spotnack/triposg-sm8750-qnn

Source: 🤗image-to-3dmit7.4 GBother✓ 30 checksumsupdated yesterday

CommentsShare your thoughts and shape the future of sovereign AI.

@spotnack is unclaimedSign in with the matching Hugging Face account or an organization member account to verify this handle.Claim @spotnack

Files and sources

SourceAvailable on 🤗

Checksums30 awaiting witness

TorrentNot listed

SwarmNot checked

Source revision: 2ef3d19f37b2…
Witnessed: Not yet
Weight files: 30
Model size: 7.4 GB
Downloads · last month: 0
Publisher: Unclaimed

Public checksum record coming on the next sync

Weights

Other

Weights7.4 GB
30 files
v2-fp32-merged-50step/dinov2_encoder.onnx 1.2 GBv2-fp32-merged-50step/vae_post/model.bin 413 MBv2-fp32-merged-50step/blk13/model.bin 281 MBv2-fp32-merged-50step/blk14/model.bin 281 MBv2-fp32-merged-50step/blk17/model.bin 281 MBv2-fp32-merged-50step/blk18/model.bin 281 MBv2-fp32-merged-50step/blk19/model.bin 281 MBv2-fp32-merged-50step/blk12/model.bin 281 MBv2-fp32-merged-50step/blk11/model.bin 281 MBv2-fp32-merged-50step/blk15/model.bin 281 MBv2-fp32-merged-50step/blk20/model.bin 281 MBv2-fp32-merged-50step/blk16/model.bin 281 MBv2-fp32-merged-50step/blk00/model.bin 247 MBv2-fp32-merged-50step/blk01/model.bin 247 MBv2-fp32-merged-50step/blk02/model.bin 247 MBv2-fp32-merged-50step/blk03/model.bin 247 MBv2-fp32-merged-50step/blk04/model.bin 247 MBv2-fp32-merged-50step/blk05/model.bin 247 MBv2-fp32-merged-50step/blk06/model.bin 247 MBv2-fp32-merged-50step/blk07/model.bin 247 MBv2-fp32-merged-50step/blk08/model.bin 247 MBv2-fp32-merged-50step/blk09/model.bin 247 MBv2-fp32-merged-50step/blk10/model.bin 247 MBv2-fp32-merged-50step/u2net.onnx 176 MBv2-fp32-merged-50step/head/model.bin 68 MBv2-fp32-merged-50step/vae_query/model.bin 27 MBv2-fp32-merged-50step/tail/model.bin 402.2 kBv2-fp32-merged-50step/dit_chain.onnx 5.3 kBv2-fp32-merged-50step/vae_query/model.onnx 0.4 kBv2-fp32-merged-50step/vae_post/model.onnx 0.3 kB

TripoSG for Snapdragon 8 Elite (QNN, on-device)

TripoSG image-to-3D, converted to run fully offline on a phone's NPU. This is the model set the Spotnack app downloads to turn a photo into a 3D mesh in about 3 minutes, with no server.

This is an unofficial conversion. It is not made or endorsed by VAST, Qualcomm or Meta.

Requirements

Chip: Snapdragon 8 Elite (SM8750, Hexagon V79) only. The compiled NPU files do not load on other chips.
RAM: 16 GB. A conversion locks about 7.1 GB of RAM for the NPU. Phones with 12 GB are untested.
Storage: 6.9 GB free for the model files.

Files

Everything lives under v2-fp32-merged-50step/ (30 files, 6.9 GB). models_manifest.json lists every file with its size and SHA-256 hash.

Part	Files	Size	Runs on
Transformer (1.44 B parameters), 23 parts merged into one session	`dit_chain.onnx`, `head/`, `blk00/` … `blk20/`, `tail/`	5.2 GB	NPU, QNN context binaries, FP32 weights
VAE decoder (latent pass + point query)	`vae_post/`, `vae_query/`	0.4 GB	NPU, QNN context binaries, weights stored at 16 bits by the compiler
DINOv2 image encoder (ViT-Large)	`dinov2_encoder.onnx`	1.1 GB	CPU, ONNX, FP32
u2net background remover	`u2net.onnx`	0.16 GB	CPU, ONNX, FP32 (unchanged rembg release file)

How it was made

The official TripoSG code at commit fc5c4099 with the VAST-AI/TripoSG weights at revision 2c1c516d was exported to ONNX (FP32) one block at a time.
Each transformer and VAE part was compiled on Qualcomm AI Hub (QAIRT 2.50, target Samsung Galaxy S25 (Family), which is SM8750) into a QNN context binary.
The 23 transformer binaries were wired into one ONNX Runtime session (dit_chain.onnx, EPContext nodes) so the NPU shares one scratch buffer. The output is bit-identical to running the 23 parts separately.

Run with ONNX Runtime 1.29 and the QNN 2.50 runtime (onnxruntime-android-qnn, qnn-runtime). The full pipeline, scripts and step-by-step rebuild guide are in the Spotnack repository (ml/triposg-mobile/, docs/reproduce_from_scratch.md).

Quality and speed (Xiaomi 15 Ultra)

50 sampling steps, guidance 7.0. A conversion takes about 176 s, fully offline.
On 5 test images, the phone's meshes pass the match thresholds against the official FP32 pipeline on a cloud GPU (Chamfer/diagonal ≤ 5.5e-3, F-score ≥ 0.917, IoU ≥ 0.958).

Limitations

Shape only: TripoSG generates geometry without color, so meshes are untextured.
Works best with one object, fully in frame, on a plain background.
SM8750 only, 16 GB RAM (see Requirements).

Licenses

TripoSG weights and code (including the DINOv2 encoder weights shipped in the TripoSG release): MIT, Copyright (c) 2025 VAST-AI-Research and contributors.
u2net (U-2-Net by Xuebin Qin et al., binary from the rembg release): Apache-2.0.
QNN context binaries: compiled by Qualcomm AI Hub from the MIT-licensed model above.