AI 3D generation
Making meshes, textures, rigs and motion with generative AI as of September 2026: what each kind of tool does, the landscape (Higgsfield, Meshy, Tripo, Rodin, Hunyuan3D, TRELLIS.2, Stability), image models for concept art and textures, prompt templates, what comes out broken and the Blender cleanup that fixes it, and licensing. This space changes monthly: vendor claims are attributed and dated, so check pricing and terms before you rely on them. Terms are defined in Core concepts; driving Blender with an LLM is Claude & GPT with Blender.
What "AI 3D" means
| Kind | In → out | Quality today | Use it for |
|---|---|---|---|
| Text → 3D | prompt → textured mesh | weakest control; the model invents the design | quick props, ideas |
| Image → 3D | one picture → textured mesh | best value: you control the design in 2D first | props, characters, statues |
| Multi-view → 3D | 2–4 views of one subject → mesh | better geometry on the hidden sides | characters, vehicles |
| AI texturing | your mesh + prompt/image → PBR textures (defined below) | good on props; seams and baked light still happen | re-skinning a clean mesh |
| Auto-retopology / remesh | dense mesh → lighter quad or triangle mesh | fine for props, rarely animation-grade edge flow | budget cuts |
| Auto-rigging | humanoid mesh → mesh + skeleton + weights | good on clear T/A-pose humanoids; fails on odd shapes | NPCs, prototypes |
| Text → motion | prompt → animation clip (FBX/BVH) | usable for generic actions (walk, wave, idle) | background characters |
| Video → mocap | phone video → animation clip | decent body motion; feet slide, hands are rough | reference, blocking |
| Scene agents | prompt → a whole editable Blender scene | blockouts and layout, not final art | staging, previz |
Mesh (vertices, edges and faces with UVs and materials) is what games and three.js render. A Gaussian splat is a cloud of millions of soft colored blobs, great for photoreal captures and viewing, but it has no surfaces to collide with, no UVs, can't be rigged or relit and isn't supported by Godot or GTA V. Some generators (Meta's SAM 3D Objects, for one) can output either; for games, always take the mesh.
| Mesh | Gaussian splat | |
|---|---|---|
| Engines | everything (glTF, FBX) | three.js via add-on loaders, dedicated viewers |
| Physics, collision | yes | no |
| Rigging, animation | yes | no (research only) |
| Editing in Blender | full | view-only add-ons |
| Best at | games, interactive web | photoreal captures, walkthroughs |
Tool landscape
As of September 2026; features and names are the vendors' own. PBR (physically based rendering) means the texture set a game engine expects: base color, roughness, metallic, normal. Topology is how the faces are laid out: "triangle soup" is a dense, irregular triangle mesh; "quads" are four-sided faces in clean loops.
| Tool | Inputs | Outputs | Topology | PBR | Rig / motion | Runs |
|---|---|---|---|---|---|---|
| Higgsfield (opens in a new tab) | image, 2–4 images | GLB | from the engine (its Blender add-on lists Meshy 5) | optional | rigging + a clip library | hosted, credits |
| Meshy (opens in a new tab) | text, image, multi-image | GLB, FBX, OBJ, USDZ, STL, .blend | quad or triangle remesh, target face count | yes | humanoid auto-rig, presets, text-to-motion | hosted, API |
| Tripo (opens in a new tab) | text, image, multi-image | GLB, FBX, OBJ and more | "Smart Mesh" quads (vendor claim) | yes | auto-rig, animations | hosted, API |
| Rodin / Hyper3D (opens in a new tab) | text, image, multi-image | GLB, FBX and more | triangle or quad | yes (vendor: up to 4K) | — | hosted, API |
