LongCat-Video-Avatar 1.5: Meituan’s Open-Source Talking Avatar System Receives Significant Update with Enhanced Efficiency and Audio Fidelity

Meituan’s LongCat team has quietly but significantly updated its open-source talking-avatar generation stack, releasing version 1.5 of LongCat-Video-Avatar. This latest iteration introduces a crucial swap in the audio encoder, a dramatic distillation of the sampling process down to a mere eight steps, and the addition of an INT8 path designed to accommodate models on more constrained GPU hardware. These under-the-hood enhancements promise to make high-quality AI-generated talking avatars more accessible and efficient, marking a notable step forward in the rapidly evolving field of generative AI for video.

LongCat-Video-Avatar 1.5: DMD2 Distillation, Wav2Vec2 Replacement, and INT8 Offloading

LongCat-Video-Avatar 1.5 is an advanced audio-driven talking-avatar generator. Its core function is to transform a single portrait image and an audio clip into a lip-synced video, complete with realistic head motion, expressive facial features, and dynamic body movements. This latest version, released on May 21, 2026, builds upon the foundation of the 13.6-billion-parameter dense LongCat-Video diffusion transformer, first introduced in October 2025. The most impactful change in v1.5 is the replacement of the Wav2Vec2 audio encoder, which was utilized in the earlier v1.0 release (December 16, 2025), with the more sophisticated Whisper-Large-v3. The development team attributes this switch to a noticeable improvement in the smoothness and naturalness of lip synchronization.

The workflow for v1.5 is defined by three new command-line flags that were not present in v1.0, each offering distinct advantages in generation and resource management.

Beneath the surface, the technical report details a technique called Cross-Chunk Latent Stitching. This innovation significantly enhances the generation of longer video sequences by eliminating redundant VAE decode/encode cycles between successive autoregressive chunks. The result is the capability for seamless, minutes-long generation without the overhead typically associated with re-encoding, a crucial improvement for applications requiring extended video output where previous tools might exhibit drift or stuttering at chunk boundaries. This technical advancement is further elaborated in their arXiv publication (arXiv:2605.26486).

One important consideration for users before investing significant GPU resources is Meituan’s performance claims. The company asserts that LongCat-Video-Avatar 1.5 achieves results on par with, or even superior to, leading competitors such as HeyGen, Kling Avatar 2.0, and OmniHuman-1.5. This assertion is backed by a benchmark evaluation involving 508 image-audio pairs, 770 crowd-sourced evaluators, and 13,240 individual judgments. However, it is critical to note that this evaluation was conducted entirely under the control of the authors and reported as win-rates. Independent corroboration from a third-party leaderboard is currently absent, so these claims should be treated as vendor-provided evidence rather than definitively settled outcomes. A comprehensive YouTube walkthrough of the release offers further insights into its capabilities.

Preparing Your Environment for LongCat-Video-Avatar 1.5: PyTorch 2.6.0, CUDA 12.4, FlashAttention-2

Setting up LongCat-Video-Avatar 1.5 begins with cloning the base repository. As the avatar functionality is an add-on layer to the foundational LongCat-Video model, users must first install the base model before layering the avatar components on top. It is strongly recommended to create a Python 3.10 virtual environment prior to any installations, as the pinned FlashAttention-2 wheel has not been tested with Python 3.11 or later, and version incompatibilities in this area are a frequent source of initial setup failures.

The dependency that most commonly causes installation issues is FlashAttention-2 v2.7.4.post1, which is a hard requirement. This must be installed from a wheel file that precisely matches CUDA 12.4 and PyTorch 2.6.0. A mismatch between the CUDA version and the default pip build is identified as the single most frequent cause of installation errors. Subsequently, both requirements.txt (for the base model) and requirements_avatar.txt (for the avatar path) must be installed. The latter file includes dependencies like librosa and assumes the presence of ffmpeg on the system.

Model weights can be downloaded using the command huggingface-cli download meituan-longcat/LongCat-Video-Avatar-1.5. It is important to note that the DMD2 distill LoRA, a roughly 1.26 GB file packaged in Comfy format, is a separate artifact. This LoRA is essential for achieving the 8-NFE distilled path and is not an optional component.

