---
license: apache-2.0
pipeline_tag: image-to-3d
tags:
  - localmesh-engine
  - image-to-3d
  - mesh
  - gltf
  - trellis
  - multi-view
  - 3d-generation
base_model: TencentARC/Pixal3D
base_model_relation: quantized
---

![LocalMesh Engine](https://raw.githubusercontent.com/Quentincls/localmesh-engine/main/assets/banner.png)

# LocalMesh Engine weights

LocalMesh Engine turns one photo, or four sides of the same subject, into a
textured `.glb`. It runs on an 8 GB NVIDIA card. The code is Apache-2.0 and
lives on [GitHub](https://github.com/Quentincls/localmesh-engine). The project
page is [local-mesh.com/localmesh-engine](https://local-mesh.com/localmesh-engine/).

TRELLIS.2 works from a single image. This engine adds a four view path and,
since version 1.2, puts the photographs' own pixels on the texture. This
repository holds every weight the engine needs, except DINOv3 and the cutout
model, laid out like the folder the engine reads. Most of it is rehosted
unchanged; the three fp8 conversions of the Pixal3D multi-view models exist
nowhere else.

## New in engine 1.2: the photographs on the texture

![The photograph, the texture drawn by the model, and the photograph projected onto it: printed text stays readable](https://raw.githubusercontent.com/Quentincls/localmesh-engine/main/assets/photo-projection.webp)

**No new weights.** Engine [v1.2.0](https://github.com/Quentincls/localmesh-engine/releases/tag/v1.2.0)
reads exactly the files in this repository. A folder downloaded for an earlier
version works as it is: there is nothing to download again.

The change is in the texture. The texture TRELLIS.2 returns is drawn by the
model: close to the photograph in colour and shape, wrong in the detail, so
printed text comes back as glyphs that read as nothing. After texturing, the
engine now projects the photographs back onto the mesh. Every texel a
photograph sees squarely takes that photograph's pixel, so a printed label, a
logo or a fine pattern reads as it does on the photograph. Where no photograph
sees the surface, underneath or on the back of a single photograph, the drawn
texture stays, its colour pulled towards the photographed surface next to it.
The photographs themselves are never retouched.

The cameras are fitted to the images. With four photographs, the engine reuses
the angles DA3-BASE (`multivue/cameras/`, listed below) measured for the shape;
when the shape was made without that measurement, on the `max` tier or without
DA3-BASE, it starts from quarter turns and checks whether the two profiles are
swapped. With one photograph, it searches the whole turn for the angle that
matches the silhouette and straightens a tilt of up to 20 degrees; nothing
under `multivue/` is needed, so the single photo download gets the projection
too. One elevation is shared by all views, and the mesh is never mirrored to
fit. Each texel takes the one photograph that sees it most squarely, so
nothing is averaged into a blur. On a cassette generated from four
photographs, 59 % of the surface takes a photograph pixel, and its printed
label reads as it does on the photographs.

Beside `model_<seed>.glb`, the engine writes `model_<seed>.ai.glb`: the same
mesh with the drawn texture alone, to compare or to fall back on. Expect about
twice the disk per job. On an RTX 4060 Laptop 8 GB the projection adds 29 to 73
seconds per job, most of it spent fitting the cameras, and its memory peak,
3.3 GB, stays below that of the texture stage.

What it cannot do is correct a shape. Where the reconstructed mesh departs from
the photograph, a correct pixel lands on the wrong surface and that part shows
twice. For such a subject, `--no-projection` keeps the drawn texture alone. A
failure of this step never fails the job: the drawn texture is kept, and the
result says why.

## What is in this repository

Everything here is laid out exactly like the folder the engine reads, so one
command places all of it:

| Folder | Size | What it is | Whose |
|---|---|---|---|
| `TRELLIS.2-4B/` | 8.1 GB | the single photo path, in fp8 | [visualbruno](https://huggingface.co/visualbruno/TRELLIS.2-4B-FP8), MIT, rehosted unchanged |
| `microsoft/` | 148 MB | the sparse structure decoder | [Microsoft](https://huggingface.co/microsoft/TRELLIS-image-large), MIT, rehosted unchanged |
| `multivue/` | 4.8 GB | the four view path | ours, from Pixal3D, see below |

Rehosted so that one `hf download` replaces six. Taking those two from their
own repositories works exactly as well; nothing here is modified.

