ChrisColeTech/Pixal3D-GGUF

Pixal3D — GGUF build

The Pixal3D image-to-3D pipeline with its three 1.3B DiTs quantized to GGUF — ≈2.7 GB each instead of ≈5.5 GB, roughly halving the download and the resident footprint. ≈85–120 s per asset on an RTX 5090, with PBR material baked into the mesh.

What this repo is: GGUF quantizations of the Pixal3D sparse-structure, shape and texture transformers, plus the fp16 decoders and the background remover / DINOv3 encoder the pipeline needs — weights only, not a retrain. Every sample below was generated with these exact GGUF weights.

---

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.

**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 quantization, 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 renders Pixal3D output from behind. Orbit the camera 180°, or check one contact sheet per model before trusting a shared preset.
• Material comes for free — no separate paint stage, which is the main reason to pick 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.
• Two shape DiTs ship here — _512_ and _1024_, differing in latent resolution. Both are the same 1.3B architecture; pick one.
• Sparse-conv backend — logs [SPARSE] Conv backend: flex_gemm; Attention backend: flash_attn on load and pulls a valeoai/NAF estimator from torch.hub on first run, so a fresh machine needs network access.

---

Files

Path	Size	Role
split/shape/slat_flow_img2shape_dit_1_3B_1024_bf16.gguf	2.78 GB	shape DiT, 1024 latents
split/shape/slat_flow_img2shape_dit_1_3B_512_bf16.gguf	2.78 GB	shape DiT, 512 latents
split/texture/slat_flow_imgshape2tex_dit_1_3B_1024_bf16.gguf	2.78 GB	texture DiT
split/Sparse/ss_flow_img_dit_1_3B_64_bf16.gguf	2.68 GB	sparse-structure DiT
split/decoder/shape_dec_next_dc_f16c32_fp16.safetensors	948 MB	shape decoder (fp16, not quantized)
split/decoder/tex_dec_next_dc_f16c32_fp16.safetensors	948 MB	texture decoder (fp16)
split/decoder/ss_dec_conv3d_16l8_fp16.safetensors	148 MB	sparse-structure decoder (fp16)
extras/dinov3-vitl16/	1.21 GB	DINOv3 image encoder
extras/birefnet/	444 MB	background remover
extras/moge-2-vitl/	1.31 GB	geometry estimator

---

Provenance

• Upstream base model: Pixal3D (TencentARC). Full-precision safetensors of the same weights live in ChrisColeTech/Pixal3D.
• This build: GGUF quantizations of the three transformers; decoders and extras are unquantized. Weights are not retrained here.
• License: MIT, per upstream.