Converting INRIA 3DGS PLY to antimatter15 .splat (32 bytes/Gaussian): rotation must be normalize*128+128, not per-vertex min-max; also sort by size*opacity
Context: shrinking HackPrinceton aerial splats (.../resources/<sid>/splat/splat.ply, INRIA layout, 68 bytes per Gaussian) to the compact .splat web format (32 bytes per Gaussian) for browser viewers. A home-made converter wrote files of exactly the right size, which made it look correct, but its rotation bytes can't be decoded by any viewer.
The layout per Gaussian (little-endian, as in antimatter15/splat convert.py, which most web viewers accept):
- 3 x float32 position x,y,z
- 3 x float32 scale = exp(scale_0..2) (the PLY stores log-scales)
- 4 x uint8 RGBA: clip((0.5 + 0.28209479177387814 * f_dc_i) * 255) and alpha sigmoid(opacity) * 255
- 4 x uint8 rotation: clip(q / |q| * 128 + 128), in the PLY's order rot_0..rot_3 (w, x, y, z)
The bug was the rotation: (q - q.min()) / (q.max() - q.min()) * 255 rescales each quaternion by its own min and max, so you can't decode it. The viewer decodes (byte - 128) / 128, which needs the fixed *128 + 128 mapping. Viewers also expect splats sorted large-and-opaque first: order = argsort(-exp(s0+s1+s2) / (1 + exp(-opacity))).
Vectorised version, byte-identical to the reference loop on test data (a Python loop with struct.pack per vertex is very slow on 5M Gaussians):
v = ply['vertex'].data
order = np.argsort(-np.exp(v['scale_0']+v['scale_1']+v['scale_2']) / (1+np.exp(-v['opacity'])))
v = v[order]; n = len(v)
out = np.zeros(n, dtype=[('p','<f4',3),('s','<f4',3),('c','u1',4),('r','u1',4)])
out['p'] = np.stack([v['x'],v['y'],v['z']],1)
out['s'] = np.exp(np.stack([v['scale_0'],v['scale_1'],v['scale_2']],1))
rgb = 0.5 + 0.28209479177387814*np.stack([v['f_dc_0'],v['f_dc_1'],v['f_dc_2']],1)
a = 1/(1+np.exp(-v['opacity']))
out['c'] = (np.column_stack([rgb, a])*255).clip(0,255).astype(np.uint8)
q = np.stack([v['rot_0'],v['rot_1'],v['rot_2'],v['rot_3']],1); q /= np.linalg.norm(q,axis=1,keepdims=True)
out['r'] = (q*128+128).clip(0,255).astype(np.uint8)
out.tofile('model.splat')Verify: file size is exactly 32 * vertex_count, and dropping it on antimatter15.com/splat shows the same scene as the PLY. The .splat format has no SH bands beyond DC, which costs nothing for these SH-degree-0 files.