The single most effective way to improve quality in AI 3D generation (image-to-3D) isn't tweaking the generation settings. It's preparing the image you feed in. Anything that doesn't appear in a single image has to be filled in by the AI's guesswork, so the groundwork you do beforehand, such as narrowing it down to one subject, including the full body, reducing overlapping parts, and separating the outline from the background, translates directly into differences in the result. In this article, we go through seven points that make a real difference in your input image, along with how to pin down the cause when things don't go well.

Note: Figmee currently lets you generate and download figurine-style images and 3D model data (GLB / 3MF) from drawings and illustrations you upload. Sales of physical 3D-printed pieces are Coming Soon.

Why the input image changes the result

Reconstructing a 3D shape from a single flat image is a problem that, by its nature, is solved with too little information. Research on 3D generation from a single image describes this situation as an "under-constrained" setting, and takes the approach of filling in other viewpoints by relying on the geometric priors that large-scale image generation models have learned (source: Zero-1-to-3: Zero-shot One Image to 3D Object (arXiv), as of July 2026).

In other words, the AI fills in the parts it can't see using "typical shapes." This points to a clear strategy for your input image: the less the AI has to fill in, the closer the result gets to what you intended. A clear outline keeps the shape estimate from wobbling, a full-body view locks in the overall proportions, and parts that don't overlap leave no room to misread what's in front of what.

There's one more thing worth knowing: the particular difficulty of illustrations. Unlike photos, hand-drawn flat coloring offers few shading cues, so from the AI's point of view it's an input with little material to estimate volume from. Research addressing this problem points out that when the reference image uses flat colors like a hand drawing, the lack of depth cues makes it hard to synthesize a plausible 3D asset (source: Art3D: Training-Free 3D Generation from Flat-Colored Illustration (arXiv), as of July 2026). Line art and anime-style cel shading tend to struggle not because the drawing is bad, but because of the nature of the input.

7 points for your input image

1. Keep it to a single subject (about 5 min)

When several characters or props are lined up together, the AI can't tell where one 3D object ends and another begins. As a result, everything may merge into a single lump, or one of them may be dropped. Keep only the subject you want to make, and remove everything else by cutting it out or cropping. If the composition includes a pet or small items in the background, start by turning it into an image of just the main character.

2. Frame the full body or the whole object (about 5 min)

Parts cut off at the edge of the frame naturally carry no information. A standing pose with the feet cropped out leaves the shape of the legs undefined, so the AI ends up adding them by guesswork. Ideally, the composition should include everything from head to toe, or the entire object, with some margin around it.

3. Reduce overlapping parts and occlusion (about 10 min)

Arms crossed in front of the body, hair falling over the face, a prop hiding the body. With occlusions like these, information about the hidden side is completely lost, so the share of the result that depends on guesswork jumps. Simply redrawing the pose so the arms are slightly away from the body, or the hands hang down at the sides, makes the result more stable. Keep in mind that a pose close to a T-pose, where each part reads as separate, is the easiest to work with.

4. Draw or shoot from an angle close to the front (about 5 min)

Angled or low-angle compositions may be appealing as artwork, but they work against you when estimating a 3D shape. An angle close to the front makes it easier to read the left-right balance and how the parts are positioned relative to each other. Even if you want to turn a dynamic pose into 3D, it's no detour to first make one in a simple frontal pose and confirm the shape comes out well before taking on the challenge.

5. Separate the outline from the background (about 5 min)

If the boundary between the subject and the background is vague, the AI misreads the outline. If the background is close in color to the subject, or contains patterns or other elements, replace it with something close to a plain background. When you photograph a drawing, simply shooting it in front of white paper or a plain wall makes a big improvement.

6. Add light shading to the coloring (about 15 min)

As mentioned above, completely flat coloring is an input with few cues for volume. You don't need to render detailed shadows, but just a little light and dark, such as a light shadow on rounded surfaces or slightly darker recessed areas, helps the shape be interpreted more consistently. If you want to turn line-only art into 3D, try starting by making a version with color and simple shading.

7. Avoid harsh shadows, blown-out highlights, and blur (about 5 min)

This applies when your input is a photo, for example a photo of a drawing on paper. Shadows cast by strong light from an angle can't be told apart from the shape of the subject itself, and may get baked into the 3D result. Shoot under bright, even lighting, with the image in focus. We've gathered specific steps for shooting in How to Photograph Children’s Drawings for Better 3D Results.

Troubleshooting table when things go wrong

When you're not happy with a generated result, you can work backward from the symptom to figure out what to fix in the input.

SymptomLikely causeFix on the input side
The back design is differentThe image has no information about the backPrepare a rough sketch of the back / fix it by other means
Thin parts go missing or fuse togetherProtrusions or parts are too thinDraw them thicker / change the pose so they stand apart
What's in front of what is wrongOverlapping parts cause occlusionRedraw with arms and props away from the body
The whole thing looks flatFlat coloring gives few shading cuesTry a version with light shading added
The outline is blurry or swallows the backgroundThe boundary between background and subject is vagueReplace it with a plain background
The proportions fall apartThe full body isn't included / low-angle compositionUse a near-frontal composition that fits the full body

Most cases where it feels like "it can't be turned into 3D" fall under one of the rows in this table. For a detailed explanation of each cause, see Why People Say Illustrations Can't Be Turned into 3D | Cases Even AI Struggles With.

