(auto)Load LongCat Avatar Model

(auto)Load LongCat Avatar Model is the pack's front door. It loads the Avatar 1.5 DiT (in the weight format you pick), the VAE, and the required distill LoRA in one shot, and - the reason it's called "auto" - it can pull the official sharded checkpoint down from Hugging Face if it's missing. Most of the classic ComfyUI model-loader confusion in this pack lives here, so it's worth reading this page even though it has zero search impressions: every other node downstream is dead without it.

The one decision that matters: inference_weight_mode

This single dropdown decides everything, because the node has no separate diffusion_models picker - the mode implies the source:

• single_file_safetensors - looks for LongCat-Video-Avatar-1.5-int8.safetensors in ComfyUI/models/diffusion_models/. The community INT8 merged file; the most turnkey option if you already downloaded it.
• official_sharded - the official base_model/ layout under ComfyUI/models/longcat/LongCat-Video-Avatar-1.5/ (config + index + diffusion_pytorch_model-00001-of-00006.safetensors shards). This is the official checkpoint contract.
• official_int8_sharded - the official base_model_int8/ INT8 sharded layout. Lower VRAM, same contract.

In the two official modes, the node validates the index and every referenced shard before loading, and with auto_download_missing_weights on it downloads exactly the missing official assets - a fixed, bounded manifest, no arbitrary URLs or repo IDs. Note what it will not fetch: the VAE, the Whisper encoder, and the distill LoRA are never auto-downloaded, because they live in normal ComfyUI folders and the node needs them present.

The other inputs

• vae - pick LongCat-Video-Avatar-vae.safetensors from models/vae/. none is in the list; don't select it.
• lora - the required distill LoRA (longcat-avatar-dmd_lora.safetensors) from models/loras/. Avatar 1.5 is distill-required; without it you're not running the model the way it was meant to run.
• attention_mode - auto respects the checkpoint config (which currently defaults to FlashAttention 2), falling back to SDPA with a console warning if a backend is missing. sdpa is the most portable; flash_attn_2, flash_attn_3, xformers, and sageattn force specific backends and raise a clear error if the package isn't installed. sageattn_3 is reserved and currently fails fast - don't pick it.

Output is a single model (MODEL) socket feeding the sampler. That's it - VAE and LoRA are folded in by this node, so you don't need separate VAE Loader / LoRA nodes in the graph.

Install and model placement

Pack install is shared with all nine nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-LongCat-Avatar
cd ComfyUI-LongCat-Avatar
pip install -r requirements.txt

or ComfyUI Manager → search ComfyUI-LongCat-Avatar → restart. CUDA GPU required (no CPU/MPS path). For the files, either let auto_download_missing_weights fetch the sharded DiT, or place things by hand:

ComfyUI/models/longcat/LongCat-Video-Avatar-1.5/base_model/...   (official sharded)
ComfyUI/models/vae/LongCat-Video-Avatar-vae.safetensors
ComfyUI/models/loras/longcat-avatar-dmd_lora.safetensors
ComfyUI/models/audio_encoders/whisper-large-v3.safetensors

Where people get burned

The classic failure is "node loads fine but sampler gives black video" - that's usually an unsupported weight path (GGUF is not supported here) or the wrong VAE, not the loader. If the node errors on a missing shard, either re-enable auto_download_missing_weights or finish the download with hf download meituan-longcat/LongCat-Video-Avatar-1.5 --local-dir ComfyUI/models/longcat/LongCat-Video-Avatar-1.5. And remember: official modes are bf16 by default - there's no FP8/FP16 runtime switch on this node, so don't go hunting for one.