Five Checks I Run Before Shipping Generative Images to Production

By Lin Mei Huang, Multimodal & Media AI Editor

Shipping a pretty demo frame is easy. Shipping a repeatable image pipeline that legal, brand, and support can live with is not. I treat generative images like any other production asset: versioned, reviewable, and killable. Here are the five checks I refuse to skip.

1. Rights and training-data assumptions

I write down, in plain language, what license covers the model outputs for our use case and what we are not allowed to do (trademark lookalikes, celebrity likeness, competitor packaging clones). If the answer is “the vendor said it is fine” without a linkable policy section, I stop.

I think “AI made it, so nobody owns it” is not a strategy. It is a future lawsuit with better lighting.

2. Consistency under seed and prompt drift

Brand work needs sameness. I keep a small golden set: five prompts, fixed seeds where the stack allows, and a side-by-side board. Any model or LoRA upgrade must beat or match the board on identity (logo geometry, palette, character design) before it touches production.

If your pipeline cannot reproduce yesterday’s hero image within tolerance, you do not have a pipeline. You have a slot machine.

3. Safety and policy filters at the edge

I run the same prompts through our blocked-topic list and through the vendor filters, then compare misses. Edge cases I care about: minor-adjacent school settings, medical skin imagery, election-related montages. False comfort from a single vendor toggle is common.

Document who can override a block and why. Overrides without audit trails become folklore.

4. Provenance metadata

I embed or attach: model name, version, prompt hash, seed, timestamp, operator id, and license tag. Support needs this when a partner asks “where did this asset come from?” six months later. Mystery PNGs in a Drive folder fail that conversation every time.

5. Human review gates by risk tier

Not every thumbnail needs a creative director. I tier:

• Low: blog decorative backgrounds → spot check sampling.
• Medium: product pages → designer review.
• High: campaign heroes, packaging, anything implying endorsement → legal + brand sign-off.

Automate the queue; do not automate the signature.

A note on “just upscale it”

Upscalers and face-fix tools introduce a second model’s failure modes. I treat them as separate stages with their own golden set. Chaining five generative tools without intermediate checks is how tiny artifacts become print disasters.

Bottom line

Generative media is not exempt from craft. The teams that look “fast” six months in are usually the ones who invested in boring preflight: rights, consistency, safety, provenance, review. Pretty pixels are the easy part. Trust is the product.