When drawing something new vs. using existing art

If you're about to draw it, you can plan the composition with 3D in mind from the start. Front view, full body, a pose with parts held apart, a plain background. Draw one rough sketch that meets these four conditions, and turning that character into 3D becomes much easier. A practical approach is to prepare "one image for 3D" separately from your finished, fully detailed illustration.

If you're using existing art, start with adjustments you can make without redrawing. Crop it down to just the main character, fill the background with a single color, and raise the brightness and contrast. These three steps take only a few minutes in an image editing app, and it's not unusual for them alone to change the result.

In Figmee, you can swap the image used for generation, so you can work by comparing results while changing the input for the same character. Rather than aiming for perfection in one try, changing one condition at a time to get a feel for it gets you where you want to go faster in the end.

Tweaks that do little or tend to backfire

Edits made with good intentions can sometimes make the result worse. Watch out for the following four.

  • Pushing contrast to the extreme: Wanting a clearer outline is the right instinct, but going too far wipes out the midtones and erases even the light and dark that show the roundness of surfaces. It's enough that the boundary can be distinguished.
  • Applying strong sharpening: When unnatural halos appear along the outline, the halo itself can be interpreted as part of the shape.
  • Making the background pitch black: With a dark-colored subject, the background and outline become impossible to tell apart. A plain background is all you need, and light gray or white is the safer choice.
  • Forcing the image to be upscaled: Enlarging a small image doesn't add information. If anything, it blurs the outline and works against you. Pass it along at its original size.

What these have in common is that edits that "give the AI more information" help, while edits that "crush information" backfire. If in doubt, the surefire approach is to try both the unedited original and the edited version and compare.

Photos and illustrations have different weak spots

Even though both count as an "input image," photos and illustrations show their weaknesses in different ways.

With photos, the shape information itself is often sufficient, and the problems are lighting and background. The typical failures are strong shadows being interpreted as part of the subject, or background elements mixing into the outline. If you can provide even lighting and a plain background, most of these are solved.

With illustrations, lighting problems are less likely, but the scarcity of shading cues and stylization come into play instead. Characters with heavily exaggerated head-to-body ratios, or expressions that work as a flat image but contradict themselves in 3D (overlapping strands of hair, symbolic ways of drawing eyes, and so on), split the interpretation when turned into 3D. When things don't go well, trying a version with slightly toned-down stylization helps you isolate the cause.

The order to check after generation

Even with a well-prepared input, skipping the check leads to rework. Looking in the following order is efficient.

  1. Overall silhouette (about 1 min) — First, from the front, check whether the proportions are as intended.
  2. The back (about 1 min) — The part filled in by guesswork. This is where the biggest differences show up.
  3. Thin parts (about 2 min) — Fingertips, hair tips, protrusions, and held items.
  4. Color and texture seams (about 2 min) — Color mismatches tend to appear where the sides wrap around to the back.
  5. Contact surface and base (about 1 min) — Often not visible in the original image, so it tends to be guesswork.

Make it a habit to check these five points in the same order every time, and you'll be able to judge for yourself "what got better this time." If you want to deepen your understanding of how it works, also see How Image-to-3D AI Works — A Beginner's Guide to Turning Drawings and Photos into 3D.

Frequently asked questions

Q. Is higher resolution always better? A. It helps up to a point, but it doesn't keep getting better the higher you go. Extremely small images (small enough that the outline gets crushed) are at a disadvantage, but once the resolution is sufficient, composition and occlusion have a bigger impact on the result from there on. The practical priority is to fix the composition first, then pick the sharpest image you have.

Q. Should I make the background transparent? A. As long as it's plain, there's no big difference between transparent and a single color. What matters is that the boundary between the subject and the background is clear. However, if you cut out the outline roughly while making it transparent, that roughness can carry over into the 3D result. If you can't cut it out cleanly, it's safer to pass it along with the plain background as is.

Q. Can line-only art be turned into 3D? A. Generation itself works, but flat line art counts as an input with few cues for estimating volume, so the results tend to look flat. Preparing a version with color laid down and light shading on rounded surfaces helps stabilize it. Conversely, if you want to keep the black-and-white feel, another option is to adjust the colors in 3D software after generation.

Q. Why does the result differ a little each time, even with the same image? A. Because the process includes estimating and filling in the parts that aren't visible. As mentioned earlier, turning a single image into 3D is solved with insufficient information, so there's a range in how the gaps get filled. Put the other way, the more you prepare the input and narrow the range that relies on guesswork, the smaller the variation in results becomes. When the variation feels large, take it as a sign to review the composition and occlusion.

Q. I've tried several times, but the back still isn't what I wanted. A. As long as you only provide a single front-facing image, the back will remain an estimate. It's an area where even a well-crafted input has limits. If you really want to refine the back, it's more reliable to import the generated 3D model into 3D software and fix it yourself, or to switch to a method that lets you provide reference for the back (a tool that supports multiple input images, or commissioning a creator).

Summary

Almost everything you could call a "tip" for AI 3D generation lies on the input image side. Narrow it to one subject, fit the full body, reduce overlapping parts, use a near-frontal angle, and separate the outline from the background. Then add a little light and dark to flat, hand-drawn coloring. If you can prepare an image that meets these six points, the range the AI has to fill in by guesswork shrinks dramatically. When things don't go well, try isolating the cause on the input side from the symptom before changing settings. More often than not, what needs fixing isn't the generation, but the drawing you're handing over.

Placing your prepared input image side by side with the 3D model made from it also makes a process post that shows how a drawing becomes 3D. For how to lay them out, see How to Post Your Art Process: Show Your Character Going from Sketch to 3D.