{"slug": "longcat-video-avatar-1-5-cuts-inference-to-8-steps-here-s", "title": "LongCat-Video-Avatar 1.5 cuts inference to 8 steps — here's", "summary": "Meituan's LongCat team released version 1.5 of its open-source talking-avatar generator LongCat-Video-Avatar, which replaces the Wav2Vec2 audio encoder with Whisper-Large-v3 and uses DMD2-based step distillation to cut inference to 8 steps. The update also adds an INT8 path for consumer GPUs and introduces Cross-Chunk Latent Stitching for seamless long-form generation. Meituan claims parity-or-better results versus HeyGen and Kling Avatar 2.0, but the evaluation is author-controlled and not independently verified.", "body_md": "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.\n\nLongCat-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]. 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.\n\nThree new flags define the v1.5 workflow, none of which exist in v1.0:\n\n| Flag | What it does | Notes |\n|---|---|---|\n`--use_distill` |\nEnables 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 |\n`--use_int8` |\nLoads the 13.6B dense DiT in INT8 to cut VRAM pressure | Main lever for consumer GPUs |\n`--resolution` |\nSelects 480P or 720P output | New in 1.5 |\n\nUnder 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.\n\nOne 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. Credit to the source video that surfaced this release: [YouTube walkthrough](https://www.youtube.com/watch?v=Bg5qZXxGa9E).\n\nSetup begins by cloning the base repository — the avatar is a head on top of the [LongCat-Video](https://github.com/meituan-longcat/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.\n\nThe 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`\n\nfor the base model and `requirements_avatar.txt`\n\nfor the avatar path, which pulls librosa and expects ffmpeg on the system .\n\nPull the weights with `huggingface-cli download meituan-longcat/LongCat-Video-Avatar-1.5`\n\n. 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.\n\nThe reference launch uses `torchrun --nproc_per_node=2`\n\nwith `--context_parallel_size=2`\n\n, i.e. two GPUs in context-parallel . Single-GPU runs are feasible by adding `--use_int8`\n\nto cut VRAM on the 13.6B dense base, but that is not the configuration the official docs test — expect to tune offloading yourself.\n\nLongCat-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. The fourth mode is audio-conditioned continuation, covered below.\n\nA single-speaker AT2V run looks like this — the snippet is verified and prints the exact upstream CLI it wraps rather than executing the model:\n\n```\n\"\"\"Minimal LongCat-Video-Avatar 1.5 8-step inference demo.\n\nThis prints the distilled inference settings and the matching upstream CLI.\nRunning the real model requires downloading the weights and LongCat-Video repo.\n\"\"\"\n\nMODEL = \"meituan-longcat/LongCat-Video-Avatar-1.5\"\n\ndef main():\n    steps = 8\n    command = [\n        \"torchrun\",\n        \"--nproc_per_node=2\",\n        \"run_demo_avatar_single_audio_to_video.py\",\n        \"--context_parallel_size=2\",\n        \"--checkpoint_dir=./weights/LongCat-Video-Avatar-1.5\",\n        \"--stage_1=at2v\",\n        \"--input_json=assets/avatar/single_example_1.json\",\n        \"--use_distill\",\n        \"--model_type=avatar-v1.5\",\n        \"--use_int8\",\n    ]\n    print(f\"{MODEL} uses distilled inference: {steps} steps\")\n    print(\" \".join(command))\n\nif __name__ == \"__main__\":\n    main()\n```\n\nBoth `--use_distill`\n\nand `--model_type avatar-v1.5`\n\nare 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 .\n\nFor two speakers, run `run_demo_avatar_multi_audio_to_video.py`\n\n. Use `--audio_type para`\n\nfor equal-length clips (parallel merge, summed) or `--audio_type add`\n\nfor 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 .\n\nVideo 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 without a re-encode step .\n\nIf the raw `torchrun`\n\npath is too heavy, two community ComfyUI routes wrap LongCat-Video-Avatar 1.5 — neither maintained by Meituan. The dedicated [smthemex/ComfyUI_LongCat_Avatar](https://github.com/smthemex/ComfyUI_LongCat_Avatar) node clones into `ComfyUI/custom_nodes`\n\nand pulls the INT8 DiT, a `umt5_xxl_fp8_e4m3fn_scaled`\n\ntext encoder, the Whisper-large-v3 audio encoder, the `LongCat_Avatar_1.5_vae`\n\n, a distill LoRA, and a `Kim_Vocal_2.onnx`\n\nvocal separator; it is tested for single- and dual-person scenarios . The second route is Kijai's [ComfyUI-WanVideoWrapper](https://github.com/kijai/ComfyUI-WanVideoWrapper), whose [LongCat model folder](https://huggingface.co/Kijai/WanVideo_comfy/tree/main/LongCat) ships a `LongCat-Avatar-15_bf16.safetensors`\n\nat 31.7 GB plus a 1.26 GB distill LoRA.