PS1 / PSX low-poly modeling
What the style is
PS1 style = small poly budget + one painted texture atlas. It is not flat shading. The body is smooth-shaded; only hair and hard props stay faceted. Flat-shading the whole model reads as 2005 low-poly, not 1997. Detail lives in the texture, so the mesh is judged on silhouette and edge flow — never on surface detail.
Conventions (measured from real PS1-era reference models, not guessed)
- +Z up, character faces −Y, feet on Z=0, metres.
- One object, one material, one texture atlas. The clothes are the mesh; there are no separate garment objects.
- 100% smooth shading, except hair and hard-surface props.
- No normal maps, no PBR, no shape keys.
- No loose vertices, no loose edges, no zero-area faces, no non-manifold garbage.
- Quads for body shells, triangles only where a shape needs them (tips, caps).
Budgets from three shipped reference models:
| model | verts | tris | notes |
|---|---|---|---|
| rigged male basemesh | 772 | 1540 | 84 bones, 256² atlas, A-pose, 1.98 m |
| Walking (PSXProp) | 349 | 645 | all tris, 512² atlas, T-pose, 1.77 m |
| LE0GAR | 548 | 943 | 9 objects, one 256² atlas, Rigify, 2.4 m |
That is a rigged bare body. Hair volume and accessories legitimately push past it — a 1225 v / 2294 tri character with a full mane, glasses, belt and boots is still a PS1 character. Never cut silhouette to hit a number derived from a bare body.
Setup
scripts/psx_build.py is the library; the two harnesses are config-at-the-top scripts.
scripts/psx_build.py ring / ring_dir / Build / finish / check / report
scripts/preview.py ortho clay renders, shots derived from the bbox
scripts/compare.py stats table + side-by-side against reference FBX/OBJ models
scripts/measure_silhouette.py width-vs-height scan of a reference PNG
Put scripts/ on sys.path and import, or copy a harness next to your build script.
Everything runs inside Blender:
import sys
sys.path.insert(0, r"…/scripts")
from psx_build import Build, ring, finish
The API
ring(cx, cy, cz, rx, ry, n=12, e=2.0, axis="Z", phase=0.0, yoff=None)
One closed cross-section. e is the superellipse exponent: 2 = circle, higher = boxier.
Useful: 2.1 limbs and head, 2.2 legs, 2.3 torso and belt, 2.6 palm. axis="X" lays the
ring in YZ for limbs along X. yoff={index: dy} pushes single vertices — that is how a
bust bulge, lumbar curve or seat is made without adding rings.
Build.loft(rings, cap_start=..., cap_end=..., flat=..., part="…")
Build.tube(path, rx, ry, n=4, e=2.0, caps=True, part="…")
Build.box(cx, cy, cz, sx, sy, sz, flat=True, part="…")
Build.add(verts, faces, flat=..., part="…")
Build.mark() → pass to Build.mirror_x(mark)
tube() takes a path of points and orients each ring along the local tangent — use it for
chains, fingers and anything swept. rx/ry may be lists, one per point, for tapering.
finish(name, *builds, material=..., color=..., uv=True)
check(ob) # the gate: loose verts / edges / zero-area faces must be 0
report(ob) # verts, faces, tris, bbox, height
finish() merges one or more Builds into a single object, applies smooth flags per face,
makes the material, recalculates normals, adds placeholder UVs and prints the build report.
It replaces any existing object of the same name, so a build script is safe to re-run.
Workflow
- Measure the reference before modelling. Run
scripts/measure_silhouette.pyon the turnaround: it prints silhouette width at every 3% of height. Write the table down. Eyeballing proportions is the single biggest source of rounds lost — a waist 2 cm off does not read as "slightly wrong", it reads as a different character. - Block with lofts. One
Build, parts in order: torso → legs → feet → arms → hands → mirror → head → hair → accessories. Read the printed PARTS log after every build; it tells you which part is eating polys. - Render clay, not shaded.
preview.py, grey background, no textures. Judge only silhouette and the shading break where surfaces meet. - Compare against reference models, not just the image.
compare.pyscales everything to one height and prints verts/tris side by side. It catches short arms and flat depth faster than looking. - Gate.
check()clean, feet on Z=0, tri count in budget. - Save a copy.
bpy.ops.wm.save_as_mainfile(filepath=…, copy=True)writes the file without hijacking the session's current filepath.