The official reference launch configuration utilizes torchrun --nproc_per_node=2 in conjunction with --context_parallel_size=2, effectively employing two GPUs for context parallelism. While single-GPU operation is feasible by incorporating the --use_int8 flag to reduce VRAM demands on the 13.6B dense base model, this configuration is not what the official documentation tests. Users attempting single-GPU setups should anticipate the need for self-tuning and potentially slower performance.

Producing a Talking-Head Avatar with LongCat-Video-Avatar 1.5: torchrun, AT2V, and DMD2

LongCat-Video-Avatar 1.5 offers four distinct task modes, and the choice of mode significantly impacts how closely the generated avatar’s identity is preserved. The AT2V (Audio-Text-to-Video) mode generates a talking clip solely from an audio track and a text prompt, without requiring a reference image. The ATI2V/AI2V modes incorporate a portrait image (Audio-Text-Image-to-Video or Audio-Image-to-Video). These modes utilize "Reference Skip Attention" to maintain facial consistency, effectively preventing copy-paste artifacts or identity drift. For scenarios where a source photograph is available and identity consistency is paramount, ATI2V or AI2V are the preferred options. The fourth mode is audio-conditioned continuation, which allows for extending existing video segments.

A typical AT2V run for a single speaker is demonstrated by the following Python snippet, which verifies the settings and prints the exact upstream CLI command it encapsulates, rather than executing the model directly. Running the actual model necessitates downloading the weights and the LongCat-Video repository.

"""
Minimal LongCat-Video-Avatar 1.5 8-step inference demo.

This prints the distilled inference settings and the matching upstream CLI.
Running the real model requires downloading the weights and LongCat-Video repo.
"""

MODEL = "meituan-longcat/LongCat-Video-Avatar-1.5"

def main():
    steps = 8
    command = [
        "torchrun",
        "--nproc_per_node=2",
        "run_demo_avatar_single_audio_to_video.py",
        "--context_parallel_size=2",
        "--checkpoint_dir=./weights/LongCat-Video-Avatar-1.5",
        "--stage_1=at2v",
        "--input_json=assets/avatar/single_example_1.json",
        "--use_distill",
        "--model_type=avatar-v1.5",
        "--use_int8",
    ]
    print(f"MODEL uses distilled inference: steps steps")
    print(" ".join(command))

if __name__ == "__main__":
    main()

For the 8-step (8-NFE) distilled path, both the --use_distill flag and --model_type avatar-v1.5 are mandatory. The official documentation recommends setting the Audio CFG within the range of 3 to 5. Values below 3 may result in under-animation of the mouth, while values above 5 can lead to exaggerated mouth movements.

For scenarios involving two speakers, the run_demo_avatar_multi_audio_to_video.py script should be used. Users can opt for --audio_type para for equally-lengthed audio clips (parallel merge, summed audio) or --audio_type add for sequentially arranged clips with silence padding. The ordering of person1/person2 is positional. In parallel-merge mode, an L-ROPE-based audio-visual binding mechanism is employed to provide spatial speaker separation, preventing voices from bleeding identity between the two faces.

Video continuation allows for extending an existing video segment by feeding a reference clip along with new audio. The Cross-Chunk Latent Stitching technique is instrumental here, as it eliminates the redundant VAE decode/encode cycles between autoregressive chunks, ensuring a clean boundary seam without the need for a re-encoding step.

ComfyUI Packaging and Workarounds for LongCat-Video-Avatar 1.5

For users who find the raw torchrun command-line interface too demanding, community-developed ComfyUI nodes offer a more streamlined approach. It’s important to note that these nodes are not officially maintained by Meituan. One such option is the dedicated smthemex/ComfyUI_LongCat_Avatar node, which can be cloned into the ComfyUI/custom_nodes directory. This integration typically includes the INT8 DiT model, a umt5_xxl_fp8_e4m3fn_scaled text encoder, the Whisper-large-v3 audio encoder, the LongCat_Avatar_1.5_vae, a distill LoRA, and a Kim_Vocal_2.onnx vocal separator. This node has been tested for both single- and dual-person scenarios.

Another viable route is Kijai’s ComfyUI-WanVideoWrapper. Within Kijai’s LongCat model folder on Hugging Face, users can find LongCat-Avatar-15_bf16.safetensors, a substantial 31.7 GB file, alongside a 1.26 GB distill LoRA.

