StableAvatar_LoadModel

StableAvatar is an audio-driven avatar pipeline - feed it a photo of a person plus an audio file and it generates a talking-head video that matches the lips, indefinitely long. The whole thing ships as three nodes, and this is the one that loads the models. StableAvatar_LoadModel is the entry point: point it at a transformer, a VAE and an optional LoRA, and it hands you the two wires that the other two nodes actually run on.

The underlying model is StableAvatar (arXiv 2508.08248), which is built on Wan 2.1's 1.3B "Fun-InP" transformer - the lightweight, low-VRAM tier of Wan rather than the 14B that everyone actually uses for general video. That matters: you're getting the speed and VRAM footprint of the little Wan, at the cost of the 14B's quality. For lip-sync close-ups that trade is usually fine, but go in with eyes open. The pack is a wrapper by smthemex, a prolific author who ports talking-head and portrait models into ComfyUI quickly (EchoMimic, SenseNova U1, TwinFlow and a dozen more come from the same person).

What it actually loads

Internally the node reads three things out of your ComfyUI/models/ tree:

• the transformer - diffusion_pytorch_model.safetensors plus config.json from ComfyUI/models/StableAvatar/transformer/. The README stresses the filename: it must stay exactly diffusion_pytorch_model.safetensors or the loader won't find it.
• the VAE - the standard Wan2.1_VAE.pth from ComfyUI/models/vae/.
• an optional LoRA - Kijai's lightx2v speed LoRA from ComfyUI/models/loras/. It cuts you from 25–50 steps down to 10, at the usual speed-LoRA quality cost.

Outputs are two custom types that only exist inside this pack: model (MODEL_PIPE_SA) feeds the sampler, and info (MODEL_INFO_SA) feeds the predata node. If you see "model type not found" errors, it means the pack isn't installed - these aren't standard ComfyUI types.

The inputs that actually matter

A beginner sets three things:

• transformer / vae - dropdowns, populated from your model folders. If a dropdown is empty, the file is in the wrong place.
• lora - leave None for your first runs. Add lightx2v once it works, and drop steps to 10.
• GPU_memory_mode / use_mmgp - the memory story. use_mmgp presets (LowRAM_LowVRAM, HighRAM_HighVRAM, …) pick a GPU/RAM offload balance; GPU_memory_mode offers model_cpu_offload and sequential_cpu_offload plus model_cpu_offload_and_qfloat8 (fp8 weights). The README's tip: setting GPU_memory_mode to None disables mmgp's fp8 quantization. There are two offload systems here and the author recommends mmgp. If you have ~12GB+ VRAM, try None first and only reach for offload when it OOMs.

enable_teacache defaults to false - keep it that way. The README is blunt: don't enable teacache yet, it makes things "花" (breaks/bloats). weight_dtype (bfloat16 default) you can ignore until you're chasing NaNs.

Installing it

Via ComfyUI Manager, search "ComfyUI_StableAvatar". Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_StableAvatar.git
cd ComfyUI_StableAvatar
pip install -r requirements.txt

Then restart ComfyUI. The heavy lift is models, not code - you need the Wan transformer + config, Wan VAE, a umt5_xxl text encoder in models/clip, a CLIP vision model in models/clip_vision, wav2vec2-base-960h in models/StableAvatar/, and Kim_Vocal_2.onnx for vocals. The audio-separator pip package is optional and only needed for separating vocals out of songs. One nice touch from the README: if you also run EchoMimic V3, this pack detects it and reuses its shared models instead of double-downloading.

Common tripping points: wrong folder → empty dropdown; renamed transformer file → load failure; and assuming the pack's custom node types are standard ComfyUI. All three are just "check your model paths" in disguise.