David Comfort tests a 1:42 AI dialogue built from four generations

• David Comfort assembled a 1:42 conversation from four 26-to-28-second generations, using one 24-line ElevenLabs master track, according to his breakdown.
• Pitch stayed within about 3 Hz of the master on nearly every line, and timestamped prompts brought timing within about half a second, he reported.
• The final word remained inaudible, and Comfort said the approach “still didn’t work that well”.

Comfort used a black video carrying the dialogue as a reference input alongside a timed transcript. Using a silent acting clip as a video reference instead produced weak lips and a stiff, copied performance.

Comfort laid out the build in his production notes:

1. Simplify the set. Rebuild the scene in a basement with an orange-moon window, an oil lamp and a frosted empty tank.
2. Make the dialogue master. Generate all 24 lines in ElevenLabs before building the video.
3. Rework the script. Apply techniques from a reference clip and divide the scene into four generations of 26 to 28 seconds.
4. Package each generation’s references:
5. Assemble the cut. Edit to 1:42 with steady room tone, three very quiet foley hits and no music.

ElevenLabs’ Text to Dialogue API accepts text/voice-ID pairs and returns audio. Its dialogue documentation notes that several generations may be needed to get the desired result.

Comfort reported four improvements in the breakdown:

• Pitch: Nearly every line stayed within about 3 Hz of the master track.
• Timing: Timestamps for both dialogue and shots brought timing within roughly half a second.
• Reference inputs: The black video plus transcript worked well on generations one and three.
• Continuity fixes: Prompt changes addressing a duplicate Zoe, plain skin and an empty tank landed.

The master track earns its keep here.

Comfort’s failure list covered generation, performance and the final audio mix:

• The first seven-generation cut: A duplicate Zoe, cuts every two seconds, freezes after cuts and overly loud foley.
• The silent acted reference: Weak lips, a copied and stiff performance, and a drifting look.
• A mistaken voice diagnosis: What he initially read as a timbre shift in a spectral measure was actually “s” and “z” sounds.
• Mastering problems: The quiet opening was being boosted by 35 dB, while the room-tone loop swelled.
• The missing last word: Generation four’s final “You.” remained inaudible. He estimated about $1.30 for a short extra generation or $3.60 to reroll the whole fourth generation.
• Speech pace: Seedance set its own pace by a second or two, leaving long holds that needed trimming in the edit.

Separately, Comfort shared “WE WERE ALWAYS BLUE”, captioned “Funny video.” He also said he had built agents to QA and change every shot in another project, discussing composition and lighting.

Comfort’s separate composition and lighting experiment