\n\nWatch for packaging churn. Older tutorials point at a `longcat_avatar`\n\nbranch of WanVideoWrapper, while current v1.5 support has moved to the main branch — always check a guide's commit date before following it. On licensing, the [MIT terms](https://github.com/meituan-longcat/LongCat-Video) cover weights and code but explicitly exclude trademark and patent rights to the \"Meituan\" and \"LongCat\" names : commercial use of the model is allowed; using the brand names is not.\n\nTwo gaps remain before production. There is no hosted inference provider on Hugging Face — roughly 1,581 downloads last month and 44 community Spaces — so a local GPU is mandatory; [HeyGen](https://developers.heygen.com/docs/pricing) at $0.05/sec for a 720p Photo Avatar stays the zero-GPU alternative. And the project ships no consent workflow, rights management, or brand-kit tooling — governance for generated likenesses must be built entirely outside it. The concrete takeaway: LongCat-Video-Avatar 1.5 gives you open, inspectable, multi-person avatar generation for the cost of a GPU, but you own every guardrail it doesn't provide.\n\nThe reference configuration is multi-GPU: the official avatar command launches with `torchrun --nproc_per_node=2`\n\nand `--context_parallel_size=2`\n\n. Because the base is a 13.6-billion-parameter dense diffusion transformer and Kijai's bf16 build weighs about 31.7 GB, a realistic floor is two 24 GB cards or a single 40 GB+ card with `--use_int8`\n\n. A single consumer GPU with INT8 quantization and offloading is possible, but it sits outside the officially tested two-GPU config, so expect slower runs and manual tuning.\n\nv1.0 shipped December 16, 2025 with a Wav2Vec2 audio encoder; v1.5 landed May 21, 2026. The substantive changes: the audio encoder moves from Wav2Vec2 to Whisper-Large-v3 for smoother lip dynamics, DMD2-based step distillation collapses inference to 8 steps via the `--use_distill`\n\nflag, INT8 loading arrives through `--use_int8`\n\nto cut VRAM, `--resolution`\n\nexposes 480P/720P output, and Cross-Chunk Latent Stitching improves long-video temporal stability.\n\n`--use_distill`\n\nactivates the DMD2 distillation LoRA that collapses diffusion sampling to 8 NFE (8 steps). Without it, the v1.5 weights fall back to the full-chain sampling schedule inherited from v1.0 — significantly slower and not the intended v1.5 inference path. The flag is why the distilled route is 8 steps rather than the dozens a standard diffusion schedule would run, so it belongs in every v1.5 avatar command.\n\nYes. The weights, code, and contributions are distributed under the MIT license, which permits commercial use of the model itself. The one carve-out is that the MIT terms explicitly exclude trademark and patent rights to the \"Meituan\" and \"LongCat\" names — so you can build commercial products on the model, but you cannot use those brand names as your own. As always, verify the license text against your specific deployment before shipping.\n\nLongCat is self-hosted and MIT-licensed, costs $0 per video once you own the GPU, and natively supports dual-speaker scenes and audio-conditioned video continuation. HeyGen is turnkey SaaS priced around $0.05/sec for a 720p Photo Avatar, with a consent video required for every video-based Digital Twin and production governance built in. LongCat ships none of that tooling — no consent workflow, rights management, captions, or brand kits. Choose LongCat for control, privacy, and marginal cost; choose HeyGen for governance, reliability, and support.", "url": "https://wpnews.pro/news/longcat-video-avatar-1-5-cuts-inference-to-8-steps-here-s", "canonical_source": "https://dev.to/creeta/longcat-video-avatar-15-cuts-inference-to-8-steps-heres-152b", "published_at": "2026-07-21 21:52:07+00:00", "updated_at": "2026-07-21 22:00:08.429141+00:00", "lang": "en", "topics": ["artificial-intelligence", "generative-ai", "computer-vision", "ai-products", "ai-infrastructure"], "entities": ["Meituan", "LongCat", "Whisper-Large-v3", "HeyGen", "Kling Avatar 2.0", "OmniHuman-1.5", "FlashAttention-2", "PyTorch"], "alternates": {"html": "https://wpnews.pro/news/longcat-video-avatar-1-5-cuts-inference-to-8-steps-here-s", "markdown": "https://wpnews.pro/news/longcat-video-avatar-1-5-cuts-inference-to-8-steps-here-s.md", "text": "https://wpnews.pro/news/longcat-video-avatar-1-5-cuts-inference-to-8-steps-here-s.txt", "jsonld": "https://wpnews.pro/news/longcat-video-avatar-1-5-cuts-inference-to-8-steps-here-s.jsonld"}}