Drama 3 Generate:
Drama3Generate creates expressive speech and multi-speaker dialogue from written scripts using Drama 3. It supports spoken words alongside performance cues such as sighing and laughing, helping you produce character conversations, narrated scenes, and other voice performances. Reference voices can help maintain a recognizable voice for each character, while different dialogue-generation methods let you prioritize voice separation, conversational responsiveness, or continuity across a scene. Long passages can be generated in smaller sections, using selected earlier material as context to help maintain continuity. The node also offers controls for exploring different performances and an optional voice-separation technique for dialogue where characters might otherwise sound too similar. It performs the speech-generation stage of the workflow; a separate decoding stage turns the generated representation into playable audio.
Drama 3 Generate Input Parameters:
model
A required DRAMA3_MODEL connection containing the Drama 3 model used to generate speech. Connect the model supplied by Load Drama 3, optionally passing it through Drama 3 Speed Options first. This model provides the speech-generation capabilities and configuration used throughout the process, and its processing requirements affect available memory and execution speed. The node loads the model through ComfyUI’s model-management system before generation. There is no numeric range or default value for this connection.
text
A required multiline STRING containing the words and performance instructions you want spoken. Place expressive tags such as [sighs] or [laughs] where the corresponding sound should occur, and use speaker markers such as <|speaker:1|> to switch characters. The speaker number associates that part of the script with the matching numbered voice slot; for example, speaker 1 uses voice1. Line boundaries are important when generating separate takes for dialogue, so put individual turns on separate lines for a clear script structure. Longer text can be divided into smaller generation chunks depending on the other settings. The schema specifies no default text or explicit text-length limit.
voices
A required COMFY_AUTOGROW_V3 input group that holds dynamically added DRAMA3_VOICE connections named with the voice prefix. The group allows zero to 16 voice connections, so the group must exist in the interface, but you do not have to supply a reference voice. Each connected voice supplies reference speech codes and a transcript that help establish how its corresponding speaker should sound. Slot numbering matters: voice1 corresponds to <|speaker:1|>, for example. An untagged reference transcript is assigned to its slot’s speaker number, and unconnected entries are ignored. When a speaker has no supplied reference, earlier generated speech from that speaker can provide context for later takes. There is no numeric default; the applicable limits are a minimum of zero and a maximum of 16 voice connections.
seed
A required INT that initializes the random choices used during speech generation. Default: 0; minimum: 0; maximum: 18446744073709551615. Keep the seed and the rest of your workflow unchanged when comparing repeated runs, or change the seed to explore alternative deliveries of the same script. A fixed seed is not a promise of identical audio after changing generation settings: voice guidance and the speed options explicitly can change what a seed produces. The interface enables ComfyUI’s after-generation seed control, allowing you to choose how the seed behaves between executions.
temperature
A required FLOAT controlling the randomness of speech-code selection. Default: 1.0; minimum: 0.0; maximum: 2.0; adjustment step: 0.05. Lower values concentrate generation on the candidates the model ranks most highly, while higher values allow more variation. For this implementation, lower temperature is not necessarily a safer or more reliable speech setting: the source explicitly warns that lower values make takes fall silent more often. A temperature of 1.0 together with top_p 0.9 matches the defaults identified for Fish Audio’s Drama 3 server and Space, making that combination a sensible starting point.
top_p
A required FLOAT that limits sampling to a set of candidates whose combined probability reaches the chosen threshold. Default: 0.9; minimum: 0.05; maximum: 1.0; adjustment step: 0.05. In practical terms, lower values narrow the range of choices available during speech generation, while values closer to 1.0 allow a broader set of possibilities. It works alongside temperature and top_k, so changing it can alter delivery and the generated take even when the seed stays the same. The documented starting combination is top_p 0.9 with temperature 1.0.
top_k
A required INT limiting how many of the highest-ranked candidates remain available during sampling. Default: 0; minimum: 0; maximum: 4096. A value of 0 disables this particular candidate-count restriction; other sampling controls still apply. Positive values restrict selection to the specified number of candidates, but a small value can be counterproductive for Drama 3: the source explains that a few silence codes can outrank the many different ways speech could resume after a pause. Its tooltip reports that top_k 30 left 11 of 40 tested takes silent. Keep 0 as the starting value unless you have a specific reason to experiment.
