Ostgot/glasses-3d-output

Glasses 3D Output — TRELLIS.2-4B

3D GLB models generated from glasses photos using TRELLIS.2-4B on H100.

Files

Single-image (frontal photo, 1024_cascade)

File	Steps	Guidance	Rescale	Verts	GLB Size	Notes
glasses_hq.glb	12	7.5	0.5	851K	37.5 MB	Default params, good baseline
ref_a.glb	20	9.0	0.5	979K	41.2 MB	More steps + slightly higher guidance
ref_b.glb	24	7.5	0.5	961K	40.4 MB	More steps, default guidance
ref_c.glb	20	10.0	0.3	1.09M	38.7 MB	Best fidelity — thinnest temples, best bridge detail

Multi-image (3 photos averaged conditioning, 1024_cascade)

File	Steps	Guidance	Rescale	Verts	GLB Size	Notes
glasses_multi.glb	12	7.5	0.5	991K	40.8 MB	Default params, 3 photos
exp_hd.glb	24	12.0	0.5	1.21M	55.7 MB	High detail, wider guidance interval [0.3, 1.0]

Common parameters

All models use:

• Model: TRELLIS.2-4B (microsoft/TRELLIS.2-4B)
• Pipeline: 1024_cascade (512 → 1024 upscale)
• Attention: SDPA (scaled dot product attention)
• Post-processing: decimation_target=1000000, texture_size=4096, remesh=True
• Seed: 42

Sampler details

All samplers use FlowEulerGuidanceIntervalSampler with sigma_min=1e-5.

Default params from pipeline.json:

• Sparse structure: steps=12, guidance=7.5, rescale=0.7, interval=[0.6, 1.0], rescale_t=5.0
• Shape SLat: steps=12, guidance=7.5, rescale=0.5, interval=[0.6, 1.0], rescale_t=3.0
• Texture SLat: steps=12, guidance=1.0, rescale=0.0, interval=[0.6, 0.9], rescale_t=3.0

ref_a overrides: all samplers steps=20, ss/shape guidance=9.0, tex guidance=2.0 ref_b overrides: all samplers steps=24 ref_c overrides: all samplers steps=20, ss/shape guidance=10.0 + rescale=0.3, tex guidance=2.0 exp_hd overrides: all samplers steps=24, ss/shape guidance=12.0 + interval=[0.3, 1.0], tex guidance=3.0 + interval=[0.3, 1.0]

Best results

• ref_c.glb — most accurate single-image generation (closest to original photo)
• ref_a.glb — good balance of detail and accuracy
• glasses_hq.glb — solid default baseline