| Hunyuan3D (opens in a new tab) | image (2.1 open); text, image (3.x hosted) | GLB, OBJ | dense triangles | 2.1: yes | hosted 3.x lists rigging | 2.1 local (≈ 29 GB VRAM), 3.x hosted |
| TRELLIS.2 (opens in a new tab) | image | GLB | dense triangles | base color, roughness, metallic, opacity | no | local, Linux, 24 GB+ NVIDIA |
| Stable Fast 3D (opens in a new tab) / SPAR3D (opens in a new tab) | image | GLB, UV-unwrapped | low-poly, optional quad remesh | roughness, metallic | no | local, ≈ 6–10 GB VRAM, Mac (MPS) |
| SAM 3D Objects (opens in a new tab) | image + object mask | mesh or splat | dense | color texture | — | local weights, playground |
| Tool | License / commercial use (brief; read the terms) | Good for |
|---|---|---|
| Higgsfield | check your plan's terms | one place for concept image → 3D → scene, also from Claude |
| Meshy | free: CC BY 4.0 (credit Meshy); paid: you own the assets | all-rounder with the most pipeline tools |
| Tripo | free: public; its pricing page says CC BY 4.0 while its terms reserve broad rights; paid: private, commercial | stylized props, quads |
| Rodin | pay when you download; paid tiers "any use" | high-detail hero props |
| Hunyuan3D 2.1 | Tencent community license: not valid in the EU, UK or South Korea; separate license over 1M monthly users | local experiments outside those regions |
| TRELLIS.2 | MIT (code and weights) | self-hosted batches |
| SF3D / SPAR3D | Stability AI Community License: free under USD 1M annual revenue, register, credit | fast, light, runs on a Mac |
| SAM 3D Objects | SAM License (research and commercial, with use restrictions) | objects cut out of real photos |
Higgsfield
Higgsfield is a hosted creative suite (image, video, audio, 3D) sold on credits. As of September 2026 it reaches 3D in four ways; all four were checked on higgsfield.ai, not tested end to end.
| Surface | What it does | Output |
|---|---|---|
Image → 3D (web, API, MCP generate_3d) | picture, or 2–4 views, to a textured mesh; optional rigging and a library of about 680 animation clips | GLB |
| Blender add-on (plugins/blender (opens in a new tab)) | floating bar in the viewport (Blender 5.1+) with Scene Builder, 3D Model, Character Animation, Image, Video, Camera and Asset tabs; results land in your open .blend | editable objects, image planes |
Higgsfield MCP connector (https://mcp.higgsfield.ai/mcp) | the same generators as tools for Claude, ChatGPT or Cursor | job ids, downloads |
3D scene builder (MCP scene_builder_3d_*, the "3D Jutsu" workspace) | the model writes Blender Python that runs on Higgsfield's hosted Blender 5.2 worker, imports GLBs from a curated catalog and renders previews | .blend, .glb, PNG/MP4 |
- The connector's own tool notes say the mesh reproduces only what is in the source image: to add a prop or
change clothing, edit the image first (
generate_image), then convert. - Multi-view input (
multi_image_to_3d) gives better hidden sides than a single image; there's alsosam_3_3dto lift one object out of a busy photo, and3d_riggingfor an existing GLB. - The add-on also has a Blender Bridge MCP URL (
https://bridge.higgsfield.ai/mcp) so an assistant can act on your open scene; setup is in Claude & GPT with Blender. - Image and video models on Higgsfield (its pages list GPT Image, Nano Banana, Seedream, Soul; Kling and Seedance for video) are useful for the concept art, turnaround sheets and texture sources below, and video for motion reference.
- Costs show before you generate (on the add-on's Generate button;
get_coston the MCP tools).