max_frames
A required INT setting the maximum number of generated frames per text chunk, rather than for the entire script. Default: 0; minimum: 0; maximum: 8192. There are approximately 21.5 frames per second of audio, so a manual cap of 215 frames corresponds to roughly 10 seconds per chunk. Generation can finish earlier when the model reaches the end of speech. A value of 0 enables automatic sizing: each chunk receives a cap of 150 frames plus 4 frames per byte of its UTF-8 text, described in the tooltip as more than twice what speech usually takes. Because the calculation uses bytes rather than visible characters, different writing systems can produce different automatic caps for similar-looking text lengths. A manual cap that is too small can cut speech short.
split_long_text
A required BOOLEAN controlling whether long passages are divided into chunks of about 300 bytes. Default: True; options: True or False. When enabled, generation uses smaller text sections and selected earlier takes as context, following the continuation approach described for Fish Audio’s server. The implementation generally retains the first and latest relevant takes rather than feeding the entire generated history into every subsequent prompt; the exact context also depends on the speaker mode. Disabling this setting removes the approximately 300-byte chunking limit, but it does not override the separate-take behavior of the selected speaker mode. Keep it enabled as a practical starting point for long scripts.
speakers
A required COMBO choosing how the script is organized into generated takes. Default: one take per line; the modes documented in the provided source are one take per line, one take per line, hearing the line before, one take per speaker, and together. With one take per line, each line is generated separately using that speaker’s references and relevant earlier takes, helping prevent voices from bleeding into one another, but without hearing other characters’ lines. With one take per line, hearing the line before, the preceding take can also provide context when it belongs to another speaker, allowing more conversational response at the risk of voice blending. With one take per speaker, each speaker’s lines are grouped into a take and then divided at generated line markers to restore script order; missing markers can leave lines joined, missing, or out of order. With together, dialogue is generated as a shared script so characters can hear one another within the generated passage. Separate takes are decoded independently by the downstream decoder.
voice_guidance
A required FLOAT that helps keep reference voices distinct in the supported dialogue modes. Default: 0.0; minimum: 0.0; maximum: 3.0; adjustment step: 0.25. It applies to together and one take per line, hearing the line before rather than the other speaker modes. The technique generates a contrasting version conditioned on another speaker’s voice and steers the intended take away from that alternative; it therefore needs suitable reference voices, with a voice for each speaker recommended for the documented dialogue use case. A value of 0.0 disables guidance, and the tooltip recommends 1.0 as a useful starting point. Higher values are not automatically better: the reported tests found little additional benefit at 2.0 and missing words in two takes. Guided takes take approximately twice as long to render and change what a given seed produces. The feature requires one GPU and ComfyUI after v0.41.2; when the implementation detects an unsupported configuration, it logs a warning and continues without guidance.
Drama 3 Generate Output Parameters:
DRAMA3_CODES
A DRAMA3_CODES connection containing the generated speech in the encoded representation used by the Drama 3 workflow. This is an intermediate result, not a playable waveform or an audio-file path. The implementation organizes generated audio codes with information about takes and script lines so that downstream processing can handle separately generated dialogue and restore script order where applicable. Connect this output to Drama 3 Decode, together with the appropriate codec, to obtain 44.1 kHz audio. The provided decoder processes separately generated takes individually and joins them in script order with 0.3 seconds of silence between takes; its optional pause-shortening setting can further change the final timing. Consequently, the duration and spacing you hear depend on both generation and decoding settings.
Drama 3 Generate Usage Tips:
- Start with temperature 1.0, top_p 0.9, top_k 0, max_frames 0, and split_long_text enabled. These settings preserve the documented sampling defaults and let the node size generation limits automatically.
- Format dialogue with one turn per line and explicit speaker markers, such as
<|speaker:1|>Hello.followed on the next line by<|speaker:2|>Good to see you.Connect each reference voice to the slot matching its speaker number. - Use
one take per linewhen distinct character voices are more important than characters reacting to the delivery of the preceding line. It keeps other speakers’ generated takes out of each line’s prompt. - Try
one take per line, hearing the line beforeortogetherwhen you want more conversational context. If voices begin to blend, supply references for each speaker and test voice_guidance at 1.0. - Keep voice_guidance at 0.0 when you do not need it. It is not active in every speaker mode, and guided takes can take about twice as long to generate.