### What is ours, under `multivue/`

| File | Size | Source |
|---|---|---|
| `multivue/structure_mv_fp8.safetensors` + `.json` | 1.39 GB | fp8 conversion of `ckpts/ss_flow_img_dit_1_3B_64_bf16_mv.safetensors`, TencentARC/Pixal3D |
| `multivue/forme_512_mv_fp8.safetensors` + `.json` | 1.44 GB | fp8 conversion of `ckpts/slat_flow_img2shape_dit_1_3B_512_bf16_mv.safetensors`, TencentARC/Pixal3D |
| `multivue/forme_1024_mv_fp8.safetensors` + `.json` | 1.44 GB | fp8 conversion of `ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16_mv.safetensors`, TencentARC/Pixal3D |
| `multivue/champ.safetensors` | 2.7 MB | valeoai/NAF, official checkpoint, tensors unchanged |

**`structure_mv_fp8`** fuses the four encoded views into a sparse volume of
cells. Each view is projected onto the shared grid by its own camera.

**`forme_512_mv_fp8`** is the first shape pass, on the grid inherited from the
structure. **`forme_1024_mv_fp8`** is the second, and the TRELLIS.2 shape
decoder turns its latent into the mesh.

**`champ`** is the NAF field network. It returns a 512 by 512 query map, which
fills in the detail the token map loses: the token map is sixteen times coarser
than the photo.

4.28 GB, seven files, plus `multivue/cameras/`, 544 MB, rehosted unchanged:
DA3-BASE and the code that reads it (Apache-2.0), and the three small Python
packages that code needs, each under its own licence (omegaconf and
antlr4-python3-runtime BSD-3-Clause, addict MIT). Keep the three `.json`
descriptors next to their `.safetensors`: the engine builds each flow model on
the `meta` device from that descriptor, then loads the tensors in place, so the
weights are never held twice. The descriptors declare `dtype: float8_e4m3fn`.

The four view path serves the `draft`, `standard` and `high` tiers. The
multi-view weights exist at 512 and 1024 only, so `max` falls back to the
earlier way of blending the views rather than shipping a `standard` shape under
another name.

## What is not in this repository

Two sets. The engine needs both.

| Weights | Where | Licence | Note |
|---|---|---|---|
| DINOv3 ViT-L/16 | https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m | DINOv3 License, Meta | Image encoder, required on both paths. Access is gated and approved by hand, so ask for it first. |
| BiRefNet_HR | https://huggingface.co/ZhengPeng7/BiRefNet_HR | MIT | Cutout, and the mask that tells the projection which pixels belong to the subject. The engine fetches it from the Hub on the first generation if the cache is empty, and loads it with `trust_remote_code=True`, so that first run executes code from the Hub. |

TRELLIS.2-4B in fp8, the structure decoder and DA3-BASE are in this
repository, rehosted unchanged. Their own repositories work just as well:
[visualbruno/TRELLIS.2-4B-FP8](https://huggingface.co/visualbruno/TRELLIS.2-4B-FP8),
[microsoft/TRELLIS-image-large](https://huggingface.co/microsoft/TRELLIS-image-large),
[depth-anything/DA3-BASE](https://huggingface.co/depth-anything/DA3-BASE).
Without DA3-BASE the engine guesses which side each profile is on, and says so
in its result; a wrong guess puts a face at the front and at the back.

## Using them

Set `LOCALMESH_ROOT` to the folder that holds `models/`, then:

```bash
pip install -U huggingface_hub
hf download Qtn-Cls/LocalMeshEngine --local-dir "$LOCALMESH_ROOT/models"
```

13.1 GB, and every file lands exactly where the engine looks for it. There is
nothing to move afterwards. For the single photo path alone, 8.3 GB, add
`--exclude "multivue/*"`.

**Then ask Meta for DINOv3**, the one set you have to fetch yourself: it is
gated, and a human grants access. Nothing generates without it, and approval is
not instant, so send the request before anything else:
[facebook/dinov3-vitl16-pretrain-lvd1689m](https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m).

The layout the engine reads:

```
<LOCALMESH_ROOT>/models/
  TRELLIS.2-4B/                              this repository
  microsoft/TRELLIS-image-large/ckpts/       this repository
  multivue/                                  this repository
  multivue/cameras/                          this repository, DA3-BASE
  facebook/dinov3-vitl16-pretrain-lvd1689m/  from Meta, gated
  hf/                                        Hugging Face cache (HF_HOME), where BiRefNet_HR lands
```

Then, from the four sides of one subject:

```bash
python -m localmesh_engine face.png --right right.png --left left.png --back back.png --to out/
```

`--tier` picks the tier: `draft`, `standard`, `high` or `max`, written
Draft, Standard, Detailed and Extreme where these pages spell them out.
`--seed` sets the seed, `--to` the output folder. The command assumes the package is installed.
The four view path also needs `natten`. Installation, including the three CUDA
extensions that are not on PyPI, is written up in `docs/INSTALL.md` in the
GitHub repository; the four tiers, frozen, are in `docs/RECIPES.md`.

The photographs are projected onto the texture unless you ask otherwise:

```bash
python -m localmesh_engine photo.png --no-projection --to out/
```

`--no-projection` keeps the drawn texture alone; `--sans-projection` is the
same flag under its French name.

