TRELLIS 2 turns one image or a handful of views into a textured mesh in minutes on a local GPU. The question a studio actually asks is different: how many artist hours does that raw output cost before it survives a shot review, and which shots are worth paying for.
Almost every TRELLIS 2 demo stops in the same place: a turntable of a textured mesh that looks convincing for eight seconds. A shot stops when a lead signs off on a version, and between those two points sit scale, topology, UVs, maps and a comp artist who needs a matte that lines up.
What follows is editorial judgement from a small-studio point of view, not a benchmark. Where a number appears it is a typical band we would quote in a bid, not a measurement; re-derive it on your own hardware and shot list. It covers what the model returns, which assets survive without a cleanup pass, how to estimate that pass in hours, and where a desktop tool stays the right answer.
The TRELLIS family builds a structured 3D latent from images and decodes it into a mesh with appearance attached; TRELLIS 2 is the successor current community nodes wrap. In pipeline terms, geometry and material come out of one pass: a triangulated surface plus PBR maps, not a bare shape waiting for a texturing run. What you do not get is anything a pipeline calls finished: no edge loops, no UV layout anyone chose, no pivot in a useful place, no scale reference. ComfyUI exports GLB or OBJ plus map images through its usual save nodes, and checkpoint names and node behaviour move quickly, so pin the version you validated.
One image is enough for the assets studios already model quickly: matte hard-surface props, crates, tools, furniture and chunky stylised shapes at one to two metres of scale, shot flat under even light. The reconstruction holds because nothing in frame depends on what the camera could not see; rotate the object slightly and the illusion thins. Multiview input is the bigger practical step: four to six consistent views, ideally front, both sides, back and two three-quarter angles, lit the same way and without baked shadows, constrain the back and contact areas instead of leaving the model to guess. In a ComfyUI graph those views are extra conditioning inputs, so the cost is shot preparation, not compute.
Screenshots argue about a model; a bid needs hours. Convert defects into time with three multipliers: the polycount the shot can carry, whether simulation touches the asset, and how close the camera gets. A 200k-triangle surface that never fills more than a few percent of frame is a decimate and a normal check; the same surface as a hero prop in a 4K close-up is a retopology session, a UV pass and a texture projection.
| Shot type | Typical raw state | Work required | Cleanup band (hours) |
|---|---|---|---|
| Background dressing, under 5% of frame | Dense triangles, uneven normals | Weld, decimate, fix normals, re-export | 0.5–2 |
| Hero hard-surface prop, close-up | Soft silhouette, fused detail, rough atlas | Retopology, UV rework, texture projection | 3–8 |
| Character that deforms | No deformation-friendly edge flow | Retopology, rigging, weights, maps | 20–60 |
| Anything destroyed or simulated | Non-manifold edges, self-intersections | Voxel remesh, watertight pass, collider cleanup | 2–6 |
| Set extension filler seen at distance | Low detail, uneven texel density | Texture fill, LOD build | 0.5–1.5 |
Those bands are editorial judgement from small-studio bids, not measurements; they shift with the brief, the delivery format and how much of the asset the story sees. The comparison that matters is against the alternative: if a clean plate and two days of modelling beat a generator plus six hours of retopology, the generator loses.
Two things break handoffs. Scale and axis: generated meshes arrive at an arbitrary size, often normalised to roughly a unit cube, Y-up or Z-up depending on the exporter. Set scale against a known reference and confirm the axis convention your renderer expects. Pivots sit wherever the mesh origin happens to be, so move them to the base centre while the asset is one object. Materials are the other half: you get albedo, roughness, metallic, normal and ambient occlusion maps, but the shader graph in Arnold, Karma or Redshift is yours to rebuild, colour-space flags included. Bring it into layout as FBX or USD and archive the source mesh and maps.
Standing caveat: engine availability, checkpoint versions, credit pricing and third-party licence terms change without notice, and a node package that works today can break on the next ComfyUI update. Verify on the vendor or repository page before building a bid on it.