Traps — each one of these cost at least one round trip
- Separate shells always cross. Two lofts of similar radius produce a jagged seam wherever their surfaces intersect, and no amount of tweaking hides it. Let the outer garment own the joint: the trousers form the hips, the torso tapers to a small cap buried inside them, and the crossover hides under the belt (belt wider than the body). Same trick at arm→shoulder and trouser-hem→boot.
- A cap that pokes out reads as a torn edge. A cap sharing a Z with another shell makes a visible jagged ring. Keep caps strictly inside the other shell, or don't cap.
- Fill the crotch. Two overlapping leg tubes leave a deep valley at x=0. Bridge it with a narrow torso cone whose front is level with the trousers. Too thin and you get a black groove; too wide and you get a diaper.
- Hair: don't cut the face slot out of the shell. For each row, columns inside the slot get a radial scale around 0.45 plus a small +Y push, ramping to full radius roughly 10° past the slot half-angle (26–34° by row). The quad leaving the head then leaves at the cheek instead of across the face. A 30° ramp puts hair over the whole face; shrink without the push leaves a big pale inner wall beside the cheek.
- Never alternate a fringe hem by column parity.
j % 2on an even column count puts a deep tooth notch dead centre. Use a smooth function of the column angle. - Hair is a closed lofted volume, not cards — e.g. 20 columns × 11 rows, flat-shaded, with per-vertex wobble for the chopped look. Sweep the lower rows backwards or the profile reads as a hood.
- Chains float. Keep a necklace arc within a few millimetres of the chest surface or it reads as wire hanging in space in three-quarter views.
- Normalise every direction vector before scaling by a length. An unnormalised one produced a 1 m thumb.
mirror_xtakes(len(v), len(f))frommark(), and mirrored faces need reversed winding. Passing vertex indices corrupts the mesh silently.- Never
join()without deselecting all first. A leftover default cube gets absorbed into the character. Filter selection by object name. - Reference FBX imports need
parent_clear(type="CLEAR_KEEP_TRANSFORM")before the armature is deleted — rigs are commonly 0.01 scale with a 100-scale mesh parented to them. Then apply transforms and centre on the bbox centre, not the bbox min, or the models never line up. - Elevated preview cameras: the horizontal offset must scale by
cos(elev), or the shot frames empty background. - Run the gate on geometry, not on the render. A degenerate quad can look fine in a
front view and wreck the wireframe.
check()exists for that.
Review pass
preview.py gives four full-body views plus two close-ups, all derived from the object's
bbox. Look for: jagged shell crossings, caps showing as flat ledges, the crotch, garment
junctions, whether the hair profile is hair or a hood, whether the face slot is clear.
Final gate:
from psx_build import check
check(ob) # loose verts, loose edges, zero-area faces → all 0
plus: bbox z.min == 0, height, tri count, smooth flags correct.
Worked example — one real character
A 1.744 m gothic girl, T-pose, mesh only, built with this library: 1225 verts / 1079 faces / 2294 tris, 26 n-gons (caps), 4 triangles, everything else quads; feet at 0; one material, one placeholder UV map.
Measured widths from its reference sheet (z → silhouette width in metres), which is what the body tables were built from:
| z | width | z | width | |
|---|---|---|---|---|
| 1.66 | 0.245 | 1.05 | 0.242 | |
| 1.575 | 0.322 | 0.96 | 0.295 | |
| 1.49 | 0.349 | 0.875 | 0.314 | |
| 1.40 | 0.403 | 0.61 | 0.330 | |
| 1.225 | 0.265 | 0.26 | 0.452 | |
| 1.14 | 0.180 | 0.0525 | 0.449 |
The numbers matter less than the method: one table, one pass of lofts, then numeric corrections driven by renders. In that build, arm length was the one thing that needed the reference models rather than the image — the image's arms droop ~20°, so its horizontal span (1.457 m) understates the true arm length. Compare against a T-posed reference model for anything measured at an angle.
Out of scope here
Texturing, rigging and export. Placeholder UVs from smart_project are islands, not a
real atlas. When texturing starts: one 256² or 512² atlas, keep hair's inner surfaces on a
dark region so the shell edges read as depth, and paint the face — the mesh stays a mask.