## Results

![Four views to one .glb](https://raw.githubusercontent.com/Quentincls/localmesh-engine/main/assets/four-views.png)

![Gallery](https://raw.githubusercontent.com/Quentincls/localmesh-engine/main/assets/gallery.png)

![Four of the six meshes on a full turntable](https://raw.githubusercontent.com/Quentincls/localmesh-engine/main/assets/turntable.gif)

A full turn each. One pose can be chosen; a full turn cannot.

Six subjects, four photos each, `high` tier, RTX 4060 Laptop 8 GB: samurai on a
base, crowned stone head, sword in the stone, motorcycle, cassette with a clear
shell, traffic light.

| Tier | Four views, measured over the six subjects |
|---|---|
| Standard, `standard` | 6 min 30 to 8 min 30 |
| Detailed, `high` | 9 min 20 to 13 min |

Texture accounts for about 60 % of that time. These times were measured with
engine 1.1, before the projection, which adds 51 to 67 seconds with four
photographs.

## The conversion

The three flow models start from the official `*_mv` weights of
TencentARC/Pixal3D. Their tensors are stored in float32, despite the `bf16` in
their filenames. The transformer blocks are cast to `float8_e4m3fn`, and the
descriptor beside each file records that dtype, so the engine builds the model
in fp8 rather than casting after the fact. Of each model, 480 tensors are
converted; the input and output layers, the norms and modulations, and the
structure model's complex rotary table are left as they were.

Measured against the source: RMSE of 0.025 to 0.026 on the weights, and 0.027
on a projection probe.

`champ.safetensors` is not converted. It carries the tensors of the official
valeoai/NAF checkpoint, `naf_release.pth`, unchanged, re-serialised to
safetensors. The file records the source URL and its SHA-256 in its own
metadata.

## Licences and attribution

The `license` field above is the repository tag. It is Apache-2.0, the licence
of the engine code and of this card. The files themselves keep the licence of
their own source:

- The three fp8 conversions derive from **TencentARC/Pixal3D**, MIT License,
  Copyright (c) 2026 Tencent. Code: https://github.com/TencentARC/Pixal3D,
  weights: https://huggingface.co/TencentARC/Pixal3D
- `champ.safetensors` comes from **valeoai/NAF**, Apache-2.0,
  https://github.com/valeoai/NAF
- `multivue/cameras/` holds **DA3-BASE** and its source, Apache-2.0,
  https://huggingface.co/depth-anything/DA3-BASE, and under `deps/` three
  Python packages, unchanged: **omegaconf** 2.3.0 (BSD-3-Clause, Omry Yadan),
  **antlr4-python3-runtime** 4.9.3 (BSD-3-Clause, The ANTLR Project) and
  **addict** 2.4.0 (MIT, Mats Julian Olsen)

The engine also builds on:

- **TRELLIS.2**, Microsoft, MIT License, https://github.com/microsoft/TRELLIS.2
- **ComfyUI-Trellis2**, visualbruno, MIT License, for the ported multi-view path
  and the fp8 loading, https://github.com/visualbruno/ComfyUI-Trellis2
- **Depth Anything 3**, Apache-2.0, for the camera measurement,
  https://github.com/ByteDance-Seed/depth-anything-3
- **BiRefNet**, MIT License, for the cutout,
  https://github.com/ZhengPeng7/BiRefNet

Built with DINOv3. DINOv3 is the image encoder on both paths. Its weights are
not redistributed here: request them from Meta, and ship a copy of the DINOv3
License Agreement with any redistribution of your own.

## Citing

```bibtex
@software{colus2026localmeshengine,
  author  = {Colus, Quentin},
  title   = {LocalMesh Engine},
  year    = {2026},
  version = {1.2.0},
  license = {Apache-2.0},
  url     = {https://github.com/Quentincls/localmesh-engine}
}
```

The upstream work these weights rest on: TRELLIS.2 (Microsoft,
arXiv:2512.14692), Pixal3D (TencentARC, arXiv:2605.10922), NAF (valeoai),
DINOv3 (Meta), Depth Anything 3 (arXiv:2511.10647).

## Code

The engine is at **https://github.com/Quentincls/localmesh-engine**. Recipes,
installation, measurements and the provenance of every vendored file are there.
Open issues there, not here.

## The application

These weights and this engine are the generation core of LocalMesh, a Windows
application at [local-mesh.com](https://local-mesh.com): the same core with a
board, a library and a viewer, installed in one step instead of fifteen. The
engine is free and open; the application is what is sold.

![The LocalMesh board, covered in generated objects](https://raw.githubusercontent.com/Quentincls/localmesh-engine/main/assets/localmesh-app.webp)