← Blog
✎blogbericht

Dogs, clothes, cars and what is next

A cinematic photograph of a stray dog sitting beside a parked old hatchback car on an empty city street at dusk, in teal and steel with a warm amber light, the AMARBARO mark and wordmark centred over it.

Art took more redoing than any system, and a TRELLIS settings A/B cut prop time from 231 to 102 seconds.

The dog was accepted on the third try. Art had to be redone more than anything else in the game, because systems are cheap to describe and slow to make look right. This last article covers what people see and drive, the tooling that makes it, and what is left to do.

People: thin bodies, no clothes, then clothes

On 30 October the team put people on a shared skeleton using a free character pack. The bodies were superheroes in underwear only. The designer approved two thin bodies, one male and one female, on the real idle pose. Painted clothes and clothing shells were rejected and removed. A rule followed: do not add clothes or looks unasked.

Clothes came back on 2 November through Blender. Of 40 clothing items, 30 became real garments. Hats and glasses stayed on the older method. We created 38 garments by duplicating the character mesh, extruding outward along surface normals, and adding details like collars and seams. We did not use cloth simulation. A simulation of 38 shapes would not have been repeatable, and the lift plus a clearance pass was enough.

The measured cost was close to neutral:

Old garments Blender garments
Runtime mesh size 2.23 MB 1.54 MB
Draw calls, 32 characters 262 278
Triangles, 32 characters 484k 461k
Frame time, p50 1.39 ms 1.41 ms

Clipping was measured too, by posing the clothes through animation clips. The worst case was 5.3 cm, loose long tops in a deep crouch. Trousers at the knee clipped 1.5 to 2.9 cm.

Visual flaws remain: coat hems look ragged while sprinting, hoods cannot be worn up, and open jackets show a gap across the chest.

Cars that finally drive

Six new cars (sedan, hatch, estate, van, pickup, bus) plus a wheel were modelled in Blender. Triangle counts are 820 to 1,004 per car, and each file is 50 to 62 KB.

A surprise: the "driven car" had never been a vehicle box. The game moved the parked car itself, so it already had the model. What it lacked was rolling and steering wheels, collision with buildings, a prompt, a camera and a driver who did not poke through the roof.

The paint and model come from a hash of the car's id and position, so every player agrees and the baked map did not change. Collision uses a swept box cast, so a car stops on contact.

The automated driving test simulates WASD keypresses, hotwires a van, and steers directly into a wall to verify collisions. It fails if the collision cast is removed, and it fails if the steering sign is flipped. Cars, keys and chop shops are saved between sessions. Moving a car on a baked map and tram stops on one remain unverified.

The dog

The stray dog is the story of the redo. Two scripted Blender dogs, built with a skin modifier and hand-placed rings, were rejected as bad. A model from the studio pipeline, which is a text-to-image-to-3D chain, was accepted in the game. Strays then dropped from 6 to 4, and the share of dogs went from 0.6 to 0.35.

Getting there exposed pipeline breakages. One ComfyUI version produced black multi-view images (NaN values), while an older install with a different library version produced good ones. That is the kind of failure that looks like bad art and is really a software version.

The asset pipeline, with a rule from the designer

The pipeline uses Z-Image for pictures and TRELLIS.2 for 3D, then reduces the mesh to a game budget. An oil drum took about 220 seconds early on, at 8,000 faces and 0.7 MB.

Condensing mesh generation settings and optimizing geometry queries cut 3D prop generation time from 231 to 102 seconds. Two servers gave no gain at 4 props. How each setting earned its place is in Making 3D props from a picture.

Two quality rules came from review. The designer asked for batches of 10, with each source image checked and fixed before building, and every result looked at as pixels. A passing asset check proves nothing about shape. Also, TRELLIS fails on grids, wire, glass and white or glossy objects on white backgrounds. Solid matte prompts in a mid-tone work better.

A bug in the pipeline shipped the previous run's model when a graph was rejected. That was fixed.

The pack work is paused. A night batch built 468 assets, all passing the gate, and a kit of 195 poor-district assets waits for approval. A Blender asset library now holds 772 assets, tagged by licence. 38 Poly Haven and 26 Hunyuan assets are marked not sellable, and 64 of them were queued for a rebuild with TRELLIS.

A negative result we kept

An upscaler study aimed to replace the engine's FSR 2 at 67% render scale. 18 papers were read, 17 clips captured, and a pre-registered plan frozen. Training stopped after a GPU reset and the designer's call, with no verdict. Two things stayed useful: a small 16-channel network costs 1.402 ms in a Godot effect on a 7900 XTX, and FSR 2 is not repeatable, with 0 of 120 frames identical.

A texture-generation setting, MIOpen, faulted the GPU on two runs at jobs 4 to 7, so it is now off by default (details).

What is next

From the status page, planned and not in code: a named heir and status drop on death, a player-to-player market sell, stream tips and a shared listing board. Each needs a design answer first.

In-game smartphone apps (Corner for legal trade and DarkCorner for illicit goods) remain on paper. The 509-NPC work has its plan and a target of 4 ms for NPC thinking per tick. The five flaky tests need event waits. Sellable asset packs await approval and a restart. And we have not yet run a Windows build or a play session with real people, both still open.

Reacties

Nog geen reacties.

Inloggen om te reageren.