Jihanki Generator

Every vending machine is a single mesh built with bmesh: boxes for the cabinet and the details, profiles of revolution (lathe) for the cans and the PET bottles. It has everything you expect from a Japanese machine: the lit sign, the backlit display window and its rails, the dispensing hatch, the payment panel with its screen, the coin and note slots and the IC card reader.

The colours are not in the materials. Every face carries its own in a face-corner attribute, jhk_color, above 1 for the lit parts. Five shared materials read it: Paint, Plastic, Metal, Glass and Emit. So you can generate as many machines as you like without multiplying materials.

The settings live on the object: size, rows, columns, bottle rows, cabinet and sign colour, share of hot drinks and of sold-out buttons. Every change rebuilds the machine. The seed draws new products, and “Randomize Style” a whole new style. Two buttons set up the render: Cycles on the best available GPU, and a night stage with wet ground, moon, street lamp and camera.

The roadmap starts from a diagnosis of the close-up on the display window. Four things give the 3D away: sharp edges, labels that are flat colours, uniform light, and everything brand new and identical.

It lists fifty features in nine phases, each with its GitHub issue under the same number. The phases follow the dependencies: UVs unlock textures, textures unlock labels and stickers, and so on. A suggested path deals first with the four flaws of the diagnosis, from the best ratio of visual gain to effort to the worst.

Before touching realism, I added images that show how the machine is built, with the Torii Generator method: the same shot rendered twice, the real materials on one side, on the other the dark clay with its edges drawn by Freestyle, blended across a band of 30% of the width. The glass is removed from the wireframe pass: Freestyle ignores transparency and would hide the products.

A surprise when opening the code: the machine already had a 3 mm, two-segment Bevel modifier. The question was why it didn’t show.

The measurement comes down to two numbers. At the roof corner, the vertex closest to the theoretical corner sits 0.87 mm away with “Clamp Overlap”, and 2.20 mm without: the modifier was only bevelling about 1.2 mm. That clamp doesn’t work part by part. Blender computes a single limit for the whole mesh, and on a vending machine the thinnest parts decide: 3 mm for a price tag, 2 mm for the rim under the bottle cap.

The builder now writes each sharp edge’s radius into its bevel weight, and the modifier switches to weight mode, with a 1 cm width: a weight of 0.4 gives 4 mm. The clamp is off, and the builder guarantees that no bevel overflows. A part gets 4 mm, the radius of the sheet metal, or 30% of its smallest dimension if it is thinner. The modifier width becomes a global scale: doubling it doubles every radius.

A real bevel creates a new problem. The door was made of boxes side by side: two stiles, a lintel, four boxes around the hatch. With 4 mm on each side, every joint becomes an 8 mm groove, where the real door is all one piece.

A new primitive builds a pierced plate as one closed mesh. It cuts the plate into a grid from the coordinates of its outline and its holes. Every solid cell gets a front face and a back face, plus a wall on each side that borders a hole or the outside. The grid edges between two coplanar faces have a zero angle: no bevel, no shading line. That is the grid you can see in the wireframe half of the opening image.

While checking the weights, the counter showed zero bevelled edges. A face created with bm.faces.new() has a null normal until it is computed, and edge angles are read from that normal: every angle was 0. Since v0.1, the same bug had hit the function that marks sharp edges. The modifier’s hardened normals hid it in the render; the wireframe pass, which turns the modifier off, showed it.

The UVs are born with each face, like its colour, on two layers. The first is metric: one unit is one metre, on every part. A rust or grime texture will have the same density on the cabinet and on a button. Boxes are projected along their normal, seen from outside so that no texture is mirrored.

The second goes from 0 to 1, only on the parts that will carry a visual: the product labels, the price tags, the advertising panel and the sign. A face attribute says which one.

The products are unrolled: u goes around, in metres, v follows the profile rather than the height, so it doesn’t stretch over the bottle shoulders. The seam runs at the back, facing the rear of the display, and the front of the label lands in the middle.

“No stretching” is checked edge by edge: each edge’s length in UV space, divided by its length in 3D, must equal 1.

The cones shear: the two rings of a band don’t have the same radius, so not the same u for the same angle. The offset grows with the angle, hence a u that starts from the front, where you look at the products. Avoiding it entirely would take a seam at every ring; condensation or dust will hide a slight skew better than a seam.

First I needed a method that works headless, since every documentation image is rendered that way. I tested five leads in Blender 4.2 and 4.5; here are the four main ones.

The chosen method leaves Blender’s font engine to do what it does well, read the font and give the shape of the letters, and does the rest in numpy. A text object, which doesn’t even need to be in the scene, is converted to curves, sampled at 8 points per segment. The fill walks the image row by row: each outline it crosses adds +1 or −1 depending on its direction, and a cumulative sum gives, for each pixel, how many times the outlines wind around it. 4 × 4 samples per pixel smooth the edges. Everything stays in memory, with no GPU and no temporary file.

A trap was waiting for the Japanese font. Blender ships Noto Sans CJK as woff2, 11 MB compressed, and every new character costs about 100 ms, as if the whole font were decompressed for each glyph. An uncompressed font outputs a whole string in 2 to 4 ms. So this step writes in Inter, Blender’s Latin font; Japanese will come with an uncompressed subset of Noto Sans JP.

The materials are shared by every machine: one atlas per machine would have forced each of them to have its own. So the catalogue seed, a scene property, decides a shared catalogue: 48 products drawn from 12 drinks, with their brand, pattern, volume and price. The brands are eight invented names. Each machine picks from that catalogue with its own seed.

Every face that carries a visual stores its tile. In the materials, a shared node group derives the column and row from it, places the 0–1 UVs inside the tile with a pixel and a half of margin, reads the right atlas and multiplies the face colour. Labels get pure white, which lets the image through as is; posters and signs are drawn in grey, and the machine’s accent colour tints them.

The next step is the Japanese font and the text generator: katakana names, prices, the words you read on every can (500ml, 無糖, 新発売). Then come the labels themselves, with brand, name, illustration, barcode and volume. After them, the light, then age.

Beyond that, what comes next depends on the use: the front for close-ups (raised buttons, 7-segment displays, flap hatch), the context for wide shots (recycling bins, rows of machines under an awning, rain on the window), and a real-time export for Unity or Unreal. The last item is a test bench: a lookdev scene next to real photos, to measure realism instead of guessing it.