20 MB each) and says a single pre-composited sheet such as a character turnaround can serve as the main upload. That is documentation of a recommended input, not a test.
So the open question, how input type changes the result, is untested in the sources we read.
What the API documentation says about inputs
Meshy’s Image to 3D API (docs.meshy.ai/en/api/image-to-3d, read 2026-11-08) states the hard limits: JPG, JPEG, PNG or WebP; at most 20 MB; at least 32 pixels per side; at most 178,956,970 total pixels. These are accept/reject limits, not quality advice; the multi-view tutorial’s 1040-pixel figure is a recommendation. Two defaults matter for an input-type comparison. image_enhancement is on by default and, per the docs, optimizes the input image before generating; it can be set to false to keep the exact look of the input, and works only with meshy-6, meshy-7.1 or latest. remove_lighting is on by default and strips highlights and shadows from the base color, but only for the Meshy 6 texture model. So a photo with strong lighting and a flat painting may be preprocessed differently from how you assume; log both settings. For texturing from a separate image (texture_image_url), the docs warn it “might not work well when the image’s shape differs a lot from the 3D model’s geometry”; that concerns texturing an existing model, not photo versus sketch as the geometry input, so use it only as a reason to keep any texture image matched to the mesh. Free-plan note: Meshy’s pricing page marks multi-view image to 3D as not included on Free, and its help article says Free accounts get 10 downloads a month, for untextured Meshy 6 Lite models only (November 2026), so a free run can test single-image geometry only.
Design: change only the input type
Use I01 to I12 from the standard set: three subjects, each as a photo, a clean render, painted concept art and a sketch. For a meaningful comparison:
- Use the same subject across the four input types so only the input differs.
- Run each case once per tool at default settings; log plan and model version.
- For objects you own, keep the physical object or a known 3D model so the back and sides can be checked against truth.
- Do not enhance or crop images differently across tools.
Scores
| Field | Definition |
|---|---|
| input_type | photo, render, painted, sketch |
| subject_id | Which of the three subjects |
| front_match | 0 to 2, front silhouette vs input |
| back_plausibility | 0 to 2, back and sides vs reference or a sensible guess |
| detail_kept | 0 to 2, logo, text, small parts retained |
| background_leak | yes/no, floor, shadow or props baked into the mesh |
| lighting_baked | yes/no, shadows or highlights in the albedo |
| style_drift | 0 to 2, for painted and sketch inputs, how much the style was lost |
| triangles, loose_pieces, open_edges | From the script |
Anchors for 0 to 2 are on the prompt set page. Scoring blind is easiest to do by renaming files to run IDs.
Ambiguous results
Rules to set before you start, for borderline readings; they are editorial, not outcomes we have seen.
- Back is plausible but different from the truth object: score back_plausibility 1, and note “invented detail” so you do not mistake a good guess for fidelity.
- Front matches but lighting is baked: front_match can be 2 while lighting_baked is yes; do not merge them. Retest with
image_enhancementoff if the tool exposes it, and record the change. - Sketch yields a plausible shape unrelated to the drawing: style_drift 2, front_match 0. A confident model is not a faithful one.
- Scorers disagree by 1: have a second person score blind; if they still differ, use the lower score.
- Preprocessing was applied (vendor enhancement, background removal): log it as part of the tool’s behavior; do not pre-clean one input type more than another.
Pass or fail
Pass for a given input type means front_match and back_plausibility are both at least 1 on every one of the three subjects, with no background leak. A tool that passes photos and fails sketches is a legitimate result: it tells you what to feed it. Where the back is wrong on a photo, move to multi-view input.
Where to go next
If your input is a photo of a real object and you want the geometry rather than a stylized guess, compare against AI generation versus photogrammetry. For concept art of characters, see from concept art to a 3D character. For a tool shortlist, see best image-to-3D tools.