Image models for inputs
Image-to-3D quality is decided by the input image. Make it with an image model (GPT Image in ChatGPT or the
API, where the current models are gpt-image-2.5-* as of September 2026; Higgsfield's catalog; any other):
| Input | Why | Caveats |
|---|---|---|
| Concept art | pick a design before spending 3D credits | pretty is not the same as buildable; check the silhouette |
| Orthographic turnaround (front, side, back) | multi-view → 3D, and reference images for modeling by hand | views drift apart; regenerate until they agree |
| Character in A-pose or T-pose | auto-riggers need spread arms and legs | hands and feet are often mangled |
| Seamless tileable texture | walls, floors, ground | "seamless" often isn't: tile it 2×2 and fix seams |
| Trim sheet | strips of molding, panels, bolts for architecture | spacing is inexact; square the strips up in an image editor |
| Decals | stains, signs, graffiti on alpha | ask for a plain background, then remove it (no true alpha from some models) |
| Sky backdrop | background plate | 8-bit, not HDR, rarely a correct 360° panorama: light with a real HDRI (Poly Haven) |
An image model gives you one texture: albedo (base color). Derive normal and roughness maps in a material tool, or ask for them separately and check they line up. See Materials & UVs for wiring them.
Prompt templates
Replace the brackets. These work in GPT Image, Higgsfield's image models and most others.
CONCEPT (single object for image-to-3D)
A [stylized low-poly | realistic | hand-painted] [object], one object only,
centered, whole object in frame with margin, three-quarter view from slightly
above, plain light-gray background, soft even studio lighting, no cast
shadows, no text, no props, clear silhouette, solid (not transparent) materials.TURNAROUND SHEET (props, vehicles, buildings)
Orthographic reference sheet of [object]: front, left side, back and top views
in a row, same scale, aligned on a common ground line, no perspective,
flat neutral lighting, plain white background, thin gaps between views,
no labels, identical design in every view.CHARACTER TURNAROUND
Character reference sheet of [description], full body, A-pose (arms 45° down,
legs slightly apart), front, side and back views side by side, same height,
orthographic, neutral expression, flat even lighting, plain background,
fingers separated, no weapon or loose props, no motion blur, no cropping.SEAMLESS TILEABLE TEXTURE
Seamless tileable [red brick wall | mossy cobblestone | oak planks] texture,
top-down orthographic, fills the whole square frame, even diffuse lighting,
no shadows, no highlights, no vignette, no perspective, [2 m] of real-world
surface per tile, consistent scale, 1:1 aspect.TRIM SHEET
Game trim sheet texture, square, horizontal strips stacked edge to edge:
[crown molding, plain concrete band, metal panel with rivets, window frame,
brick soldier course], each strip full width, flat front view, even lighting,
no shadows, no gaps, no text.DECAL
Single [rust stain | peeling poster | graffiti tag] decal, centered on a plain
pure-white background, front view, flat lighting, soft natural edges, no
frame, no shadow, isolated.SKY BACKDROP (background only, not lighting)
Wide 360-degree equirectangular panorama of [overcast sky over low hills],
horizon exactly at the vertical middle, no sun disk, no buildings, soft gradient.Prompting for 3D
| Do | Why |
|---|---|
| One object, centered, fully in frame | cropped parts become holes or stubs |
| Plain background, flat light, no cast shadows | shadows get baked into the texture (see below) |
| Simple, readable silhouette | the generator guesses the unseen sides from the outline |
| Avoid thin parts (antennae, straps, wires, hair strands) | they come out blobby or detached |
| Avoid glass, transparency, mirrors | reconstructed as solid, oddly colored shells |
