(For SOTAI Squad) Stop Paying for AI Avatars! Build Your Own with LongCat Avatar (ComfyUI Workflow)
Added 2026-01-07 15:19:19 +0000 UTCLink video: https://youtu.be/VBa1ii-DH2c
In this video, I will show you exactly how to build the LongCat Avatar workflow to create natural lip-sync videos that can last up to several minutes! π€―
Forget about the short 5-second limit of other tools. With this ComfyUI setup, we are unlocking the ability to generate long-form talking head videos with incredible realism. I will also share my personal tips and tricks to optimize your results and speed up the rendering process.
βοΈ Build Workflow
Step 1 - Install the custom node
Install WanVideoWrapper node: https://github.com/kijai/ComfyUI-WanVideoWrapper
Search in ComfyUI Manager or instal via Git clone and pip install
Move to custom_node/WanVideoWrapper folder: git switch longcat_avatar
Install ComfyUI-MelBandRoformer: https://github.com/kijai/ComfyUI-MelBandRoFormer
Install ComfyUI-Logic: https://github.com/theUpsider/ComfyUI-Logic
Search in ComfyUI Manager or instal via Git clone
Step 2 - Download the model and place it in the correct folder
Place the file in: "models/unet/LongCat-Avatar-single_fp8_e4m3nf_scaled_mixed_KJ.safetensors"
Place the file in: "models/loras/LongCat_distill_lora_alpha64_bf16.safetensors"
https://huggingface.co/Wan-AI/Wan2.1-T2V-14B/blob/main/Wan2.1_VAE.pth
Place the file in:
"models/vae/Wan2.1_VAE.pth"
https://huggingface.co/Kijai/WanVideo_comfy/blob/main/umt5-xxl-enc-fp8_e4m3fn.safetensors
Place the file in:
"models/text_encoders/umt5-xxl-enc-fp8_e4m3fn.safetensors"
https://huggingface.co/Kijai/MelBandRoFormer_comfy/tree/main
Place the file in: "models/diffusion_models/MelBandRoFormer_comfy/MelBandRoformer_fp32.safetensors"
Please check the workflow in the attachments.
Comments
Hello, based on the error you described, I suspect that your prompt may not match the content of the image you uploaded. Please double-check and make sure you havenβt modified anything in the workflow other than the inputs. If the prompt does match the input image, please send me the workflow youβre using along with the input files so I can take a closer look. You can email them to me at sotai.work@gmail.com
SOTAI
2026-02-04 04:22:57 +0000 UTCHi! I found this tutorial via YouTube search. I have it running on my machine and it's working (kinda). It doesn't really follow my input image at all. On the output video I see the initial frame is my input image but the next frame is a totally new image that seems like it is generated entirely from the text prompt. I didn't make any changes to your example JSON with the exception of the input image, input sound file, prompt and the audio duration value (left side of workflow) to match my audio file length which is 15s. Any thoughts?
SinthCore
2026-02-03 15:17:20 +0000 UTC
