Asset Generator for 2D & 3D

A Blender add-on that generates 2D billboard/concept assets from text prompts and can convert them into textured 3D meshes — all in one pipeline, without leaving Blender.

• 2D generation: Z-Image-Turbo generates an image per prompt line, BiRefNet_HR removes the background and splits multi-object images into separate cropped assets, each imported as an upright image-plane object and marked as a Blender asset.

• 3D conversion: selected image-plane assets are sent through an image-to-3D pipeline to produce a textured .glb mesh, imported back into the scene next to the source plane. Two backends are selectable:
  • TRELLIS.2 (Microsoft, TRELLIS.2-4B) — the default; proven, needs no camera parameters.
  • Pixal3D (TencentARC) — pixel-aligned conditioning that better preserves fine surface detail, at the cost of a manual camera FOV guess (there's no real camera to estimate from for flat AI illustrations).

Both stages run in isolated background subprocesses with their own progress bar, so Blender's UI stays responsive during generation.

Requirements

• Windows, Blender 5.2 (bundled Python 3.13).
• An NVIDIA GPU with a recent driver (CUDA 12.8 runtime is installed automatically as part of the dependency setup — no separate CUDA Toolkit install is required to run the add-on).
• ~10–15 GB of free disk space for the model weights, CUDA-enabled PyTorch, and the TRELLIS.2 runtime. The optional Pixal3D backend downloads an additional ~24 GB checkpoint (TencentARC/Pixal3D) the first time it's used.
• 2D generation alone needs no extra setup beyond the dependency install below. 3D conversion additionally needs the compiled CUDA extensions (o-voxel, cumesh, flexgemm, nvdiffrast, flash_attn) — prebuilt wheels for these are bundled or fetched automatically (see Prebuilt CUDA extensions below), so end users do not need a CUDA Toolkit or MSVC install for this either.

Installation

1. Download: https://github.com/tin2tin/Asset_Generator-2D-3D/archive/refs/heads/main.zip and install the repo as a zip via Edit > Preferences > Add-ons > Install...).
2. Enable "Asset Generator (2D/3D)" in the add-ons list.
3. Open its preferences and click "Install Dependencies". This runs in the background and:
   • Installs a CUDA-enabled PyTorch build (torch==2.9.1+cu128) into an isolated addon_packages/ folder (added to sys.path at runtime — it does not touch Blender's own Python environment).
   • Installs the full pinned dependency set from requirements.txt.
   • Clones the TRELLIS.2 and Pixal3D repos (for the 3D pipeline code).
   • Installs the compiled CUDA extensions from prebuilt wheels (falling back to a from-source build only if no matching wheel is found).
4. Watch progress in the Preferences panel and the system console. When it reports "Done — CUDA OK", you're ready to generate.

Dependencies already installed and matching the pinned versions are skipped automatically on subsequent runs — you can re-run "Install 2D + 3D Dependencies" any time to repair or update without a full reinstall.

To remove everything the installer downloaded (to free disk space or start over), use "Nuclear Wipe" in preferences — this deletes addon_packages/ only.

Usage

Both panels live in the 3D Viewport sidebar, under the "2D Asset" tab.

2D Asset Generator

1. Choose Prompt (a single prompt + asset name) or Text-Block (one prompt per line in a Blender text block — useful for batch-generating many assets in one pass).
2. Click "Generate 2D Assets".
3. Each generated image is background-removed, split into per-object crops if it contains multiple subjects, and imported as an upright image-plane object with a transparent material, marked as a Blender asset with a thumbnail preview.

3D Conversion

1. Select one or more of the generated 2D image-plane assets in the viewport.
2. Pick a backend — TRELLIS.2 (default) or Pixal3D. For Pixal3D, set the Camera FOV (0.2 rad is a sane default for flat illustrations) and Mesh Scale in the camera box that appears.
3. Click "Convert Selected to 3D".
4. Each selected plane's texture is run through the chosen backend; the resulting textured mesh is imported as a .glb next to the original plane.
5. Optionally expand Advanced Settings to tune the sparse-structure / shape / texture sampler steps, guidance, seed, and max token budget per stage.

How it works

• Isolated environment: all Python packages are installed into addon_packages/ inside this add-on's folder rather than Blender's own site-packages, so the heavy ML stack can't conflict with Blender's bundled libraries (or vice versa). It's added to sys.path at runtime, not a real venv.
• Subprocess isolation: generation runs in a separate python.exe process (not inside Blender's process), so a crash or CUDA OOM in the model code can't take down the Blender session. Progress lines are parsed from the subprocess's stdout to drive the UI progress bar.
• VRAM discipline: within each subprocess, models are explicitly released (del + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect()) as soon as they're no longer needed — the 2D script never holds the diffusion pipeline and the segmenter in VRAM at the same time, and the 3D script frees per-mesh intermediates after every conversion.
• Frozen dependencies: requirements.txt pins the full transitive closure of runtime packages and installs everything with --no-deps, so nothing can silently pull in a second (CPU-only) torch/numpy that would shadow the CUDA build.

Prebuilt CUDA extensions

The compiled extensions (o-voxel, cumesh, flexgemm, nvdiffrast, flash_attn, and natten — the last two shared with the Pixal3D backend) are ordinarily built from source against a specific CUDA Toolkit + MSVC version — impractical to ask end users to set up. Instead:

• o-voxel, cumesh, flexgemm, and nvdiffrast wheels (built for cu128 / torch2.9 / cp313 / win_amd64) are bundled directly in wheels/ and installed straight from there.
• flash_attn's wheel (~240 MB) and natten's wheel (~130 MB) are too large to commit to git — so they're instead downloaded on demand from the same PozzettiAndrea/cuda-wheels release index at install time.
• If a bundled/remote wheel doesn't match your Python/CUDA/torch combination, the installer falls back to cloning and compiling the extension from source (requires the CUDA Toolkit and MSVC Build Tools with the C++ workload installed).

Known limitations

• Windows only (the CUDA-torch install, wheel selection, and DLL-search-path setup are Windows-specific).
• Requires an NVIDIA GPU — there is no CPU fallback for either generation stage.
• 3D conversion needs a GPU with enough VRAM for TRELLIS.2's 4B-parameter model (low_vram mode is enabled by default to reduce peak usage).

Credits

Built on top of:

• Z-Image-Turbo (Tongyi-MAI) for 2D image generation.
• BiRefNet_HR (ZhengPeng7) for background removal / segmentation.
• TRELLIS.2 (Microsoft) for image-to-3D mesh generation.
• Pixal3D (TencentARC) as the alternative pixel-aligned image-to-3D backend.
• PozzettiAndrea/cuda-wheels for prebuilt Windows CUDA extension wheels.