Open ComfyUI workflow turns text into a lip-synced talking video with Drama 3 and MiniMax H3
A free ComfyUI workflow from Skysay chains Fish Audio Drama 3 for expressive speech with MiniMax H3 for a lip-synced talking video, all on your own GPU. It ships as MIT-licensed graphs plus a script that runs both stages.
What it does
Stage A takes your text and a reference voice through Fish Audio Drama 3 (via the ComfyUI-Drama3 nodes) to make an expressive take. Drama 3 is then unloaded, so the two models are never in memory at the same time. Stage B feeds a face image and that take into MiniMax H3's reference-to-video node, anchors the audio at frame 0 for the lip sync, and muxes the video with the original take.
The repo has the combined graph, the two stages as separate graphs, and a Python script that runs stage A, frees memory, then runs stage B on a ComfyUI server.
Results and limits
On a 96 GB RTX PRO 6000, the speech stage took 18 to 42 seconds and the video stage about 130 to 190 seconds for a 7 to 9 second clip. The demo uses a synthetic face and the public-domain LJ Speech voice as the reference; there are no numbers for smaller cards yet.
Worth knowing
It needs ComfyUI 0.40 or newer for the core MiniMax H3 nodes, plus the ComfyUI-Drama3 custom nodes. The README asks you to read both model licenses first: the Fish Audio Community License allows free commercial use only under US$500k revenue. Users in the EU, UK, US and South Korea need to request a license from MiniMax via the H3 License Request Form (platform.minimax.io/h3-license) before use. Only use faces and voices you have consent for.
Try it yourself
Set B5_TEXT to "[laughing] I can't believe that worked. [whispering] Don't tell anyone.", B5_FACE to [a face image you have consent to use] and B5_PROMPT to "[character] talking to the camera in [setting], natural head movement", then run b5_wf.py against your ComfyUI server.