Tessera

3D in, game-ready sprite sheets out. No image model anywhere in the path.

Tessera is the fourth host of the ASCII universal kernel (com.astral.ascii.kernel/1), after ASCII, ASCII Studio for Blender and ASCII Studio for Unity. It does not fork that kernel: it copies the payload, verifies the fingerprint at load, and fails loudly rather than running against a kernel that has drifted.

Run it

open ~/Tessera/dist/Tessera.app          # the app
python3 -m tessera studio                # same thing, in a terminal
python3 -m tessera doctor                # check kernel, render host, library
python3 -m tessera library ninja --rigged --animated
python3 -m tessera config

The app binds to 127.0.0.1:8733 only. Logs go to ~/Library/Logs/Tessera.log.

Three editions, one core

Edition	Status	Why it exists
standalone	working	Python 3, standard library only. Runs anywhere the kernel does.
Blender	capture working	The DCC renders with real materials, textures and lighting first.
Unity	planned	Physics- and solver-driven subjects (SlimeShot) that have no baked clip.

Feature work lands once, in the shared layer. An edition never forks it.

What is configuration, not code

Nothing here knows a path on anyone's machine. tessera/config.py layers flag → TESSERA_CONFIG → ~/.config/tessera/config.json → profiles/*.json. profiles/workstation.json describes this workstation — a BlendKit library on a Linux box, Blender over SSH with DRI_PRIME=1. Someone else drops in their own profile and no code changes.

The library index is cached locally so browsing does not shell out per keystroke; asset paths stay remote, because that is where the render happens.

Things learned the hard way, now encoded

• The scene range is not the action range. Ninja Girl's scene says 1..250; her action ends at 75. Samples past the end return the held final pose — duplicate cells, and no error anywhere. Timing reads action.frame_range.
• A cycle's last frame is its first. loop=True drops the endpoint. Keeping it puts a stutter in every loop.
• Track the root, not the bounding box. A bbox centroid moves when a limb extends, so an overhead sword drags the camera and the character appears to sink.
• Lights ride the camera. A world-fixed key lights the front and silhouettes the back, so one character reads as two across a turnaround.
• AgX is a bug here. Blender's default view transform drains saturation by design. The product rule is that colour comes from the model.
• Cell aspect 0.5 is a terminal value. Left alone it squashes every sprite by 1.43x, which reads as a modelling error and gets debugged as one.
• Luminance-driven coverage is wrong for sprites. A red torso is never white so it never picks a solid tile. fill_mode="solid" runs the kernel twice — silhouette for coverage, alpha-masked colour for colour.

Checks are not optional

tessera/qa.py runs on every sheet and each check can FAIL: cells non-empty, coverage consistent, anchor identical in every cell, bbox inside its cell, orbit varies, frames distinct, output not crushed to black. The app shows the result and the runner refuses to deliver a failing sheet. Verified against a known-bad capture, where it fails 4 of 7 — a guard that passes everything is worthless.

Tile sets are data

text-ramp (5x7, 10 glyphs, the canonical kernel) · quadrant-block (4x4, 16 blocks) · pixel-hd (2x2 carrying all 16 possible patterns, so silhouette reproduction is lossless). Default is pixel-hd: this is a sprite generator, and it should not look like ASCII unless someone asks for that.

tests: python3 tests/test_sheet.py && python3 tests/test_timing.py