Keep an AI production notebook: costs, source material and privacy belong beside the prompt

Treat the generated asset as one part of a project

A musician might use a generated background for a lyric graphic; an artist might explore layouts before photographing an actual work. The practical question is how to keep those experiments understandable when a collaborator or customer asks what was made and how.

This desk review uses Adobe's documentation and C2PA's provenance explanation checked September 19, 2026. It does not test model quality or certify an output's rights.

Create a short production note before generating: the intended use, the material supplied, the tool and model selected, and the part you expect to replace or revise yourself. Keep original artwork and recordings intact.

The finished package should make it possible to distinguish your source material, generated additions and later human editing without reconstructing the process from memory.

Read the model and plan together

Adobe's Firefly plan comparison distinguishes standard and premium features, monthly generative credits, and access to partner models. Its listed Standard allocation is 2,000 credits per month and Pro is 4,000.

A credit allocation is not a guaranteed number of finished client assets: different operations and models can consume different amounts, and discarded attempts still affect a working budget.

The price page returned Swiss-franc pricing in this review, so we are not presenting it as a verified U.S. checkout quote. Before subscribing, select the correct region and billing term, check renewal conditions and identify the features included in that exact plan.

A limited trial or promotional allowance should be entered separately from the recurring production cost.

Our suggested cost measure is the amount spent to finish one usable deliverable, including revision time. Set an experiment allowance, stop when it is reached and review why attempts were rejected.

That makes a tool comparison more informative than counting how many images it can produce.

Training assurances are narrower than a complete privacy promise

Adobe's Firefly FAQ says its current Firefly generative models were trained using licensed and public-domain material, and says it does not train on Creative Cloud subscribers' personal content. These are Adobe's statements about its models and practices.

Do not automatically extend them to every partner model available inside the same interface.

Adobe's privacy policy also describes collection involving content sent through online features, including prompts and related material. A no-training statement therefore does not mean that an upload never leaves your device or is never processed.

Read the relevant product and account settings before providing unreleased recordings, private correspondence or a collaborator's identifiable material.

For ordinary experiments, use your own nonconfidential sample assets and remove information the task does not need. If another person supplied the material, agree on the intended upload and use first.

Keep a record of that agreement with the project rather than treating an application's ability to accept a file as permission to use it.

Provenance helps explain the work

C2PA describes Content Credentials as signed, tamper-evident information about an asset's origin and changes. The system can record tools and processes used; it is not a mechanism that restricts copying.

Its FAQ also explains that complete ingredient verification depends on access to those ingredients' data.

For a local release, preserve credentials where supported, but also keep a plain-language credit note and the relevant source files. A technical history does not replace checking the claims in a poster, identifying a fictional scene or confirming that a pictured person approved the intended presentation.

Before publishing, inspect the actual exported file and its caption together. The audience should understand the creative contribution without needing to decode the production software.