You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By accessing these TensorRT engine weights, you agree to the terms of the AVTR-1 Community License and Acceptable Use Policy.

AVTR-1 pre-built TensorRT engines (RTX 40-series)

Pre-built TensorRT fp16 engines for the AVTR-1 live-dialogue avatar pipeline. These let you skip the local engine build entirely — clone, download the remaining runtime ONNX weights, and run.

⚠️ RTX 40-series only (sm_89 / Ada Lovelace). TensorRT engines are compute-capability specific. These were built on an NVIDIA RTX 4060 Ti (sm_89) without a broader hardware-compatibility level, so they load only on RTX 40-series GPUs (4060 / 4060 Ti / 4070 / 4080 / 4090). 30-series (sm_86) and 50-series (sm_120) users must build engines on their own machine (pixi run build-trt-engines).

renderer_runtime_artifacts_cc/
    decoder_b5_fp16.engine         108.1 MB   SPADE decoder (b=5 dynamic)
    warp_network_b5_fp16.engine     99.3 MB   dense-motion warp net (b=5 dynamic)
    modnet_b5_fp16.engine           16.8 MB   portrait matting (b=5 dynamic)
    stitch_network_b5_fp16.engine    0.4 MB   stitching layer
speech2motion_runtime_artifacts_cc/
    hubert_lbs_fp16.engine         607.7 MB   audio features
    avtr1_encode_fp16.engine        20.8 MB   encode_conditions
    avtr1_decode_fp16.engine       279.2 MB   guided CFG decode pass
    avtr1_normalizer.safetensors        2 KB   motion-vector normalization buffers
renderer_runtime_artifacts/
    libgrid_sample_3d_plugin.dll   (Windows, built from source)  custom grid_sample 3D TRT plugin
    libgrid_sample_3d_plugin.so    (Linux, upstream build)       (loaded separately; required by the warp engine)

How these were built

Built with TensorRT 10.11.0.33, CUDA 12.x, on Windows with an RTX 4060 Ti (sm_89), fp16 throughout. The workflow (from scripts/ in the AVTR-1 repo) was:

Download source weights (pixi run download): AVTR-1 weights from avaturn-live/avtr-1 and LivePortrait ONNX graphs from digital-avatar/ditto-talkinghead.
Renderer engines — python scripts/build_renderer_engines.py Four engines parsed directly into the TensorRT builder with a dynamic [1..5] batch profile. Each ONNX graph is patched in memory ("surgery") to make it TRT-parsable at batch 5:
- decoder — Resize surgery on the SPADE conditioning path
- warp_network — Tile/Reshape/Concat surgery on the dense-motion stack
- modnet — SE-block Reshape surgery
- stitch_network — Reshape fixes + redundant 2× ScatterND → single Concat
The warp engine depends on the grid_sample 3D plugin — see step 3.
grid_sample 3D plugin — built from source The warp network uses PyTorch's 3D grid_sample (bilinear, unaligned corners), which TensorRT has no built-in layer for, so TensorRT can't parse that subgraph without help. We compiled a custom TRT plugin from the Apache-2.0 upstream repo SeanWangJS/grid-sample3d-trt-plugin to produce the Windows libgrid_sample_3d_plugin.dll:
- cloned the upstream CUDA/CMake source and configured a CMake + MSVC build against CUDA 12.8 and the bundled TensorRT 10.11 headers/libs;
- compiled the CUDA kernels with nvcc using CUDA_ARCHITECTURES="70;80;86;89", so the DLL's kernels cover sm70, sm80, sm86 (30-series) and sm89 (40-series). Only the engines are sm89-specific — the plugin is portable across these architectures;
- the plugin self-registers via REGISTER_TENSORRT_PLUGIN when the DLL is loaded, so the ONNX parser sees the custom op. The warp builder loads it with ctypes.cdll.LoadLibrary(...) followed by trt.init_libnvinfer_plugins(...); the warp engine build fails outright without it (verified).
The upstream Linux build is shipped as-is (libgrid_sample_3d_plugin.so, compiled for x86_64 Linux against CUDA 12 / TRT 10). On Windows the .dll from this repo is required; the loader picks .dll on Windows and .so on Linux.
AVTR1 engines — python scripts/build_avtr1_engines.py Exports the scripted checkpoint (train-39.scripted.pt) into two ONNX graphs in memory — encode_conditions and the guided CFG decode pass — and parses each into a TRT engine:
- fp16 with LayerNorm-flavoured ops pinned to fp32 for numerical stability
- CFG guidance weight tensors are engine inputs (not baked in), so guidance strength is tunable per request at runtime
- motion normalization buffers are lifted off the scripted wrapper into the avtr1_normalizer.safetensors sidecar
HuBERT engine — python scripts/build_hubert_engine.py From hubert-lbs-avtr1.onnx, dynamic batch + length profile (batch [1..2], 3240..12960 frames) covering one chunk of speech + active listening in parallel.

Caveats

sm_89 only. No hardware_compatibility_level was set on the renderer and HuBERT builds (and --ampere-plus defaults off on the AVTR1 build), so these engines will not load on Ampere (30-series) or Blackwell (50-series).
Requires the matching grid_sample 3D plugin for the warp engine; the loader picks .dll on Windows and .so on Linux.
Engines are deterministic outputs of the build scripts — rebuild any time via pixi run build-trt-engines if you target a different GPU.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

bentay85
/

avtr-1-tensorrt-ada

You need to agree to share your contact information to access this model

AVTR-1 pre-built TensorRT engines (RTX 40-series)

Contents

How these were built

Caveats