Pixal3D — image-to-3D with PBR materials (full precision)
High-fidelity image-to-mesh with material baked in. ≈85–120 s per asset on
an RTX 5090, returning ≈15–17 MB .glb files that already carry shading.
What this repo is: the Pixal3D pipeline at full precision — the three 1.3B transformers as ≈5.5 GB bf16 safetensors under
ckpts/. Weights only, not a retrain. A GGUF build of the same weights (≈2.7 GB per transformer) lives inChrisColeTech/Pixal3D-GGUF.
Samples
Single image in, textured mesh out. Source images were rendered with Z-Image
Turbo; the meshes are Blender Workbench renders of the raw .glb — no
cleanup, no retopology, no separate paint stage.
Note on these samples: they were rendered from the GGUF build of these weights, not from the full-precision files in this repo. Timings and mesh sizes come from that run. The architecture and settings are identical; expect the full-precision path to differ only marginally, but the numbers below are not measured on it.
 |  |
| **source** — `a cute chunky retro robot toy standing upright, rounded metal body, simple friendly face, studio product photograph on a plain white background` | **mesh** — 17.1 MB `.glb`, generated in 119 s, seed 42 |
 |  |
| **source** — `an ornate ceramic teapot with a curved spout and handle, glazed blue and white porcelain, studio product photograph` | **mesh** — 15.5 MB `.glb`, generated in 85 s, seed 42 |
Turntable
Four views, 90° apart.
The chest display and dial, the ear knobs and the boot flare all survive, and the surface reads as brushed metal — material is produced with the geometry rather than in a later pass.
Recommended settings
| Parameter | Production value | Meaning |
|---|---|---|
seed |
any | Deterministic per seed |
| input | one RGB(A) image | Centred subject, plain background |
| output | .glb |
Mesh with baked material |
Supported modes: img2mesh, txt2mesh, and the _textured variants
Notes and gotchas
- ⚠ Meshes export rotated 180° relative to some other generators. Pixal3D puts the subject's front along the opposite axis from TripoSG/TRELLIS.2, so a fixed camera that frames those correctly will render Pixal3D output from behind. If you batch-render thumbnails, orbit the camera 180° (or check one contact sheet per model before trusting a shared preset).
- Material comes for free — no separate paint stage needed, which is the main reason to choose this over a geometry-only model.
- Loads in a low-VRAM mode by default, reporting ≈0 GiB resident after load and streaming weights in as it runs. Generation still wants the card largely to itself; sharing a GPU with another model-serving process produces failures that look like crashes rather than clean OOM errors.
- Sparse-conv backend — logs
[SPARSE] Conv backend: flex_gemm; Attention backend: flash_attnon load and pulls avaleoai/NAFestimator fromtorch.hubon first run, so a fresh machine needs network access.
Files
| Path | Size | Role |
|---|---|---|
ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16.safetensors |
5.55 GB | shape DiT, 1024 latents |
ckpts/slat_flow_img2shape_dit_1_3B_512_bf16.safetensors |
5.55 GB | shape DiT, 512 latents |
ckpts/slat_flow_imgshape2tex_dit_1_3B_1024_bf16.safetensors |
5.55 GB | texture DiT |
ckpts/ss_flow_img_dit_1_3B_64_bf16.safetensors |
5.36 GB | sparse-structure DiT |
ckpts/shape_dec_next_dc_f16c32_fp16.safetensors |
948 MB | shape decoder |
ckpts/tex_dec_next_dc_f16c32_fp16.safetensors |
948 MB | texture decoder |
ckpts/ss_dec_conv3d_16l8_fp16.safetensors |
148 MB | sparse-structure decoder |
For roughly half the download and resident footprint, use the GGUF build:
ChrisColeTech/Pixal3D-GGUF.
Provenance
- Upstream base model: Pixal3D (TencentARC)
- This build: redistributed as-is at full precision. Weights are not retrained here.
- License: MIT, per upstream.