v1.2.1

Podcast speaker assignment fix for #3.

• Added generation_mode to the Podcast node. The new default, per_turn, generates each turn using only its named speaker's reference and joins the audio.
• Repeated speakers reuse their own encoded reference, including when different recordings have identical transcripts. Long turns split automatically with a separate max_steps limit per fragment.
• native preserves the original continuous multi-speaker generation. Per-turn generation resets context between turns, so conversational pacing can change. Reference quality still affects voice similarity.
• Updated English/Chinese documentation, tooltips and the podcast example workflow. Existing workflows remain compatible.

Validation: 41 regression tests and all 11 example workflows passed. Isolated original Tiny and Tiny INT8 controls followed speaker 1 -> 2 -> 1 and retained all dialogue words. Native-mode parity checks matched v1.2.0 exactly.

After updating, restart ComfyUI and use generation_mode=per_turn in the Podcast node. No model download or conversion is required.

v1.2.0

What's changed

Long voice-design and freeform text now generates in short fragments instead of one continuous decode, addressing the runaway repetition reported in #2.

• Voice design and freeform use up to 24 characters or words per fragment, with a separate max_steps limit for each fragment.
• Later fragments reuse the first fragment's speaker embedding, audio prompt latent and transcript to retain the designed voice.
• Generated references exclude trailing silence even when output trimming is disabled. Alignment padding goes before the speech so generation history ends with speech.
• Original style instructions and user-provided reference inputs are preserved.
• Updated English and Chinese documentation, cookbook instructions, tooltips and example-workflow notes.

Other single-speaker tasks retain their existing 50-unit splitting limit. Podcast remains one continuous generation.

Validation

33 regression tests, 11 workflow checks and upstream prompt comparisons passed. Isolated GPU checks passed on 0.5B BF16, 0.5B INT8 ConvRot and Tiny INT8 ConvRot with AIMDO enabled and 10 flow steps. Short voice-design and freeform outputs remained bit-identical to v1.1.0.

A separate 411-character Japanese voice-design test completed as 65 seconds of audio across 22 fragments without reaching a fragment's step limit. Japanese speech accuracy still varies with the text, voice description and seed.

Full changelog: v1.1.0...v1.2.0

v1.1.0

What's changed

• Long text is automatically split for voice cloning, built-in voices (ip), basic attributes, emotion and dialect tasks. Each fragment gets a fresh generation cache and its own max_steps limit before the audio is joined.
• Splitting preserves vocal tags, punctuation, whitespace, accented words, numbers and dotted initials. Long unpunctuated Japanese and Korean text is split correctly.
• Reference audio is encoded once per request, and the seed continues across fragments for reproducible generation.
• Persistent terminal progress bars show the current fragment number.
• Each audio fragment is checked for invalid values before silence trimming.
• English and Chinese documentation and the max_steps tooltip explain the new behavior.

Voice design, freeform and podcast inputs remain a single generation. Short inputs are preserved unchanged.

Validation

28 regression tests and 11 workflow checks passed. Isolated GPU checks passed on both 0.5B INT8 and Tiny 9B INT8 with AIMDO enabled; short built-in-voice and cloning outputs were bit-identical to the previous runtime. Multi-fragment cloning smoke tests also passed.

Full changelog: v1.0.0...v1.1.0

v1.0.0

v1.0.0

Initial release of Ming 3.0 TTS for ComfyUI.

• Support 0.5B and Tiny 9B models with original merged and INT8 ConvRot weights.
• Voice cloning, voice design, built-in voices, freeform direction, speed/pitch/volume, emotion and dialect controls.
• Tiny multi-speaker podcasts with separate reference voices and transcripts.
• Local Whisper transcription connected to cloning and podcast workflows.
• Optional FLA acceleration for Tiny, with a PyTorch fallback and backend reporting.
• Model downloads, extra model paths, symlinks and Windows junction support.
• 11 example workflows with Markdown guides, English/Chinese READMEs and upstream cookbooks.
• Setting tooltips, seed 42 defaults and persistent generation progress.

Download the models · Setup and usage