1038lab/ComfyUI-QwenTTS
1
0
Fork 0
mirror of https://github.com/1038lab/ComfyUI-QwenTTS.git synced 2026-11-04 23:37:45 +00:00
Code Issues 21 Projects Packages Wiki Activity

[GH-ISSUE #8] why its not realtime and low latency as advertised by qwen tts? i am using 4090 #8

New issue
Open
opened 2026-10-30 21:03:22 +00:00 by bsakaguchi · 1 comment
bsakaguchi commented 2026-10-30 21:03:22 +00:00
Owner
Copy link

Originally created by @HuntingSuccubus on GitHub (Jan 26, 2026).
Original GitHub issue: https://github.com/1038lab/ComfyUI-QwenTTS/issues/8

Originally created by @HuntingSuccubus on GitHub (Jan 26, 2026). Original GitHub issue: https://github.com/1038lab/ComfyUI-QwenTTS/issues/8
bsakaguchi commented 2026-10-30 21:03:25 +00:00
Author
Owner
Copy link

@1038lab commented on GitHub (Jan 30, 2026):

Qwen3-TTS can support streaming internally, but ComfyUI nodes are typically “batch”: they compute first, then output the final AUDIO. So you won’t hear audio “as it’s being generated” unless ComfyUI itself supports incremental audio playback for that node.

That said, you can still get much lower latency:

  1. Use 0.6B model for faster generation.
  2. Set do_sample = False (more stable + usually faster).
  3. Reduce max_new_tokens (this is the #1 latency driver). Try 512–1024.
  4. On CUDA, install FlashAttention 2 (optional): 
    pip install flash-attn --no-build-isolation
  5. First run is always slower due to model load / warmup.

If you share:

  • which model (1.7B CustomVoice / Base / VoiceDesign)
  • your max_new_tokens and text length
    I can suggest a good “speed preset”.
<!-- gh-comment-id:3821874024 --> @1038lab commented on GitHub (Jan 30, 2026): Qwen3-TTS can support streaming internally, but **ComfyUI nodes are typically “batch”**: they compute first, then output the final `AUDIO`. So you won’t hear audio “as it’s being generated” unless ComfyUI itself supports incremental audio playback for that node. That said, you can still get *much* lower latency: 1) Use **0.6B** model for faster generation. 2) Set **do_sample = False** (more stable + usually faster). 3) Reduce **max_new_tokens** (this is the #1 latency driver). Try 512–1024. 4) On CUDA, install **FlashAttention 2** (optional): ```bash pip install flash-attn --no-build-isolation ``` 5) First run is always slower due to model load / warmup. If you share: - which model (1.7B CustomVoice / Base / VoiceDesign) - your `max_new_tokens` and text length I can suggest a good “speed preset”.
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
No results found.
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
1038lab/ComfyUI-QwenTTS#8
No description provided.