A renderer is forgiving: non-manifold edges, an open bottom and a few self-intersections stay invisible in a beauty render if the normals point outwards, and turning a patch into triangles costs nothing at distance. A solver is not: destruction, cloth and volume containers want a closed, consistently wound volume, and open interiors produce jitter, blow-ups and floating fragments that cost more than the cleanup you skipped. Printing is a third test: an FDM wall wants roughly 0.8–1.2 mm with a 0.4 mm nozzle, resin parts 0.6–1.0 mm, and overhangs past about 45 degrees need support. Decide which test the asset must pass before buying cleanup.
The expensive decision is not which generator wins in general, it is whether this asset is worth cleaning. This deployment of Open3DAI runs one local procedural generation engine and can connect optional third-party engine adapters with your own API keys, so the same prompt or reference image can be compared side by side before you commit to a cleanup pass. It also covers text-to-3D, image-to-3D, multiview input and AI texturing, and exports GLB, glTF, OBJ, STL, PLY and 3MF, with 48 format converters, 12 browser tools and a REST API. The boundary matters as much: no auto-retopology, UV unwrapping, rigging, animation or photogrammetry, no parametric CAD solids, and no promise of watertight or print-ready output without cleanup. Those stay desktop work.
| Stage | Tool class | What it decides |
|---|---|---|
| Concept and blockout | Generator, local or hosted | Whether the shape reads at all |
| Cleanup and retopology | Desktop retopo tool or DCC | Whether the mesh survives close-ups |
| UVs and maps | Substance Painter, Mari or a DCC | Texel density and seam placement |
| Simulation | Houdini or a DCC solver | Whether the volume can solve |
| Comp | Nuke or equivalent | Whether mattes and AOVs line up |
Against these criteria, TRELLIS 2 is production-usable for background dressing, set extension filler, previz and look-dev plates, where a decimate and a normal check are the whole cleanup; for hero props, characters and anything simulated it is a starting point, not a finish. The honest summary is that image-to-3D generation compresses the modelling stage and leaves the finishing stage where it was: someone still has to make the mesh move and hold up in close-up. A week is enough to test that claim: take five shots you have already finished, regenerate their assets, and log real hours through cleanup, handoff and one comp iteration. Compare that number with what the shots actually cost; it is worth more than any demo reel, this review included.
For background dressing, set extension filler and previz, yes: a decimate and a normal check are usually the whole cleanup. For hero props, characters and anything that deforms or breaks, treat the output as a starting point and budget the artist hours from the cleanup table above.
Typical editorial bands run from about half an hour for a distant filler asset to several days for a deforming character. The three multipliers are the polycount the shot can carry, whether simulation touches the asset, and how close the camera gets.
Not usually. Renderers tolerate non-manifold edges and an open bottom when the normals are consistent. Solvers do not: destruction, cloth and volume work need a closed volume, which normally means a voxel remesh or a boolean union before the simulation is built.
Yes, by splitting the work. Run generation on a hosted endpoint and keep the desktop machine for cleanup, or accept lower resolution settings locally. On 12–16 GB of VRAM, expect single-asset batches and more silhouette error to fix afterwards; high-resolution settings and batches generally want a 24 GB card or more.
Send the same prompt or image to several engines, compare the meshes side by side, then export GLB, glTF, OBJ, STL, PLY or 3MF. Creating an account grants free credits and needs no card.
TRELLIS 2 is a 4B-parameter image-to-3D model, and its ComfyUI wrapper depends on compiled CUDA extensions that must match your Python, torch and CUDA versions. This hub covers what it produces, what a machine needs, and how both workflow builds are assembled.
A head-to-head of the three open 3D generators people actually run locally: architecture, mesh fidelity, VRAM and time budgets, licence terms and ComfyUI node support, plus a decision framework you can apply to your own images.
Hunyuan3D 2.x is a shape diffusion model plus a separate texturing stage, and ComfyUI can drive both. Here is how the graph is wired, what each route costs in VRAM, and what the usual error messages mean.