| Name the style: low-poly, stylized, realistic | sets both geometry and texture detail |
Set a face budget where the tool allows (Meshy target_polycount, Tripo face limit, SF3D remesh) | saves a decimation pass |
| Feed 2–4 matching views when supported | fixes invented backs |
| Ask for quad or low-poly output only if you will rig or edit it | quads cost detail |
| Generate 3–4 variants, keep the best | results vary a lot between seeds |
What comes out broken
| Problem | Symptom | Fix in Blender |
|---|---|---|
| Dense triangle soup | 100k–1M triangles, no edge loops | Decimate modifier for props; retopology for anything you deform or edit (Modeling) |
| Baked-in lighting | shadows and highlights painted into the albedo; looks wrong under your lights | delight it (paint out, or regenerate with flat light); vendors' "delighting" options help |
| No, or messy, UVs | hundreds of tiny UV islands; texture can't be edited | Smart UV Project or seams + unwrap, then bake the old texture onto the new UVs (Materials & UVs) |
| Wrong scale, rotation, origin | 40 m tall mug, lying on its side, pivot in space | set real size, apply transforms, origin to the bottom center (script below) |
| Non-manifold, holes | mesh isn't watertight; bad for 3D printing, physics, booleans | Select All by Trait → Non Manifold, Fill, Merge by Distance |
| Floaters | small detached junk pieces | Select Linked, delete loose parts; separate by loose parts to find them |
| Symmetry drift | left and right differ | cut in half, Mirror modifier, re-bake |
| Mushy details | text, faces, hands, mechanical parts blurred | model those parts by hand, or use a normal map |
| Bad rigs | wrong joints, weights bleeding (arm moves the torso) | re-rig, fix weights (Animation); auto-rigs expect clean humanoids |
| Too many materials | several materials per small prop = several draw calls | join, bake to one material and one texture set |
Cleanup pipeline
generate (image → 3D, 3–4 variants) pick the one with the best silhouette
│
import GLB into Blender 5.x File > Import > glTF 2.0
│
scale, orient, origin real meters, Z up, feet on 0, apply transforms
│
clean merge by distance, delete floaters, fill holes
│
retopo or decimate props: Decimate; characters: retopology
│
UV unwrap new, clean UVs on the low-poly mesh
│
bake old texture + high-poly detail → new maps
│
material check one Principled BSDF, image textures only
│
rig (optional) vendor auto-rig, Mixamo, or Rigify
│
export GLB for web and Godot, Sollumz for GTA VA first-pass script for the "scale, orient, origin" and "decimate" steps: run it in the Text Editor with the imported mesh active (bpy is covered in Python scripting).
import bpy
import bmesh
from mathutils import Matrix, Vector
def tris(ob: bpy.types.Object) -> int:
return sum(len(p.vertices) - 2 for p in ob.data.polygons)
def normalize(ob: bpy.types.Object, height: float) -> None:
"""Bake transforms into the mesh, scale it to `height`
meters, origin at bottom center, placed at 0,0,0."""
me = ob.data
mw = ob.matrix_world.copy()
ob.parent = None # glTF imports may add a parent
me.transform(mw)
ob.matrix_world = Matrix.Identity(4)
lo = Vector([min(v.co[i] for v in me.vertices)
for i in range(3)])
hi = Vector([max(v.co[i] for v in me.vertices)
for i in range(3)])
k = height / (hi.z - lo.z)
base = Vector(((lo.x + hi.x) / 2, (lo.y + hi.y) / 2,
lo.z))
me.transform(Matrix.Scale(k, 4) @
Matrix.Translation(-base))
bm = bmesh.new()
bm.from_mesh(me)
bmesh.ops.remove_doubles(bm, verts=bm.verts,
dist=0.0001) # weld seams
bm.to_mesh(me)
bm.free()
def decimate(ob: bpy.types.Object, target: int) -> None:
"""Non-destructive Decimate modifier to ~target tris."""
mod = ob.modifiers.new("Decimate", "DECIMATE")
mod.ratio = min(1.0, target / max(tris(ob), 1))
ob = bpy.context.active_object
normalize(ob, height=1.0) # this prop is 1 m tall
decimate(ob, target=5000)Budgets depend on the target: see three.js export, Godot export and GTA V export.
