LongCat-Video-Avatar 1.6 for ComfyUI: bf16, int8, W6A8, W4A8
The DiT of LongCat-Video-Avatar 1.6 (Meituan
LongCat) as single ComfyUI files, for the native ComfyUI nodes of ComfyUI-LongCat-Avatar. The bf16 file is the
release's base_model/ shards merged into one file, keys unchanged. The three smaller files are ComfyUI's own int8
ConvRot, W6A8 and W4A8 layouts (Comfy Kitchen kernels), made from it by plain round-to-nearest: no GPTQ, no
calibration.
| File | Size | Layer output error (random inputs) |
|---|---|---|
longcat_video_avatar_1.6_bf16.safetensors |
29.5 GiB | - |
longcat_video_avatar_1.6_int8_convrot.safetensors |
14.9 GiB | 1.4% |
longcat_video_avatar_1.6_w6a8.safetensors |
12.3 GiB | 2.1% |
longcat_video_avatar_1.6_w4a8.safetensors |
8.8 GiB | 6.4% |
Block linears are in the named format, the adaLN modulation layers in int8 ConvRot, and the embedders, audio projection and final layer in bf16.
The int8 file is the one the nodes were tested with: on an RTX 4090 with --use-ck-attention, the distilled 8-step
mode runs at 5.1 to 5.7 s per step at 768x512, 93 frames. W6A8 and W4A8 have not been timed with the current nodes,
and no file's video quality has been measured against bf16 beyond the layer error above.
Use
The folders here mirror ComfyUI/models/. Take one transformer and the three files below it:
| Folder | File | Size | |
|---|---|---|---|
diffusion_models/longcat-video-avatar/ |
one of the four files above | pick it in Load LongCat Avatar Model | |
loras/longcat-video-avatar/ |
longcat_video_avatar_1.6_dmd_lora_bf16.safetensors |
1.55 GiB | the release's loraA/ (DMD), for the distilled node |
loras/longcat-video-avatar/ |
longcat_video_avatar_1.6_opsd_lora_bf16.safetensors |
1.55 GiB | the release's loraB/ (OPSD), for the distilled node |
audio_encoders/ |
whisper_large_v3_encoder_fp16.safetensors |
1.19 GiB | Whisper large-v3, encoder only |
The LoRAs are the release's, rewritten as standard LoRA files. The text encoder and VAE are not here: they are
Wan 2.1's (umt5_xxl, wan_2.1_vae.safetensors), byte-identical to LongCat-Video's.
License and credits
MIT, as the original. LongCat-Video-Avatar 1.6 is by the Meituan LongCat team; these files only change the storage format of its weights. The Whisper large-v3 encoder is OpenAI's, Apache 2.0, taken from the copy in the release.
Model tree for PedroMarinhoDev/LongCat-Video-Avatar-1.6-ComfyUI
Base model
meituan-longcat/LongCat-Video