--- license: mit pipeline_tag: image-to-3d base_model: VAST-AI/TripoSG-scribble tags: - lux3d - rust - candle - safetensors - canonical-weights - 3d - image-to-3d - mesh - triposg - triposg-scribble - scribble --- # Lux3D TripoSG-scribble Canonical Weights Canonicalized `safetensors` export of the [`VAST-AI/TripoSG-scribble`](https://huggingface.co/VAST-AI/TripoSG-scribble) checkpoint used by the Lux3D runtime. ## Model Details - Model family: `triposg` - Variant: `triposg-scribble` - Canonical file: `model.safetensors` (4,705,388,818 bytes, about 4.38 GiB) - Tensor count: `1735` - Parameter storage: F16 (same precision as the upstream shards) - SHA-256: `0c31d6c251c89ea18556ad4e15fd8d499405fc868fb002d5491fea87dbb705b1` TripoSG-scribble is the dual-conditioning variant of TripoSG: - a DINOv2-large scribble image encoder (24 layers, width 1024, patch 14); - a 21-block rectified-flow DiT with two cross-attention branches: CLIP text (`cross_attention_dim=768`) and scribble image (`cross_attention_2_dim=1024`); - an SDF VAE decoder; - a CLIP text encoder and its byte-level BPE tokenizer; - 512 latent tokens, 16 steps, CFG-distilled (`guidance_scale=0`), dense and hierarchical octree depth 8. ## Included Files - `model.safetensors` — merged weights with `transformer.*`, `vae.*`, `image_encoder.*` and `text_encoder.*` prefixes - `resolved_config.json` — every component config, the pipeline defaults, and the embedded CLIP `vocab.json` / `merges.txt` / tokenizer configs - `manifest.json` — family, variant, provenance, source checksums, tensor count - `checksums.json` — SHA-256 and size for every required package file No raw diffusers checkpoint or tokenizer files are needed at run time. ## Usage With Lux3D ```powershell cargo run -p lux3d-cli --features cuda -- run triposg --mode scribble ` --prompt "a cat with wings" ` --model-path ` --source --output ``` The package can also contain a quantized `model.gguf`; the canonical loader prefers it over `model.safetensors`. The standalone GGUF packages in `TripoSG-scribble-GGUF` are a separate, smaller distribution. ## Provenance - Upstream model: [`VAST-AI/TripoSG-scribble`](https://huggingface.co/VAST-AI/TripoSG-scribble) - Upstream code: [`VAST-AI-Research/TripoSG`](https://github.com/VAST-AI-Research/TripoSG) (MIT) - Canonicalization: `weights normalize ... triposg` (auto-detects the `text_encoder/model.safetensors` shard) or `tools/python_baseline/normalize_weights.py` The raw checkpoint is split across `transformer/`, `vae/`, `image_encoder_dinov2/`, `text_encoder/` and `tokenizer/`. Normalization merges the four shards under one safetensors file, records every source checksum in `manifest.json`, and embeds the tokenizer documents in `resolved_config.json`. No tensor values are modified. ## Validation On CPU, a canonical package and the raw checkpoint produced a byte-identical OBJ for the same input, prompt and step count. The canonical loader test constructs DINOv2, the dual-cross-attention DiT, the VAE decoder and the CLIP text encoder, and parses the embedded tokenizer. ## License TripoSG-scribble code and weights are distributed under the **MIT License**. The embedded CLIP text encoder, DINOv2 image encoder and VAE components keep their upstream terms.