AI or by hand
| Asset | Use AI for | Do by hand | Verdict |
|---|---|---|---|
| Background props (crates, rocks, bins, plants) | the mesh and texture | decimate, one material | AI wins |
| Hero props (the sword, the product on a web page) | concept, first mesh, texture ideas | retopo, clean UVs, bake, details | AI draft, human finish |
| Characters | concept, turnaround, first mesh, auto-rig for prototypes | retopology for deformation, face, hands, weights | AI for NPCs; heroes by hand |
| Buildings, interiors | facade concepts, textures, trim sheets, decals, scripted layouts via an LLM | blockout, modular kits on a grid, procedural nodes (Buildings & scenes) | AI is weak at straight walls, right angles and openings |
| Terrain | height-map ideas, ground textures | sculpt, geometry nodes, displacement | mostly procedural |
| Vehicles | concept, reference sheets | hard-surface modeling, separate wheels and doors | by hand (parts must move) |
| Animation | text-to-motion for generic loops, video mocap for blocking | keyframing, cleanup of foot slide, anything acted | AI for crowds |
Image-to-3D output is one fused lump: doors don't open, wheels don't turn, and a building has no interior. If a part must move or be entered, model it or cut it out and rebuild it.
Rigging and animation
| Step | AI options (September 2026) | Watch for |
|---|---|---|
| Auto-rig | Meshy Rigging (textured humanoids, face toward +Z per its docs, ≤ 300k faces), Tripo, Higgsfield 3d_rigging, Mixamo | non-humanoids fail; check shoulders, hips, fingers |
| Preset clips | Meshy animation library, Higgsfield's clip library, Mixamo | skeleton names differ between vendors |
| Text → motion | Meshy Text to Motion (2–10 s clips, FBX or BVH), DeepMotion SayMotion | generic motion; retarget onto your rig |
| Video → mocap | DeepMotion Animate 3D, Rokoko Vision (FBX; free tier is short) | foot sliding, jitter, hand errors |
| Retarget and clean up | Blender NLA and Graph Editor (Animation) | scale 0.01 and mixamorig: bone names on Mixamo FBX |
Mixamo is still free with an Adobe ID as of September 2026 but isn't actively supported; the Mixamo Rig (opens in a new tab) extension adds a control rig in Blender. Rigging concepts: Animation & rigging.
Licensing and ethics
| Issue | What to know |
|---|---|
| Free tiers | often public and CC BY 4.0 (you must credit) or with rights kept by the vendor; paid tiers usually give private assets and commercial use |
| Open weights ≠ open use | Hunyuan3D 2.1's license excludes the EU, UK and South Korea; Stability's is free only under USD 1M revenue; TRELLIS.2 is MIT |
| Copyright of outputs | purely AI-generated work may not be protectable in some countries (the US Copyright Office requires human authorship); your edits can be |
| Other people's IP | don't upload characters, logos or photos you don't own to make "their" model; a generator's output can still infringe |
| Steam | the Content Survey asks you to disclose AI content that ships and is seen by players (pre-generated) and content generated while the game runs (live, with guardrails); AI used only as an efficiency tool is out of scope |
| Records | keep prompts, source images, tool, plan and date per asset, in case a store, client or licensor asks |
| People | it's your call, but say so when AI made the art, and don't pass it off as someone's hand work |
Recipes
Prop from a concept image with Higgsfield
Use for a single web or game prop in under an hour.
1. Image model: CONCEPT template → 3–4 options → pick one; regenerate a
side and back view if the tool takes multi-view input.
2. Higgsfield image → 3D (web, add-on "3D Model" tab, or ask Claude with the
Higgsfield connector: "turn this image into a 3D model"). Download the GLB.
3. Blender: File > Import > glTF 2.0. Run normalize() + decimate() above
(target 5–20k tris for web).
4. Check the albedo for baked shadows; paint them out or regenerate.
5. If you will edit the texture: Smart UV Project on a copy, bake the
original texture across ("Selected to Active").
6. Export GLB → optimize and load: /3d/export/web-threejs/.Character for Godot: turnaround → Meshy or Tripo → Mixamo
Use for NPCs and prototypes, not a hero character.
1. CHARACTER TURNAROUND template (A-pose, fingers apart, no props).
2. Meshy or Tripo multi-image → 3D, humanoid, quad or low-poly output,
~10–30k faces.
3. Blender: normalize(height=1.75), delete floaters, Mirror modifier if the
halves drifted, one material, check the face and hands.
4. Rig: vendor auto-rig, or export FBX → Mixamo auto-rigger → download with
skin; add clips "without skin".