- Place performance tags exactly where you want the event to occur, for example
I thought we were finished. [sighs] Let’s try again.Listen to the result rather than assuming every cue will be realized identically across seeds. - Leave max_frames at 0 until you have a reason to impose a duration ceiling. If you set it manually, estimate approximately 21.5 frames per second for each chunk and leave enough room for pauses and expressive delivery.
- Keep split_long_text enabled for long passages. Its approximate 300-byte chunk size is not a 300-character limit, especially for text containing multibyte characters.
- Change only one setting at a time when comparing performances, and keep the seed fixed. Remember that changing speed options or enabling voice guidance can alter the take despite an unchanged seed.
- Connect the result to Drama 3 Decode before evaluating silence or timing. For an untrimmed listening comparison, disable that decoder’s shorten_long_pauses option, which otherwise shortens long pauses and trims leading and trailing silence.
Drama 3 Generate Common Errors and Solutions:
Drama 3: voice_guidance needs one GPU and ComfyUI after v0.41.2; generating without it.
- Explanation: This is an explicit warning from the source. The current execution configuration does not support voice guidance, so generation continues with guidance disabled.
- Solution: Use a single-GPU configuration and update ComfyUI to a version after v0.41.2 if you need guidance. Otherwise, set voice_guidance to 0.0 and continue without it.
Drama 3: speaker %d's take marked %d of its %d line breaks, so its last %d lines are joined or missing, out of script order; another seed or one take per line avoids this.
- Explanation: This is the source warning template; actual logs substitute numbers for the placeholders. In
one take per speakermode, the generated take did not include enough line markers to reliably separate all of that speaker’s lines and restore script order. - Solution: Try another seed or switch to
one take per line. Listen through the affected section and check for joined, missing, or misplaced lines before accepting the result.
Silent or mostly silent takes
- Explanation: This is a possible result rather than a quoted exception. The supplied tooltips specifically identify low temperature and restrictive top_k sampling as causes of increased silence: highly ranked silence candidates can crowd out the many possible speech continuations.
- Solution: Restore temperature to 1.0, top_p to 0.9, and top_k to 0, then try another seed. Confirm that the script contains words to speak and evaluate the decoded audio.
Speech ends before the script is finished
- Explanation: This is an output symptom, not a documented error string. A manually chosen max_frames limit may stop a chunk before all its speech has been generated. The model can also end a take before reaching the cap.
- Solution: Set max_frames to 0 for automatic sizing or increase the manual limit. Keep split_long_text enabled for long passages, and try another seed if speech still ends early.
A character sounds like another speaker
- Explanation: This is a dialogue-quality issue rather than a runtime exception. Hearing another speaker’s previous take or generating dialogue together can allow reference voices to blend; incorrect speaker-to-slot assignments can also condition a line on the wrong reference.
- Solution: Verify that each speaker marker matches its intended voice slot. Use
one take per linefor stronger separation, or supply a reference for each speaker and try voice_guidance 1.0 in a supported contextual dialogue mode.
Increasing voice_guidance does not change the result
- Explanation: Guidance is deliberately disabled outside its supported speaker modes. It also needs an applicable contrasting reference voice; a positive value alone does not guarantee that a guided comparison is available.
- Solution: Select
togetherorone take per line, hearing the line before, supply distinct references for the speakers, and check the console for the compatibility warning. Start at 1.0 rather than immediately increasing the value.
Missing words after increasing voice_guidance
- Explanation: This is a possible quality issue documented in the tooltip’s test results, not a guaranteed failure. Stronger guidance can affect speech content, and 2.0 provided little extra voice-separation benefit while costing words in two tested takes.
- Solution: Reduce voice_guidance toward 1.0 or disable it, then compare another take. If voice separation remains difficult, use
one take per lineinstead.
CUDA out of memory or another device-memory failure
- Explanation: The supplied excerpt does not show a specific memory-error string, but generation loads the model and allocates session resources for reference material, text, generated frames, and concurrent work. Insufficient GPU memory can therefore prevent execution.
- Solution: Close other GPU-heavy tasks, keep long-text splitting enabled, and test a shorter script with fewer reference connections. If you use a large manual max_frames value, reduce it appropriately; disable voice guidance while checking available capacity.
The output cannot be played as audio
- Explanation: This is a workflow mismatch, not a generation error. DRAMA3_CODES contains encoded speech data rather than the AUDIO representation required for playback or saving as an audio file.
- Solution: Connect DRAMA3_CODES to Drama 3 Decode and provide its codec connection. Send the decoder’s AUDIO output to your audio playback or saving node.
