InfiniteTalk | Image-to-Speech AI Workflow (All Files + Setup Guide Included)
🗣️ InfiniteTalk Workflow – Make Any Image Speak!
This setup turns a single portrait or character image into a talking AI avatar using the InfiniteTalk system.
It includes everything you need for image-to-speech, from text-to-voice generation to synced lip animation.
✨ Features
• Fully working InfiniteTalk pipeline
• Converts any image → realistic talking video
• Built-in voice generation (Text-to-Speech)
• Lip-sync and animation nodes included
• Ready-to-export video output
⚙️ Requirements
All required components and Hugging Face model links are listed directly inside the workflow.
🔗 When you open the workflow, follow the notes in the comments — they show exactly where to download the model files and where to place them.
💾 Download the Workflow
[Insert your workflow link here]
🧠 How to Use
Open the workflow JSON in ComfyUI
Drop your portrait into the image input
Enter your dialogue or script text
Run the workflow and export your talking video
🎯 Perfect for AI creators, VTubers, storytellers, or anyone who wants to bring portraits to life!
💬 Join the Community https://discord.gg/qGSnE27NQn
Description
FAQ
Comments (6)
Hay so I managed to get this working locally on my 16GB card and I'm surprised how good it works at 480x832. However, it cuts at 40 seconds and I can't spot anything that indicates why that would be. Apologies if this is a stupid question, it's my first time using this. I can absolutely figure ways to cut up audio and do last-frame-in loops if that's necessary. But figured I'd ask just in case I missed something obvious.
can i put other action movement loras in this workflow without breaking it,
For what it's worth, I have added at least one LoRa and it still works well - I usually use a face LoRa, but I used an action LoRa and it worked.
I've been using this workflow for a bit and I am ery pleased with it. In particular, it is very good with close up lip sync, though less so at more distant images.
I am wondering if there is a variant without the gguf files and if it would be more accuate.
Regardless, this workflow is very solid. WIth an 5090 and 64GB ram I have managed to generate a 37 second video that looked very, very good.
Thanks for the hard work!
what is the audio input needed for? do you just put any audio?
Thank you very much for this interesting workflow; hdean recommended it to me.
However, I'm having a problem with it. The WanVideo Model Loader can't find wan2.1-i2v-14b-480p-Q4_0.gguf, even though the file is in the correct folder. I tried switching to Safetensors, but that isn't accepted: WanVideoModelLoader.loadmodel() got an unexpected keyword argument ‘lora’
Since I'm pretty new to this, I'm not sure what to do next.