For three years I've been prompting the same character from slightly different angles, hoping this time the lighting will cooperate. AI image generators are incredible at "facing left, dramatic side light, golden hour" on the fifteenth try. What they can't do is give me that same character three seconds later in a different pose, under a different key light, from a camera angle I just decided I want. That's a 3D problem. And I'm a 2D artist who's never opened Blender.
Last month I spent two weeks doing nothing but testing AI 3D generation tools. Meshy AI, Luma AI's Genie feature, Tripo3D, and CSM.ai. The pitch for all of them is essentially the same: describe or photograph something, and get back a usable 3D mesh. No modeling skills required. The gap between 2D and 3D, closed.
Here's what I actually found.
Why I Wanted 3D in the First Place
My work is primarily illustrative. Characters in environments, often Ghanaian figures, architectural scenes, decorative compositions. I work in Procreate and Photoshop, and lately I've been using AI image generation for reference and texture iteration. The problem I kept running into: once I've designed a character or object I love, I can't reuse it. Every new pose is a new prompt lottery. Every new angle is an hour of coaxing out of Midjourney or Stable Diffusion.
What I actually wanted was a workflow where I design a character, generate a 3D version, bring it into a renderer, pose and light it myself, and then composite the render into a 2D illustration. Photographers call this using 3D for lighting reference. For illustrators it's closer to building your own poseable mannequin that actually looks like your character, not some generic wooden doll from an art supply store.
The question was whether the AI 3D tools were at the point where this workflow was actually viable.
Meshy AI: The Most Useful One I Tested
Meshy AI (meshy.ai) handles both text-to-3D and image-to-3D, which makes it immediately more useful than tools that only do one. The image-to-3D workflow is what I spent most of my time with: you upload a reference image, and Meshy generates a full 3D mesh you can download in GLB, OBJ, FBX, or other formats.
For objects, it's genuinely impressive. I fed it photos of a wooden stool, a ceramic pot, and a carved wooden mask. The stool came back with clean geometry, good proportions, and enough surface detail to be usable as a reference object in a composition. The mask was more complex and the AI struggled with the undercutting in the carving, but the basic form was there.
Characters are where things get complicated. Meshy's text-to-3D generates humanoid figures that are, charitably, competent. The proportions are readable, the pose is neutral and riggable, and the texture baking is decent. But human faces come back looking like someone tried to reconstruct a face from a police sketch. Soft, slightly wrong, emotionally vacant. This isn't specific to Meshy; it's a pattern across all four tools I tested. The face problem in AI 3D is roughly where the face problem in AI images was in 2022.
On pricing: Meshy runs on a credit system. The free tier gives you 200 credits monthly, which buys about 10-15 generation attempts depending on quality settings. The Pro plan is $20/month for 2,000 credits. The Max plan at $48/month gives you 8,000 credits and faster generation. For regular creative use, Pro is workable. For heavy production use, you'll burn through credits faster than you expect.
The feature I use most is image-to-3D with texture refinement. You upload a 2D drawing, Meshy generates the mesh, and then there's a second pass where you can use an additional reference image to improve the texture quality. For getting a character design into 3D space with its color palette intact, this workflow gets surprisingly far. The mesh won't win any topology awards, but for lighting reference it's more than enough.
Luma AI Genie: Best for Reference Photography
Luma AI's Genie feature lives inside Luma Dream Machine (lumalabs.ai), and it's a different kind of tool than Meshy. Where Meshy is primarily generation-from-prompt or generation-from-single-image, Luma's 3D output is rooted in its video and multi-view processing capabilities. The results have a photogrammetry feel: if you feed it something photographically grounded, the 3D reconstruction is spatially coherent in a way pure generation isn't.
My main use case for Luma became turning physical reference objects into 3D assets. If you have an actual object you want in 3D, and you can photograph it from multiple angles (or find multi-view reference photos online), Luma does a better job than any of the other tools at capturing spatial accuracy. The prop reconstruction quality is noticeably better than Meshy's for photographic input.
The limitation is that it's harder to use for purely generated characters or non-photographic input. Luma wants photographs, not drawings or AI images. When I fed it my own illustration as the source, the results were blurry and geometrically confused. When I fed it actual product photos of objects I wanted to reference, the results were often directly usable.
For a working illustrator, the Luma workflow looks like this: find or take reference photos of a real-world prop or object, reconstruct it in Luma, bring the resulting mesh into Blender or your renderer of choice, light and position it, render, composite. It's more steps than I'd like, but it produces better spatial grounding than anything I could get from prompting alone.
Tripo3D: Fast, for When Fast Is What You Need
Tripo3D (tripo3d.ai) is the speed option. Generation times are noticeably faster than Meshy or Luma, often under a minute for a basic mesh, and the interface is straightforward. You type a description or upload an image and get a mesh back quickly. For iterating on object shapes, this is genuinely useful.
The quality tradeoff is real. Tripo3D's meshes tend to be lower poly and less detailed than Meshy's equivalent outputs. Textures are flatter. The results work fine for quick blocking, for checking whether a shape idea makes compositional sense, or for thumbnailing. They're not where you want to end up for final production.
The free tier is generous for quick experimentation. The paid tiers start around $10-15/month for serious usage. If you're doing fast ideation and don't need the highest quality output, Tripo3D earns its place. I use it the way I'd use a rough sketch: fast and disposable, for checking the idea before committing to a more careful execution.
One thing Tripo3D does well that I didn't expect: architectural and vehicle forms. Clean geometric objects without organic complexity come back looking better than you'd expect for the generation speed. Rooms, furniture, simple vehicles; these all performed well. If your 3D needs are object-heavy rather than character-heavy, Tripo's speed-to-quality ratio works in your favor.
