Case Study: GenAI Finetuning at YouTube

The hurdle

YouTube’s DreamScreen let creators generate any image from a text prompt. But a blank prompt box is just another blank canvas: most users knew a “good” output when they saw it, but they could not write the input needed to generate a “good” output in a prompt.

To complicate things further, the base model is inconsistent by nature: without style conditioning, the same prompt could swing from photoreal to cartoonish to off-model on a different run. Even the best prompt engineers struggled to maintain consistency between iterations on nearly identical prompts.

Better UI wouldn't fix inconsistent outputs, and a better model alone wouldn't fix the blank-canvas problem. We needed a product layer — curated, selectable styles — that removed prompt engineering entirely while guaranteeing a consumer-grade quality bar.

The approach

Models that create high-quality images require high-quality training data. Getting there meant treating visual craft as a first-class input, not an afterthought — and staffing the team accordingly.

I tapped an Art Director to own the creative quality bar and translate it into a spec the model could actually learn from — someone who could review outputs against craft standards, and adjust direction in real time as we saw what the model responded to. Then, I engaged creative agency Anyways, whose global network of illustrators gave us production capacity and stylistic range we couldn't have built in-house fast enough. Together, we commissioned multiple style datasets — oil pastel, modern polygon, anime, pixel art, and more — each with 70+ bespoke illustrations spanning a diverse range of people, places, and things. The Art Director set the bar; the agency delivered the scale to hit it. That let the core team stay focused on fine-tuning and product while the training data was engineered with equal rigor.

The result

The presets hit a high visual bar while generating diverse, prompt-aligned outputs at low latency. Six weeks post-launch, over [redacted]% of published generations used our fine-tuned presets — proof that users chose curated consistency over open-ended prompting, and that the dataset investment paid off in the model itself.

The below snapshot illustrates why. The base model (far left) defaults to a literal, photorealistic reading of the prompt — a person in a dinosaur costume on a real bike, poorly framed, only slightly recognizable as SF. It has no visual identity: no intentional color, linework, or craft signature, so it reads as a novelty prop rather than a usable creative asset. The four fine-tuned styles, by contrast, each apply a distinct, recognizable aesthetic that a creator can trust and drop straight into finished content and, just as importantly, continue to create with a consistent look and feel.