How-to · Prompts
From AI Concept Art to a 3D Character: Turnarounds and Generation
The chain from AI concept art and turnaround to a rigged 3D character: input limits, rigging constraints and engine checks from documentation, unmeasured parts flagged.
A character is the hardest thing to take from a picture to a game, because every step has a downstream requirement: the input needs usable views, the mesh needs sane topology, and the rig has constraints that some generators state outright. This guide follows the chain from concept art to an engine-ready character using documented limits, and marks the stages where no source gives a number.
Stage 1: the reference image
Many pipelines start with an image model. Meshy’s July 2025 artist story describes using ChatGPT for a first image, then Midjourney with Omni Reference to refine it over several generations, before using a result as the 3D input. That is one artist’s account on a vendor site, not a measured method. What matters downstream is the view set. Meshy’s multi-view tutorial accepts one to four images (main plus Left, Back, Right) and allows a single reference sheet with front, side and back views as the main image. It asks for a plain background, even lighting and at least 1040 x 1040 pixels, and Multi-View needs Pro or above and runs only on Meshy 7. Details are in preparing images for image-to-3D.
Practical consequence: if your image model drifts between views, the inputs may disagree, which a generator may not be able to reconcile. We have not measured how often this happens.
Stage 2: generation
Pick a generator on documented facts: input formats, whether it rigs, plan, and license. The character generator shortlist does that; the character rigging comparison sets out how to check your own, and we have not run it ourselves. Note that rights differ by plan; check the license matrix before you build a cast on a free tier.
Generator settings that bear on a character
Meshy’s Image to 3D API (accessed 2026-11-08) documents several parameters that matter for a character, and they are the only such settings we read from a vendor:
pose_mode:a-pose,t-pose, or empty for no specific pose. If your target rigger or engine wants a rest pose, this is the documented control, though the page does not say how closely the output follows it.image_enhancement(defaulttrue): the page describes it as optimizing the input image before generating, sofalseskips that step. It works only withmeshy-6,meshy-7.1orlatest.remove_lighting(defaulttrue): takes effect only with the Meshy 6 texture model; it returns 400 on unsupported combinations such as Smart Topology.should_remeshandtopology(triangleorquad), withtarget_polycountof 100 to 300,000 for remesh (default 30,000);topology: quadwith Smart Topology returns 400.should_texturemust staytrueif you plan to rig, since the rigging endpoint accepts textured models only.- Image input: JPG, JPEG, PNG or WebP, at most 20 MB, at least 32 px per side, as a public URL or base64 data URI. That floor is the API’s; the web tutorial asks for 1040 x 1040.
Stage 3: topology and size
A rig deforms the mesh, so topology around joints matters, but no source we read gives a quad-count or edge-flow target for characters. Documented limits that do exist:
- Meshy’s auto-rigging API accepts a face limit of 380,000 triangles and its documentation says to remesh to 300,000 or 150,000 quad faces first.
- Roblox’s specifications say individual meshes cannot exceed 20,000 triangles, and a vertex cannot be influenced by more than four bones.
For the method choice see retopology for AI 3D models.
Stage 4: rigging
Documented constraints, accessed 2026-11-08:
| Rigger | Documented constraints |
|---|---|
| Meshy auto-rig API | Textured models only; humanoid (biped) only; face must point +Z; .glb input; height_meters default 1.7; outputs rigged FBX and GLB plus walking and running clips. A web-app Smart Rig for non-humanoids is mentioned in a search snippet only. |
| Tripo rig API | Inputs GLB, FBX, OBJ, STL; output GLB or FBX; rig types include biped and quadruped; spec options differ by model version (v3.0 adds actorcore, unity, unreal, vrm). Rig Check is free; rigging costs 25 credits. |
| Reallusion AccuRIG | Free stand-alone download; exports FBX and USD; input formats and limits not stated on the page we read. |
| Mixamo | Accepts FBX, OBJ and ZIP per a search snippet only; Adobe’s help pages returned 403, so limits and terms are unverified. |
Humanoid-only is a real constraint for creature work; see non-humanoid and quadruped rigs and Mixamo with AI models. Pose matters: Godot’s export guidance says to put bones in rest pose or T-pose before export, apply transforms, triangulate, and use right-handed Y-up. A concept sheet drawn in a dynamic pose may not give a rest-pose mesh; that is a risk, not a measured outcome.
When rigging fails
Meshy’s rigging API documents these failures (the docs give categories, not exact message text):
| Symptom | Documented cause | Fix |
|---|---|---|
| 422, pose estimation failed | The model may not be a valid humanoid, or does not face +Z | Rotate so the face points +Z (the docs’ only stated remedies); a clearer pose is general practice, not from the docs |
| 400, face count exceeded | More than 380,000 triangles | Remesh to 300,000 or fewer, or 150,000 or fewer with topology: quad |
| 400, invalid model format | model_url must be a GLB, up to 400 MB by URL or 100 MB as a data URI |
Convert to GLB |
| Rig rejected as unsupported | The docs list untextured and non-humanoid models as unsupported; they do not name the error code | Texture it; use a different rigger for creatures |
| 402 | Insufficient credits, or free plan | Add credits or upgrade |
| Textures lost | The deprecated texture_image_url replaces the model’s textures and drops PBR maps |
Omit it |
Large files over 100 MB are limited to 10,000,000 faces and no Draco compression. Outputs on success are rigged FBX and GLB plus walking and running clips.
Stage 5: engine check
Import the rigged file into your engine and check scale, orientation, skin weights and materials. Godot documents that models with blend shapes need glTF export with “Export Deformation Bones Only” enabled in Blender’s armature options. No source we read tests an AI-rigged character across engines; that is what the character tests are for.
What this guide cannot tell you
How often the chain succeeds, how many credits it burns, and how long cleanup takes are unmeasured. Cost per usable asset shows how to log them. A rigged character also adds credit lines (Meshy’s API pricing lists rigging at 5 credits while its web app lists 0; Tripo’s rigging page says 25), which sit in the once-per-keeper term, not the retry term. For the broader rigging topic, start at rigging.
See also the AI 3D generation guides hub.