Native Hunyuan3D 2.1 Full image-to-mesh extension for Modly, optimized for low-VRAM NVIDIA GPUs with INT8, FP8 and FP16 inference.
Native Modly extension for generating 3D meshes from an image with the full Hunyuan Shape 2.1 model (3.3B) on NVIDIA GPUs with limited VRAM. Author and maintainer: AlefK1708 Version: 0.5.0 Repository: AlefK1708/modly-hunyuan3d-21-lowvram Model: Hunyuan3D Shape 2.1 Full - this is not Mini, Turbo, or Turbo Mini. Backend: native Modly Python integration; no ComfyUI, Gradio, or external server is used. Output: GLB geometry only, with no texture generation.
The Hunyuan Shape 2.1 - Full Low VRAM node accepts an image and returns a mesh output, so it can be connected normally to other compatible nodes in the Modly Workflow.
Main features: Full Hunyuan Shape 2.1 with the complete Safetensors checkpoint. INT8 recommended for GPUs with 8 GB of VRAM, experimental FP8, and FP16. meta initialization, memory-mapped Safetensors loading, and MMGP offloading to reduce RAM peaks. Persistent, versioned INT8/FP8 disk cache that skips repeated runtime quantization. Chunk-streamed dense-grid reconstruction without allocating a complete XYZ cube in RAM. Optional CPU background removal with rembg. Native GLB export without Blender. Attempts to release the heavy pipeline and RAM/VRAM caches after every generation, including errors and cancellations.
The extension returns the geometry produced by Hunyuan Shape without automatic polygon reduction. If you want a lower-poly mesh, connect a dedicated mesh optimization node later in the Modly Workflow.
This release adds a safe persistent cache for quantized weights: The first use of INT8 or FP8 still converts the original FP16 checkpoint locally, then saves the transformer and conditioner in MMGP/Quanto-aware Safetensors files. Later generations load those already-quantized files directly through read-only memory mapping, avoiding the repeated conversion pass and its temporary RAM pressure. INT8 and FP8 have separate caches under the model directory: quantizedcache/int8/ and quantizedcache/fp8/. A manifest binds each cache to the source checkpoint, Hunyuan configuration, precision, cache format, PyTorch, MMGP, and Quanto versions. A mismatch automatically triggers a safe rebuild. Component files are written to unique temporary paths and the manifest is published last. Interrupted or incomplete writes are therefore never treated as valid. If the model directory is read-only, the disk is full, or a cache cannot be loaded, generation falls back to the original checkpoint. Cache creation is an optimization and does not become a new runtime requirement.
Read the full README on GitHub →