Test plan · Generators
Low-Poly and Stylized AI 3D: How to Compare Native Modes and Decimation
Native low-poly modes versus decimating a detailed AI mesh: what Meshy, Hunyuan, Rodin and Sloyd document, and how to compare silhouette against a triangle budget.
For a stylized or low-poly asset you have two routes: ask the generator for a low-poly result directly, or generate something detailed and reduce it afterward. Which is better depends on whether the silhouette survives at your triangle budget, and that is a measurement you can make in an afternoon. This page sets out what vendors document for each route and how to compare them, as one of the test protocols. The “afternoon” is an estimate of effort for a handful of prompts, not a measured time.
The two routes
Native low-poly or remesh controls produce a mesh built to a budget. Documented examples: Meshy’s API lists a Remesh operation at 5 credits (the web-app docs list Remesh at 0) and mesh-only generation separately from textured; Hunyuan3D’s hosted API has been listed with a LowPoly generation type and a FaceCount add-on on Tencent Cloud’s billing page, which blocked our last automated check (November 2026), so confirm them and their credit cost there; Rodin’s Gen-2.5 API page defines mesh_mode as Raw (triangles, the Gen-2.5 default) or Quad, with quality presets of 4,000 to 50,000 faces in Quad and 20,000 to 1,000,000 in Raw, and a Smart Low-Poly feature is described in an 80.lv interview rather than in that API page; Sloyd’s homepage says you can set target polygon counts and use quads. These are documented features, not evidence of quality.
Post-decimation takes a detailed output and reduces triangles with a modifier or retopology tool. It costs no credits but costs time and can destroy thin features, text and flat-shaded edges. Methods are covered in retopology for AI 3D models.
One more variable: flat-shaded, hand-painted styles depend on texture and shading as much as geometry, so a low-poly mesh with a photoreal texture can still read as the wrong style.
Documented budget controls in Meshy’s API
Meshy’s Text to 3D API page (read 2026-11-08) gives checkable ranges. With remesh on (should_remesh: true), target_polycount accepts 100 to 300,000 with a default of 30,000, and topology is triangle (default) or quad. decimation_mode takes 1 (ultra) to 4 (low) and, when set, target_polycount is ignored. Smart Topology models are separate: meshy-t2 accepts 100 to 15,000 (default 4,000) and meshy-t2.5 100 to 30,000 (default 15,000), output is triangle-only, and out-of-range values return 400. On this page should_remesh defaults to true for meshy-6-lite and false for other models, and is ignored with Smart Topology; the changelog’s 2026-01-22 entry says true for other models than meshy-6, so set it explicitly. The page shows a banner that lowpoly retires on 2026-10-30. The changelog says model_type: "lowpoly" was deprecated in August 2026, and its 2026-10-30 entry treats model_type: "lowpoly" and ai_model: "meshy-t1" as T1 requests: some now run on Smart Topology, the rest return 400, and from 2026-11-30 every T1 request returns 400. An older “lowpoly” workflow will therefore not keep behaving as written. Meshy’s own advice on that page is “For a low-poly game asset, use a Smart Topology model instead of remeshing a standard model”, which is a vendor recommendation, not evidence. What this means for a budget: the defaults (30,000 for remesh, 4,000 or 15,000 for Smart Topology) are the vendor’s, not yours, so set the number explicitly. As one hard ceiling to compare against, Roblox states “Individual meshes can not exceed 20,000 triangles” (create.roblox.com, modeling specifications), which a 30,000 default would break. Whether the reduced meshes keep their silhouette is exactly what this page asks you to measure.
What is not available
We could not read Tripo’s low-poly feature page (it returned HTTP 403), so nothing about its polygon ranges or seam claims appears here. No source we read compares a native low-poly mode with the same asset decimated afterward. Vendor feature pages describe controls; none we read reports silhouette loss.
Protocol
- Choose a triangle budget per prompt before generating, for example from the polycount budget calculator for your platform, and record it.
- For S01 to S05 in the standard set, generate each of: (a) the tool’s native low-poly or remesh mode at your budget, (b) default output decimated to the same budget with the same decimation method for every tool.
- Render all outputs from front, side and a thumbnail at roughly 64 pixels tall, flat-shaded, no texture.
- Score silhouette 0 to 2 (anchors on the prompt set page), and record triangles, loose_pieces and open_edges.
- Compare texture style separately: flat color, hand-painted look, or photographic.
Fields for this page: route (native/decimated), budget_triangles, achieved_triangles, silhouette_score, thin_feature_lost (yes/no, which), texture_style_match (0 to 2), credits_used, minutes_spent.
Choosing the decimation method
Use one method for every tool so the route comparison is fair. Blender’s Decimate modifier has three modes: Collapse (a Ratio from 1.0, unchanged, to 0.0, all faces removed, with an optional Triangulate; with it off, untouched quads and n-gons can make the final count higher than the ratio suggests), Un-Subdivide (Iterations; even numbers are usually best) and Planar (Angle Limit, with delimits for Normal, Material, Seam, Sharp and UVs). Record the mode and ratio. The manual gives only broad guidance (Un-Subdivide is intended for mainly grid-based topology, Planar for mainly flat surfaces) and says nothing about AI meshes, so if the first mode fails on an asset, try a second on that asset and record both. Ratio does not promise an exact count: read the resulting triangle count and adjust.
Ambiguous results
Rules to set before you start, for borderline readings; they are editorial, not outcomes we have seen.
- Triangles slightly over budget (say within 5%, a tolerance we suggest, not a vendor figure): fail, adjust the ratio and rerun; do not round.
- Silhouette readable at 64 pixels but a thin feature merged: that is a fail for the thin-feature field and a pass for silhouette. Report both.
- Native mode ignored your target: log achieved_triangles as returned and note it; a mode that overshoots the budget fails on budget, whatever it looks like.
- Texture and geometry disagree on style: score texture_style_match separately; do not let a good mesh pass a mismatched look.
- Quad output: count triangles after triangulating (a quad is two), or the budget comparison between routes is not like for like.
Pass or fail
Pass means achieved triangles are at or under budget, the 64-pixel thumbnail is still readable as the prompted subject, and no thin feature (sword blade, antlers, fence rails) has been lost or merged. A route wins for a prompt only if it passes; if both pass, the cheaper in credits plus minutes wins. If neither passes, the budget may be too tight for the subject, which is also a result.
How this feeds your pipeline
If decimation works for your style, the cost of the route is time and the LOD workflow is the same. If native modes win, check that their output also passes the raw mesh checks, since a low triangle count says nothing about holes. For a shortlist of generators by job, see best AI 3D generator for game assets.