Entry 23

Faces, the milk spray and the lens splash in the opening

2026-10-1525mvideotooling

The opening had no moving faces, and the user saw no milk from Tommy's bottle. Both come from one mechanism. The stream's per-object "animation" and "state" fields pick frames that the game copies into VRAM over a block of texture.

What the original does:

• Face banks are files 531-550, picked by a mask word per database (0x800b91a0; the opening's is 0xd515). 0x80074b9c loads them, each with a 64-byte header skipped.
• 0x80074ec0(frame, id) is the mouth setter and 0x80075304(id, frame) the eye setter. Each maps an id to a VRAM rect and a bank and does a LoadImage of that frame.
• An actor record resets both to frame 0 and then applies its fields. A pose blinks first (0x80075698), then applies its fields.
• The blink is per object (+0xa5 state, +0xa7 count). It uses a rand of its own at 0x800c4c48, which play_cine seeds with the cinematic's number. It starts with a 1 in 32 chance, then shows eye frame 1 for three records and frame 2 for two.
• Ids 88 and 112 are not characters. They are the milk spray (records 599-608) and the splash over the lens (from 830). The stream animates their textures through the mouth field.

The port:

• face::Faces and face::Blinks hold the banks and the blinks.
• Set::faces applies each record in order, including records the display skips, onto the stage's VRAM.
• Stage::vram_rev makes the renderer re-upload VRAM, through a new Renderer::load_vram_rect.

Checks against the oracle:

• Every face rect at records 300, 604 and 840 matches, except blink timing.
• cine_blink shows why the blinks differ. The oracle drops records when a frame runs long: 104 of the first 840, starting at 89. Each drop skips a pose pass and its rand calls. Up to the first drop, the rand matches step for step.

The splash then drew opaque. The oracle drew it with packet code 0x26. Finding where the semi bit comes from took several tries:

• First guess: flag bit 25, as in the C face emitters (0x80022964 and kin). Hooks on those, and on AddPrim, never fired for cinematic models.
• A LP flags word (+0x14, which becomes object +0x0) is 0 for both objects, so not that.
• The model's bytes in RAM equal the file's, so nothing patches them.
• The answer is the hand-written model renderer at 0x80010000. It keeps (flags & 0x80) << 18 in gp and ORs it into every packet code. Face flag bit 7 is semi-transparency, and an untextured semi face takes its blend from bits 5-6.

model_triangles now sets Vtx::SEMI from bit 7. The stage splits out the semi triangles for set_semi. The splash at record 839 now blends over the room like the oracle's.

Also found on the way: 0x800221cc picks a face's ordering-table depth (first, min, max or average vertex, minus 64), which is where the draw-order bug will be fixed.