LongCat-Video-Avatar 1.5 cuts inference to 8 steps — here’s how

LongCat-Video-Avatar 1.5 将推理步数缩减至 8 步——操作指南

Meituan’s LongCat team shipped a quiet but meaningful update to its open-source talking-avatar stack: version 1.5 swaps the audio encoder, distills sampling down to 8 steps, and adds an INT8 path to fit the model on tighter GPUs. Here’s what actually changed under the hood, and which flags you now have to pass. 美团 LongCat 团队为其开源数字人（talking-avatar）技术栈发布了一次低调但意义重大的更新：1.5 版本更换了音频编码器，将采样过程蒸馏至 8 步，并增加了 INT8 路径，以便在显存更小的 GPU 上运行。以下是其底层技术的具体变化以及你需要使用的命令行参数。

LongCat-Video-Avatar 1.5: DMD2 distillation, Wav2Vec2 replacement, and INT8 offloading

LongCat-Video-Avatar 1.5：DMD2 蒸馏、Wav2Vec2 替换与 INT8 卸载

LongCat-Video-Avatar 1.5 is an audio-driven talking-avatar generator that turns one portrait plus an audio clip into a lip-synced video with head motion, expression, and body dynamics. Released May 21, 2026, it is an avatar head built on the 13.6-billion-parameter dense LongCat-Video diffusion transformer [base model, Oct 2025]. LongCat-Video-Avatar 1.5 是一款音频驱动的数字人生成器，它能将一张肖像照和一段音频转化为包含头部运动、表情和身体动态的口型同步视频。该版本于 2026 年 5 月 21 日发布，是基于 136 亿参数的 LongCat-Video 稠密扩散 Transformer（2025 年 10 月基座模型）构建的数字人头部模型。

The headline change: v1.5 replaces the Wav2Vec2 audio encoder used in v1.0 (Dec 16, 2025) with Whisper-Large-v3, which the team attributes to smoother, more natural lip dynamics. 核心变化：v1.5 版本将 v1.0（2025 年 12 月 16 日发布）中使用的 Wav2Vec2 音频编码器替换为 Whisper-Large-v3，团队表示这带来了更平滑、更自然的口型动态。

Three new flags define the v1.5 workflow, none of which exist in v1.0: v1.5 的工作流由三个新参数定义，这些参数在 v1.0 中均不存在：

Flag	What it does	Notes
--use_distill	Enables DMD2-based step distillation, collapsing generation to 8 NFE	Mandatory for v1.5 — omitting it falls back to the full-chain v1.0 sampling schedule, not an 8-step path
--use_int8	Loads the 13.6B dense DiT in INT8 to cut VRAM pressure	Main lever for consumer GPUs
--resolution	Selects 480P or 720P output	New in 1.5

参数	功能	备注
--use_distill	启用基于 DMD2 的步数蒸馏，将生成过程压缩至 8 NFE	v1.5 强制要求 — 省略此参数将回退至 v1.0 的全链采样计划，而非 8 步路径
--use_int8	以 INT8 格式加载 136 亿参数的稠密 DiT，以降低显存压力	消费级 GPU 的主要优化手段
--resolution	选择 480P 或 720P 输出	1.5 版本新增

Under the hood, the technique report describes Cross-Chunk Latent Stitching, which removes redundant VAE decode/encode cycles between autoregressive chunks, enabling seamless minutes-long generation without re-encoding overhead [arXiv:2605.26486]. That matters for long-form output where earlier tools drift or stutter at chunk boundaries. 在底层技术上，技术报告描述了“跨块潜在拼接”（Cross-Chunk Latent Stitching）技术，它消除了自回归块之间冗余的 VAE 解码/编码循环，实现了长达数分钟的无缝生成，且没有重编码开销 [arXiv:2605.26486]。这对于长视频输出至关重要，因为早期的工具在块边界处容易出现漂移或卡顿。

One caveat worth flagging before you invest a GPU-hour: Meituan claims parity-or-better results versus HeyGen, Kling Avatar 2.0, and OmniHuman-1.5 across 508 image-audio pairs, 770 crowdsourced evaluators, and 13,240 judgments [human-eval benchmark]. That evaluation is entirely author-controlled and reported as win-rates; no independent leaderboard corroborates the superiority claim yet. Treat it as vendor evidence, not a settled result. 在投入 GPU 算力之前，有一点需要注意：美团声称在 508 对图音组合、770 名众包评估员和 13,240 次判断的 [人工评估基准] 中，其结果优于或持平于 HeyGen、可灵数字人 2.0 和 OmniHuman-1.5。该评估完全由作者控制并以胜率形式报告；目前尚无独立的排行榜证实其优越性。请将其视为厂商提供的证据，而非定论。

---

Preparing your environment for LongCat-Video-Avatar 1.5: PyTorch 2.6.0, CUDA 12.4, FlashAttention-2

为 LongCat-Video-Avatar 1.5 准备环境：PyTorch 2.6.0, CUDA 12.4, FlashAttention-2

Setup begins by cloning the base repository — the avatar is a head on top of the LongCat-Video foundation model, so you install that first and add the avatar layer on top. Create a Python 3.10 virtual environment before anything else; the pinned FlashAttention-2 wheel is not tested against 3.11+, and version drift there is a common early failure. 安装首先需要克隆基础仓库——该数字人是构建在 LongCat-Video 基础模型之上的头部模型，因此你需要先安装基础模型，再在其上添加数字人层。请务必首先创建一个 Python 3.10 的虚拟环境；指定的 FlashAttention-2 wheel 文件未在 3.11+ 版本上进行测试，版本偏差是早期常见的失败原因。