CSM.ai: When You Need It to Actually Be Good
CSM.ai (csm.ai) is the production-quality option. The meshes are cleaner, the topology is more respectable, and the texture quality is higher than any of the others. It's also slower and more expensive, and the workflow is more involved. This is a tool aimed at studios and professional developers, not casual creative experimenters.
The image-to-3D pipeline at CSM is specifically designed for game-asset production: clean quads, reasonable poly counts, baked textures with diffuse and normal maps separated out. If you're bringing an AI-generated 3D model into a real-time engine or a production renderer, CSM is the one that gives you something you can actually work with at a technical level.
For my compositing workflow, CSM's output quality is clearly the best. The normals are cleaner, the silhouettes hold up better at different angles, and the surface detail bakes more accurately. The cost is time and credits. CSM's free tier is limited. Paid plans start around $39/month for meaningful usage, and generation takes several minutes per asset rather than seconds.
My honest read: for a 2D artist using 3D as a lighting reference tool, CSM is overkill unless you're doing a lot of it and quality really matters for your compositing. For someone who needs production-ready assets regularly, it's probably the right tool. It's the one professional game studios would actually use.
The Cultural Gap Nobody Talks About
All four tools have the same problem, and none of them say anything about it in their marketing materials.
The training data for AI 3D generation skews heavily Western. You can feel this the moment you try to generate anything culturally specific that isn't European or East Asian. I tested each tool with the same set of prompts: a kente cloth draped figure, an adinkra-patterned stool, an Akan gold weight, a traditional Ghanaian kente headwrap. The results were consistently generic at best and culturally confused at worst.
Meshy's kente came back as a vaguely geometric patterned fabric that shared nothing with actual kente weave structure. The adinkra stool could have been any vaguely African-styled decorative object from a stock photo site. The gold weight looked like a malformed bronze. None of the tools had clearly ever processed significant amounts of data on these objects, and the gap between what I described and what came back was much wider than the gap I get when prompting for a chair or a medieval sword.
This isn't just an aesthetic disappointment. It means these tools are significantly less useful for artists whose practice is rooted in cultures underrepresented in tech training data. My work is Ghanaian. When I want to generate reference objects from that visual tradition, I can't get the same quality and accuracy that a European artist would get generating a Greek urn or a Victorian armchair. The baseline quality of the tools is unequal depending on who you are and what you're making.
The workaround, such as it is, is to use Luma's photogrammetry approach with actual photographs of real objects rather than relying on generation. If you have access to real examples of culturally specific objects to photograph, Luma will reconstruct them accurately. But that's a higher friction workflow, and it puts the burden of filling the training data gap onto the artist rather than the tool.
My Actual Workflow: How I Use These Tools Now
After two weeks of testing, I settled on something that works. It's not what I imagined going in. It's more useful than I expected in some ways, and more limited in others.
My current workflow:
- Character bodies: Use Meshy AI image-to-3D with my own character illustration as input, then run the texture refinement pass. Accept that the face will need to be composited from a 2D render or simply kept out of the shot.
- Environmental props (real-world objects): Photograph or source multi-view reference photos, reconstruct in Luma. Much better spatial accuracy than pure generation.
- Quick shape blocking: Tripo3D for fast iterations when I'm still figuring out composition and don't need final quality.
- High-stakes assets: CSM.ai when the object needs to be in a lot of final renders and quality actually matters.
- Rendering and lighting: Export to Blender. Set up a three-point light rig. Render to PNG with a transparent background or a plain backdrop. Import to Photoshop as a lighting reference layer, then paint over it.
The compositing step is the key one. I'm not trying to make a 3D render look good on its own. I'm using the render the same way an illustrator would use a photograph on a lightbox: as reference for where light falls, where shadows go, how a form reads at this angle. The actual illustration is still painted by hand, over the rendered reference. The 3D is scaffolding, not the finished building.
This workflow genuinely changed how I work. Not in a sudden way, but incrementally. I can now explore a character in ten different lighting scenarios in an afternoon instead of spending three days iterating prompts hoping to get the right result. The fact that the meshes aren't perfect doesn't matter when you're using them as reference, not as the final output.
The Honest Verdict
What these tools can do for a 2D artist right now: give you poseable, lightable approximations of objects and characters for use as compositional and lighting reference. For that use case, they work. The quality gap between "usable for reference" and "publication-ready 3D asset" is still large, but if your goal is reference, the current tools are enough to genuinely change your workflow.
What they can't do yet: generate convincing human faces, faithfully reconstruct culturally specific objects from non-Western traditions, produce production-ready meshes quickly and cheaply, or handle complex organic forms with the fidelity you'd expect from a skilled 3D artist.
If you're a 2D artist curious about whether these tools are worth your time, the answer is yes, conditionally. The condition is that you approach them as reference tools, not final output generators. Once you let go of the expectation that the mesh will be usable directly, and accept that it's scaffolding for your actual work, the tools become genuinely useful pretty quickly.
Start with Meshy AI's free tier and your own drawings. See if the image-to-3D workflow produces anything useful for your practice. If it does, you'll figure out from there which parts of the pipeline to invest in further.
For artists like me, working from a visual tradition that isn't well-represented in these training sets, expect more friction and more workarounds. The tools are worth using anyway. But don't expect them to understand your visual culture the way they understand a mid-century modern chair or a fantasy knight's armor. That gap is real, and nobody in the marketing materials is telling you about it.