5. Blender: import the FBX, fix 0.01 scale, rename or keep mixamorig: bones,
push each clip to its own NLA track, name them (Idle, Walk, Run).
6. Export GLB; in Godot the clips arrive in an AnimationPlayer:
/3d/export/godot/.Tileable brick texture with an image model
Use for walls, floors and ground in Blender, three.js or Godot.
1. SEAMLESS TILEABLE TEXTURE template, "red brick wall, 2 m per tile".
2. Check tiling: place the image 2×2 (any image editor, or a Blender plane with
the UVs scaled by 2). Fix seams by offsetting 50% and painting the cross.
3. Make normal and roughness maps from it (Materials & UVs covers the nodes),
or generate them and confirm the bricks line up.
4. Principled BSDF: Base Color (sRGB), Roughness and Normal (Non-Color).
5. UV the wall to real size so one tile covers 2 m; bake or export as GLB.A building: blockout plus AI surfaces
Use when image-to-3D gives you a melted house.
1. TURNAROUND SHEET of the facade for reference; generate trim sheet and
decals with the templates above.
2. Block out walls, floors, openings on a 1 m grid by hand, from a modular
kit, or with Claude + MCP (/3d/ai-workflows/llm-blender/).
3. Map the trim sheet onto moldings and window frames; tileable textures on
walls; decals for dirt and signs.
4. Use image-to-3D only for small props that dress it (planters, lamps).
5. Export per target; collision meshes stay simple boxes.A shelf of background props, locally
Use when you need 30 rocks, crates or bins and don't want to pay per model.
1. Image model: one CONCEPT image per prop, same style words in every prompt.
2. Run Stable Fast 3D (≈ 6 GB VRAM, runs on Apple Silicon) or TRELLIS.2
(24 GB NVIDIA) on the folder; mind each model's license.
3. Blender: import all, run normalize() + decimate(target=2000) per object,
join each prop into one material.
4. Batch-export one GLB per object (script in /3d/ai-workflows/llm-blender/).An NPC wave with text-to-motion
Use for a generic gesture you don't want to keyframe.
1. Rig the character (auto-rig or Mixamo) and export it.
2. Meshy Text to Motion: "a character waving with the right hand", 3 s;
or pick a clip from Higgsfield's animation library.
3. Apply it to your rig in the vendor tool, or import the FBX/BVH into Blender
and retarget onto your armature.
4. Clean foot slide and the first/last frames so it loops; export.References
- Higgsfield for Blender (opens in a new tab) and its blog post (opens in a new tab): add-on tabs, install, Bridge MCP URL
- Higgsfield MCP (opens in a new tab), Scene Builder (opens in a new tab), 3D Jutsu announcement (opens in a new tab): connector, hosted scene building
- Meshy docs (opens in a new tab): image, multi-image and text to 3D, Remesh, Rigging, Animation, Text to Motion; Meshy pricing (opens in a new tab) for license by plan
- Tripo (opens in a new tab) and its terms (opens in a new tab); Hyper3D Rodin pricing (opens in a new tab) and API docs (opens in a new tab)
- Hunyuan3D 2.1 (opens in a new tab) with its license text (opens in a new tab)
- TRELLIS.2 (opens in a new tab), Stable Fast 3D (opens in a new tab), SPAR3D (opens in a new tab), Stability AI Community License (opens in a new tab)
- Meta SAM 3D (opens in a new tab): single-image objects, mesh or splat
- Blender Manual: glTF 2.0 (opens in a new tab): import and export settings
- Steamworks Content Survey (opens in a new tab): the AI disclosure section
- US Copyright Office: Copyright and AI (opens in a new tab): human authorship reports
- Mixamo (opens in a new tab), Mixamo Rig extension (opens in a new tab), DeepMotion (opens in a new tab), Rokoko Vision (opens in a new tab)
- Poly Haven (opens in a new tab): CC0 HDRIs, textures and models for real lighting and filler assets