Generate Ads
Base and trained Inkling each receive the same product brief and write a complete ad. Useful for discovery, but ad writing and expressive direction are confounded.
Can a model hear an ad—and teach the delivery back?
We trained a small adapter on the relationship between ad scripts, audible delivery, and expressive direction. The training pipeline worked. The quality question is still open.
The question, the pipeline, the evidence, and the honest answer so far.
01 / The question
Can delivery become reusable training data?
Great UGC ads do more than contain persuasive words. A creator rushes through a hook, lets a proof point breathe, smiles through a reveal, or lowers their voice before the close. Those decisions are audible, but most ad-writing systems throw them away.
This project asks whether Inkling—an audio-language model—can listen to those performance decisions, describe them as expressive text, and then learn to add useful direction to new scripts without hearing the original audio.
Audio in. Performance language out. Then teach the relationship back without pretending the model learned more than the evidence shows.The project in one sentence
Selection completeTraining completeCheckpoint savedQuality evaluation pendingFish listening test pending
02 / Pipeline
The experiment has four distinct jobs
Keeping the jobs separate matters. A transcript tells us what was said. Inkling interprets how it was delivered. A LoRA adapter learns a narrow transformation. Fish Audio eventually turns the resulting direction back into speech.
01 / ObserveWinning ad audio
A reviewer chooses an ad because its performance is worth studying—not because the model can see its business metrics.
02 / LabelExpressive text
Inkling produces word-preserving direction: pace, emphasis, pauses, emotion, and physical reactions.
03 / LearnRank-32 LoRA
The adapter trains on script and context → the same words with expressive direction.
04 / RenderFish Audio S2-Pro
The planned renderer speaks base and trained outputs with the same voice for a blind A/B test.
Important boundary
Inkling does not generate the final audio in this project. It produces text and expressive markup. Fish Audio is downstream.
03 / Evidence
What actually happened
The source library contains more than 3,000 collected UGC ads. Fifty were selected through a local media-review dashboard. Forty-two labels passed the final training checks; eight were quarantined instead of being silently forced into the dataset.
Source pool3,000+Locally held ads
Human-selected cohort50Performance examples
Accepted targets42Used in training
Quarantined8Kept out of the run
The accepted set contained 31 text-only rows and 11 audio-conditioned rows. Training ran in two rank-32 passes, each with 42 optimizer steps. Both passes completed and saved remote sampler and training-state checkpoints.
What this proves
The selection, labeling, rendering, training, saving, and inference plumbing can run end to end.
What this does not prove
A completed training run does not prove the model improved. That requires a controlled comparison on material the adapter did not train on.
04 / Answer so far
The LoRA is real. The verdict is not.
The adapter can be loaded and asked to create an ad or add expressive markup. But the original supervision mostly taught existing script → same script with delivery direction. Asking it to invent a complete campaign from a product brief is a broader task and therefore an exploratory test, not the cleanest measure of learning.
Compare Markup
Both receive one locked script. Spoken words must stay identical; only delivery direction may change. This is closer to the training objective.
The next meaningful result is not another screenshot of text. It is a blind same-voice listening test in Fish Audio, followed by an untouched evaluation set large enough to resist anecdotal conclusions.