Qwen3-TTS in ComfyUI: Multi-Speaker Dialogue, 3-Second Voice Cloning, and One Sampler You Have Not Seen
Voice generation finishing inside ComfyUI changes the shape of a video pipeline: audio and picture become one graph instead of two tools and an export. This release puts Qwen3-TTS in the graph, with cloning, multi-speake
Voice generation finishing inside ComfyUI changes the shape of a video pipeline: audio and picture become one graph instead of two tools and an export. This release puts Qwen3-TTS in the graph, with cloning, multi-speaker dialogue, and a lip-sync stage downstream.
The four graphs
| Workflow | Nodes |
|---|---|
| Qwen3 TTS preset voices + custom voice design | 11 |
| Qwen3 TTS voice cloning | 40 |
| Qwen3 TTS multi-speaker dialogue | 21 |
| LTX-2 lip-sync | 83 |
Voice cloning from three seconds
Three seconds is short. What it can capture:
- Timbre β the fundamental quality of the voice
- Rough pitch range β whether it sits high or low
- Speaking rate tendency
What three seconds cannot capture:
- Prosody across a long read β how the speaker handles paragraph-level emphasis
- Emotional range β three seconds is one mood
- Pronunciation of unusual words
So a cloned voice at three seconds is recognisably the same person on a similar sentence, and drifts on a different kind of sentence. That is the honest boundary.
The 11-node voice design graph
The smallest graph and arguably the most useful: preset voices plus custom voice design in eleven nodes.
Voice design means specifying a voice by description rather than cloning from a sample. For content where you need a voice that does not exist in your recordings, that is the path.
The 21-node multi-speaker graph
Multi-speaker dialogue is the interesting capability: two or more voices in one audio track, in a single generation.
That solves a real problem. Generating each speaker separately and concatenating produces clean audio with no conversational rhythm β each line starts and ends in isolation. A single generation with speaker assignment can produce overlap, backchannel, and natural timing.
The gradient_estimation sampler
The LTX-2 lip-sync graph uses KSamplerSelect β gradient_estimation.
That is not a sampler that appears in any other workflow covered in this series. It is a different solver family β gradient estimation rather than an Euler or ancestral variant.
Why it might be here: lip-sync is a constrained problem. The audio specifies the mouth timing exactly, so the sampler is solving for a video that matches a fixed temporal constraint. A gradient-estimation approach handles constrained optimisation differently from a stochastic sampler.
If you port settings from other LTX workflows into this one, the sampler will not behave the same way. Treat it as its own configuration.
The LTX-2 stack
| Node | Value |
|---|---|
CheckpointLoaderSimple |
ltx-2-19b-dev-fp8.safetensors |
LTXVAudioVAELoader |
ltx-2-19b-dev-fp8.safetensors |
LoraLoaderModelOnly |
ltx-2-19b-distilled-lora-384 at 1.0 |
EmptyLTXVLatentVideo |
768Γ512, 97 frames |
EmptyImage |
1280Γ720, colour 0
|
Note the checkpoint and audio VAE match β both ltx-2-19b-dev-fp8. That is the correct configuration, and it is worth noting given how often the mismatch appears elsewhere in this family.
EmptyImage at colour 0 is black β a placeholder, meaning the first frame is not being supplied from an image.
Where this breaks
- Three seconds is a narrow clone. Long reads will drift.
-
The
gradient_estimationsampler means settings from other LTX graphs do not transfer. - Qwen3-TTS and IndexTTS2 compatibility β the release compares them, but a comparison across two TTS engines needs matched text and matched speaker, which is easy to get wrong.
- LTX-2 (19B), not 2.3. Older generation.
- 97 frames is about four seconds of lip-sync per segment.
Practical approach
Start with the 11-node preset voice graph to confirm the TTS works and to hear what the presets sound like. Then clone one speaker and read a paragraph β that tells you how far cloning holds. Multi-speaker dialogue last, because it is the most complex to evaluate.
Workflows
Four graphs, published on RunningHub with invite code rh-v1111:
- Qwen3 TTS preset voices + custom voice design
- Qwen3 TTS voice cloning
- Qwen3 TTS multi-speaker dialogue
- LTX-2 lip-sync
Files
https://huggingface.co/AI-KSK/ksk-video-166-comfyui-workflows
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- The "3-second" cloning figure is the tool's stated minimum sample length.
- Independent AI-KSK release. Not affiliated with Alibaba, Lightricks or RunningHub.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.