BenchLabs Models
Three tiny diffusion models, running live with on-demand GPU (ZeroGPU). All are under 45M trained parameters — the whole model fits in a PNG image if you're curious — see the model pages linked below.
A larger MMDiT model conditioned by T5 and CLIP. 256x256 output.
| Prompt | Detail (sampling steps) | Prompt strength (CFG) | Seed |
|---|
Good at bulky objects: chairs, tables, cars, mushrooms. Thin objects (swords, keys) don't survive 32³ voxelization, in the training data or the model, so expect a blob rather than a blade.
| Prompt | Detail (sampling steps) | Prompt strength (CFG) | Occupancy threshold | Seed |
|---|
17.8 seconds of sound at 22 kHz from a text prompt. The same tiny DiT + rectified-flow recipe as the other tabs, applied to a mel spectrogram image instead of pixels. Conditioned by the CLAP text tower; audio comes back via Griffin-Lim (32 iterations), so expect lo-fi, slightly phasey sound — there's no learned vocoder in v1.
| Prompt | Detail (sampling steps) | Prompt strength (CFG) | Seed |
|---|
Good at material and light: food, landscapes, skies, interiors. Weak on faces, hands, and anything needing precise structure or text.
| Prompt | Detail (sampling steps) | Prompt strength (CFG) | Seed |
|---|
Models: PixelModel v5 · PixelModel v6 · VoxelModel v1 · AudioModel v1