Users should exercise caution regarding packaging updates. Older tutorials might reference a longcat_avatar branch of WanVideoWrapper, whereas current v1.5 support has transitioned to the main branch. It is always advisable to check the commit date of any guide before following it.

Regarding licensing, the underlying code and weights are distributed under the MIT license, which permits commercial use. However, this license explicitly excludes trademark and patent rights associated with the "Meituan" and "LongCat" brand names. This means that while commercial applications built upon the model are allowed, using the brand names themselves is prohibited.

Significant hurdles remain before LongCat-Video-Avatar 1.5 can be considered a fully production-ready solution. Currently, there is no hosted inference provider available on Hugging Face, necessitating a local GPU for operation. As an alternative for users without suitable hardware, services like HeyGen, priced at approximately $0.05 per second for a 720p Photo Avatar, offer a GPU-free option. Furthermore, the project does not include built-in consent workflows, rights management tools, or brand-kit functionalities. Governance for generated likenesses must therefore be implemented entirely outside the scope of the LongCat system. The key takeaway is that LongCat-Video-Avatar 1.5 provides open, inspectable, multi-person avatar generation capabilities at the cost of a GPU, but users are solely responsible for implementing all necessary guardrails and ethical considerations.

Frequently Asked Questions

What GPU do I need to run LongCat-Video-Avatar 1.5?
The officially recommended configuration involves multiple GPUs, with the avatar command launching using torchrun --nproc_per_node=2 and --context_parallel_size=2. Given that the base model is a 13.6-billion-parameter dense diffusion transformer and Kijai’s bf16 build is approximately 31.7 GB, a practical minimum requirement would be two 24 GB GPUs or a single GPU with 40 GB or more of VRAM when utilizing the --use_int8 flag. While running on a single consumer GPU with INT8 quantization and offloading is technically possible, it falls outside the officially tested configurations, and users should anticipate slower performance and the need for manual tuning.

What changed between LongCat-Video-Avatar v1.0 and v1.5?
Version 1.0, released on December 16, 2025, utilized a Wav2Vec2 audio encoder. Version 1.5, released on May 21, 2026, introduces several substantial upgrades. The audio encoder has been upgraded from Wav2Vec2 to Whisper-Large-v3, resulting in smoother lip dynamics. The introduction of DMD2-based step distillation, activated by the --use_distill flag, collapses inference to just 8 steps. The --use_int8 flag enables INT8 loading, reducing VRAM requirements. The --resolution flag now allows users to select between 480P and 720P output. Finally, Cross-Chunk Latent Stitching has been implemented to enhance temporal stability in longer video generations.

What does the --use_distill flag do and why is it required for v1.5?
The --use_distill flag activates the DMD2 distillation LoRA, which significantly reduces the diffusion sampling process to just 8 NFE (number of function evaluations, or steps). Without this flag, the v1.5 weights revert to the full sampling schedule inherited from v1.0, which is considerably slower and not the intended inference path for v1.5. This flag is the reason behind the 8-step generation time, a substantial improvement over the dozens of steps typically required by standard diffusion schedules. Therefore, it must be included in every v1.5 avatar generation command.

Is LongCat-Video-Avatar MIT-licensed for commercial use?
Yes, the weights, code, and contributions are released under the MIT license, which permits commercial use of the model itself. However, the MIT terms explicitly exclude trademark and patent rights for the "Meituan" and "LongCat" brand names. This means that while users can build and distribute commercial products based on the model, they are not permitted to use these specific brand names as their own. As with any licensing agreement, it is prudent to review the license text thoroughly in relation to your specific deployment plans.

How does LongCat-Video-Avatar 1.5 compare to HeyGen for talking avatars?
LongCat-Video-Avatar 1.5 is a self-hosted solution that is MIT-licensed and offers zero cost per video once the necessary GPU hardware is acquired. It natively supports dual-speaker scenes and audio-conditioned video continuation. In contrast, HeyGen is a Software-as-a-Service (SaaS) platform priced at approximately $0.05 per second for a 720p Photo Avatar. HeyGen also requires a consent video for each Digital Twin and includes built-in production governance features. LongCat, however, lacks these integrated tools, such as consent workflows, rights management, captioning, or brand kits. The choice between them depends on user priorities: LongCat is ideal for those prioritizing control, privacy, and minimal ongoing costs, while HeyGen is suited for users who value governance, reliability, and comprehensive support.