Entry 10

Model chunks, the frame section, and a clean walk over every archive

2026-09-22formatrenderdecomptooling

Filestools/ccs.pytools/ccsmodel.pytools/test_ccs.pydocs/formats/ccs.mddocs/formats/ccs-model.md

7 left the chunk walk stuck after Model chunks in 176 of 1,023 files and did not know what 0xccccff01 was. Both are now read out of the code, and every CCSF file on the disc walks to its exact last byte:

Only two chunk kinds need a rule other than the size field: textures (7) and one kind of model. tools/ccs.py survey reports how many chunks disagree with their size field: all 4,306 textures in DATA.BIN, and 190 models.

The Model chunk

ccStream::Decode_Model (INF SLUS_202.67:0x0014bce0) reads a header - object, f32 vertexScale, u16 mtype, u16 mmatNum, u16 flag, s16 zoffs, a byte that is 0x70 in every file, padding - and then mmatNum "mmats", each a run of vertices sharing one material. mtype picks one of three layouts:

• mtype & 4: Decode_Mmat02 (0x0014a600), drawn by ccModel::DrawBoneType. With offsetNum 0 every vertex rides one clump node ("bone"); otherwise each vertex is one or more weighted entries, one per node, each holding the position in that node's space ("skin"). A weight is 9 bits (256 = 1.0) of a u16 that also carries an end-of-vertex bit and a 6-bit node slot.
• mtype & 8: DecodeShadowModel (0x00140920), positions and a triangle list for the shadow volume.
• otherwise: Decode_Mmat01 (0x0014b890), rigid vertices in the owning object's space, with normals, colours and ST each present unless an mtype bit (0x40, 0x200, 0x400) turns them off.

The full layout is on the model page. The size field disagrees only for mtype & 4 models, and always by exactly 15 words per mmat more than the decoder reads; why is not known.

Units, and how they were settled

• Positions are s16 * vertexScale / 4096. Decode_Mmat02 converts with sceVu0ITOF4Vector and a vertexScale / 256 multiply, and ccBbox_SetBox (0x001388c0) uses / 4096 for every model type. Rigid vertices go to the VU as s16 with vertexScale folded into the matrix (ccSetMatrixPacket); that the microcode then does an ITOF12 is inferred from the bounding-box factor, not read from VU code.
• Normals are s8 / 64; their measured length is 64. The fourth byte of the normal word is the GS strip flag: 1 starts a strip without drawing, 0 draws a triangle with the two vertices before it. All 1,212,751 strips open with two 1s. Strips run on across the 48-vertex VU batches (ccModelDmaTag_SetVertex, 0x0013dae0), because the GS keeps its vertex queue.
• ST is u16 / 256: the STROW offset in ccSetModelPacket wraps one texture repeat at 256, 88% of mmats have UVs inside 0..1, and v = T / 256 lines up with the PNGs of 8 as they are - T = 0 is the first stored row, which is the bottom of the picture. That answers how the UVs deal with bottom-up rows: they simply count from the stored first row.
• Colours are RGBA with 0x80 = 1.0.
• The GS does not cull. With winding alternating from each strip's start, 1.88 million triangles face their stored normals and 0.60 million do not; some models are authored inside out.

Posing

Bone and skin vertices mean nothing until clump nodes are posed. Local matrix = T * Rx * Ry * Rz * S with rotation in degrees (SetMatrix_PosRotZYXScale, 0x00138120); world = parent world * local (_SetLWMatrix__7ccCoordFv, 0x00138380); DrawBoneType (0x0013f860) builds each bone matrix as inv(model world) * node world. That node slot i is the i-th node of the model's clump is inferred, and borne out by poses that come out right. Axes are Z-up, right-handed.

The Anime chunk's sub-stream walks exactly by its size field in all 4,091 chunks; the one controller record decoded so far, 0x0102, gives position, rotation and scale controllers (constant, or n keyed values).

The frame section

DecodeFrameChunk (0x0014e1b0) reads a u16 kind, pads to 4 - skipping the 0xcccc half - and dispatches on its own table. The Frame chunk (0x0005, Decode_Frame) that ends the setup section holds a u32 frame count. Top (0xccccff01, DecodeF_Top, 0x0014e490) holds a u32 frame number and opens each frame; frame -1 ends the section, and -2 ends it and resets the scene (only str7300.tmp). The header u32 that 7 could not name, at ccStream+0x16c, is the length of the seek table at ccStream+0x48 that records each frame's Top chunk.

Every file in DATA.BIN has an empty frame section: Frame = 1, then the end Top. Animated DATA.BIN scenes (x74*cam, the c*body characters) animate through Anime chunks in the setup section instead. The frame section is for the cutscenes in STREAM/, which come in pairs: strNNNNe.tmp carries models and textures, and strNNNN.tmp carries object stubs, audio (Pcm) and the frames - every frame has a camera and an ambient-light chunk. Every frame-section decoder was read, and each reads exactly its size field; checked over roughly 10.4 million frame chunks.

Other answers

• The # in front of a source path in the index (3) marks an object defined in another file that is loaded alongside, as the stream pairs are.
• The Setup chunk is 0 words in all 1,289 files and Decode_Setup reads nothing; the reference page's "0 to 2" came from the broken walk of 3 and is corrected.
• Layouts read along the way: Obj (object, parent, model, and a shadow model from version 0x96), Clump (object, count, node list), ExtObj (object, parent, and the target the animation actually drives), Material (object, texture, f32 transparency, u16 crop U/V), Bbox (a hand-placed culling box, two f32[3] corners).

With models passable, tools/ccstex.py dump now reaches all 4,306 TEX_ objects in DATA.BIN, where 8 reached 3,985.

Tools and checks

tools/ccsmodel.py decodes models and writes Wavefront OBJ, optionally posed from an Anime chunk (--anime) or a stream frame (--frame N --frames-from strNNNN.tmp), with an MTL naming the ccstex.py PNGs. ccsmodel.py check decodes every model in an archive and requires each decode to end exactly at the walker's model_end, indices in range, no NaNs, strips opening correctly and skin weights summing to 1. DATA.BIN: 17,254 models - 16,407 rigid, 3,129 shadow, 2,994 bone and 185 skin mmats - 5,417,645 vertices and 2,819,008 triangles, no problems. STRCMN 1,344 and STR1 6,029 models, no problems. tools/test_ccs.py (8 tests) holds the survey totals and one model's counts.

Looked at by eye:

• cw1hst00, a weapon: 118 vertices and 68 triangles, a rapier with a curved guard; its shadow model is 17 vertices and 26 triangles.
• cbu1body posed with ANM_cbu1nut0: 21 bone mmats plus a skin of 1,137 vertices, an armoured figure with sword and shield standing from z -0.34 to 163.2, inside its hand-placed box (+-90, +-80, 0 to 230). The face mesh's UVs land on the face texture.
• town01 posed with ANM_sr1town1a: a recognisable town.
• Kite and Helba from str0130e + str0130, posed at frames 100 and 1,000.

109 of 168 posed skinned models fit inside their Bbox; the rest miss by a few units, mostly feet just below z = 0. A wrong scale would miss by a factor.