TripoSG-scribble β€” standalone GGUF packages (Lux3D)

Standalone GGUF packages of TripoSG-scribble, produced from the raw VAST-AI/TripoSG-scribble checkpoint by Lux3D.

Each file contains the complete model:

  • DINOv2-large scribble image encoder (image_encoder.*)
  • dual-cross-attention rectified-flow DiT (transformer.*)
  • SDF VAE decoder (vae.*)
  • CLIP text encoder (text_encoder.*)
  • CLIP tokenizer (vocab.json, merges.txt, tokenizer configs)
  • all component configs, pipeline defaults and scheduler config

No raw checkpoint, safetensors shards or tokenizer files are required. Pass the .gguf file directly, or put it in a directory as model.gguf and pass the directory.

Files

file dtype size quantized matrices status
triposg-scribble-q8_0.gguf Q8_0 DiT, F16 other components 2.916 GiB 308 recommended
triposg-scribble-q6_k.gguf Q6_K DiT, F16 other components 2.537 GiB 307 compact fallback
triposg-scribble-q5_k.gguf Q5_K DiT, F16 other components 2.329 GiB 307 preview
triposg-scribble-q4_k.gguf Q4_K DiT, F16 other components 2.133 GiB 307 preview

Every file stores 1,735 tensors. Only rank-2 linear *.weight matrices inside the DiT are quantized; the CLIP text encoder, VAE decoder and DINOv2 image encoder stay in F16. Norms, biases, embeddings, convolutions and rows that do not divide into the dtype block also stay in F16. Q8_0 quantizes 308 matrices; the K-quants use 256-wide blocks and skip one matrix whose row length is not divisible by 256 (307 quantized). SHA-256 and exact byte sizes are in checksums.json.

The raw F16 checkpoint is 4.38 GiB. The standalone Q8_0 package is 2.916 GiB and removes the need to download the 3.13 GiB float DiT plus the other components separately.

Optional component quantization

The shipped files quantize only the DiT. The VAE decoder can also be Q8_0 with effectively no change: on CUDA, against the transformer-only Q8_0 mesh, CD rel is 0.006%, F@1% = 1.000 and voxel IoU = 0.991; the package drops to 2.70 GiB. Build that variant with:

lux3d-cli weights quantize triposg --scribble --standalone \
  --source /path/to/TripoSG-scribble \
  --output triposg-scribble-q8_0-vae.gguf \
  --dtype q8_0 --fallback-dtype f16 \
  --prefix transformer. --prefix vae.

DINOv2 (image_encoder.) and CLIP (text_encoder.) Q8_0 are supported but measurably drift the conditioning (CD rel 0.18%/0.16%, F@1% 0.94/0.96 against transformer-only Q8_0), so they are not used in the default files.

Evaluation

Reference: the raw checkpoint's native F16 DiT promoted to F32, same CUDA backend, seed, steps and input. Input: cat_with_wings.png, prompt "a cat with wings", 16 steps, 512 tokens, dense/hierarchical depth 8, seed 42. The reference mesh has 80,109 vertices / 159,976 faces. Metrics are drift against the float reference, not ground-truth accuracy.

variant CD rel F@0.5% F@1% voxel IoU normal vertices faces max drift mean drift
q8_0 0.461% 0.626 0.907 0.268 0.925 82,145 164,044 0.0669 0.0131
q6_k 0.603% 0.515 0.812 0.204 0.908 84,865 169,440 0.1043 0.0171
q5_k 0.557% 0.523 0.857 0.203 0.920 76,732 153,204 0.0654 0.0158
q4_k 0.744% 0.444 0.742 0.167 0.893 83,340 166,368 0.1421 0.0211

The CFG-distilled scribble DiT is more quantization-sensitive than the image TripoSG model. Only Q8_0 keeps F@1% above 0.90; every K-quant changes the surface topology substantially. Prefer Q8_0 for production, Q6_K as a compact fallback, and treat Q5_K/Q4_K as preview-grade.

Standalone and DiT-only builds produce byte-identical meshes for the same quantized DiT and input; the table above is shared with the DiT-only package.

Usage

Download one standalone file:

hf download oxide-lab/TripoSG-scribble-GGUF \
  triposg-scribble-q8_0.gguf --local-dir .

Run it directly; no base checkpoint is needed:

lux3d-cli run triposg --device cuda --mode scribble \
  --prompt "a cat with wings" \
  --model-path ./triposg-scribble-q8_0.gguf \
  --source ./scribble.png \
  --output scribble.glb

Or use the package-directory form:

mkdir -p triposg-scribble-q8_0
mv triposg-scribble-q8_0.gguf triposg-scribble-q8_0/model.gguf
lux3d-cli run triposg --device cuda --mode scribble \
  --prompt "a cat with wings" \
  --model-path ./triposg-scribble-q8_0 \
  --source ./scribble.png \
  --output scribble.glb

Mesh output dispatches on the extension: .glb (primary), .obj, .ply, .stl.

Conversion

lux3d-cli weights quantize triposg --scribble --standalone \
  --source /path/to/TripoSG-scribble \
  --output triposg-scribble-q8_0.gguf \
  --dtype q8_0 --fallback-dtype f16

--source is the raw checkpoint directory. The embedded lux3d.resolved_config metadata carries every component config, the pipeline defaults and the CLIP tokenizer documents, which is what makes the output a standalone package.

License

TripoSG-scribble code and weights are distributed under the MIT License; see LICENSE. Model weights remain subject to the upstream license. The package also embeds the upstream CLIP text encoder, DINOv2 image encoder and VAE components; their upstream terms continue to apply.

Downloads last month
51
GGUF
Model size
2B params
Architecture
triposg
Hardware compatibility
Log In to add your hardware

6-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for oxide-lab/TripoSG-scribble-GGUF

Quantized
(1)
this model

Collection including oxide-lab/TripoSG-scribble-GGUF