The dependency the setup most often breaks on is FlashAttention-2 v2.7.4.post1, a hard requirement. Install it from the wheel that matches CUDA 12.4 against PyTorch 2.6.0 — a CUDA mismatch with the default pip build is the single most frequent install error. Then install both requirements files: requirements.txt for the base model and requirements_avatar.txt for the avatar path, which pulls librosa and expects ffmpeg on the system. 安装过程中最容易出错的依赖是 FlashAttention-2 v2.7.4.post1，这是一个硬性要求。请从匹配 CUDA 12.4 和 PyTorch 2.6.0 的 wheel 文件中安装它——与默认 pip 构建的 CUDA 版本不匹配是安装中最常见的错误。然后安装两个需求文件：用于基础模型的 requirements.txt 和用于数字人路径的 requirements_avatar.txt（后者会拉取 librosa 并要求系统安装 ffmpeg）。

Pull the weights with huggingface-cli download meituan-longcat/LongCat-Video-Avatar-1.5. Note the DMD2 distill LoRA — roughly 1.26 GB in the Comfy-packaged form — is a separate artifact and is required for the 8-NFE distilled path, not an optional extra. 使用 huggingface-cli download meituan-longcat/LongCat-Video-Avatar-1.5 拉取权重。请注意，DMD2 蒸馏 LoRA（在 Comfy 打包形式下约为 1.26 GB）是一个独立的组件，对于 8-NFE 蒸馏路径是必需的，而非可选的额外内容。

The reference launch uses torchrun --nproc_per_node=2 with --context_parallel_size=2, i.e. two GPUs in context-parallel. Single-GPU runs are feasible by adding --use_int8 to cut VRAM on the 13.6B dense base, but that is not the configuration the official docs test — expect to tune offloading yourself. 参考启动方式使用 torchrun --nproc_per_node=2 并配合 --context_parallel_size=2，即在上下文并行模式下使用两块 GPU。通过添加 --use_int8 来降低 136 亿参数基座模型的显存占用，单 GPU 运行也是可行的，但这并非官方文档测试的配置——请做好自行调整卸载策略的准备。

---

Producing a talking-head avatar with LongCat-Video-Avatar 1.5: torchrun, AT2V, and DMD2

使用 LongCat-Video-Avatar 1.5 生成数字人：torchrun, AT2V 与 DMD2

LongCat-Video-Avatar 1.5 exposes four task modes, and picking the right one decides how tightly identity is preserved. AT2V (Audio-Text-to-Video) generates a talking clip from an audio track plus a text prompt with no reference image. ATI2V/AI2V add a portrait (Audio-Text-Image or Audio-Image-to-Video), which routes through Reference Skip Attention to hold the face steady without copy-paste artifacts or identity drift. Prefer ATI2V or AI2V when you have a source photo and care about consistency. LongCat-Video-Avatar 1.5 提供了四种任务模式，选择正确的模式决定了身份特征的保留程度。AT2V（音频-文本转视频）通过音频轨道加文本提示词生成说话片段，无需参考图像。ATI2V/AI2V 增加了肖像输入（音频-文本-图像或音频-图像转视频），通过“参考跳跃注意力”（Reference Skip Attention）机制保持面部稳定，避免复制粘贴伪影或身份漂移。当你拥有源照片并关注一致性时，请优先选择 ATI2V 或 AI2V。

The fourth mode is audio-conditioned continuation, covered below. A single-speaker AT2V run looks like this — the snippet is verified and prints the exact upstream CLI it wraps rather than executing the model: 第四种模式是音频条件续写，详见下文。单人 AT2V 的运行方式如下——该代码片段已验证，它会打印出其封装的精确上游 CLI 命令，而不是直接执行模型：

"""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()

Both --use_distill and --model_type avatar-v1.5 are mandatory for the 8-step (8-NFE) distilled path. Set Audio CFG in the 3–5 range per the official docs: below 3 under-animates the mouth, above 5 over-exaggerates it. 对于 8 步（8-NFE）蒸馏路径，--use_distill 和 --model_type avatar-v1.5 都是强制参数。根据官方文档，将 Audio CFG 设置在 3–5 之间：低于 3 会导致口型动画不足，高于 5 则会过度夸张。

For two speakers, run run_demo_avatar_multi_audio_to_video.py. Use --audio_type para for equal-length clips (parallel merge, summed) or --audio_type add for unequal clips sequenced with silence padding; person1/person2 ordering is positional. In parallel-merge mode, L-ROPE-based audio-visual binding supplies spatial speaker separation so voices do not bleed identity between the two faces. 对于双人对话，请运行 run_demo_avatar_multi_audio_to_video.py。使用 --audio_type para 处理等长音频片段（并行合并、求和），或使用 --audio_type add 处理不等长片段（通过静音填充进行序列化）；person1/person2 的顺序由位置决定。在并行合并模式下，基于 L-ROPE 的视听绑定提供了空间说话人分离，确保声音不会在两个面部之间产生身份混淆。

Video continuation feeds a reference clip plus new audio to extend an existing segment. Here Cross-Chunk Latent Stitching removes the redundant VAE decode/encode cycle between autoregressive chunks, so the boundary seam stays clean. 视频续写功能通过输入参考片段和新音频来扩展现有视频段。在此过程中，“跨块潜在拼接”技术消除了自回归块之间冗余的 VAE 解码/编码循环，从而保持了边界缝合的平滑。