Firefly vs Midjourney: Better Output, Bigger Legal Risk

Matt Supple, a high-end retoucher with over 25 years of experience across New York, Sydney, and New Zealand and founder of Fred NY, spent a week testing AI image tools head-to-head. His finding: Midjourney produces more artistically capable work than Adobe Firefly, and the gap traces directly to their training data. Firefly uses a narrow, licensed dataset. Midjourney trains on whatever its crawlers can reach. That difference in approach produces a blunt trade-off for any agency or brand: better output carries more legal exposure.

What happened

Factor	Detail
Author	Matt Supple, retoucher and founder of Fred NY, 25+ years experience
Test method	Same references, same briefs, same intent fed to both tools
Firefly training data	Adobe Stock licensed library, openly licensed content, public domain
Midjourney training data	No published dataset; crawled social platforms, portfolios, stock sites
Firefly indemnification	Adobe defends enterprise customers against copyright claims on outputs
Midjourney indemnification	None; terms disclaim all responsibility for IP status of outputs
Firefly asterisk	In 2024 it emerged some training images were AI-generated, some sourced from Midjourney, uploaded to Adobe Stock by contributors

Supple ran the same creative brief through both tools and found Midjourney’s results “more striking, more original, more alive.” He argues this is a direct, measurable consequence of the training approach: a licensed dataset is a smaller, safer sandbox, and a model trained on it will reflect those constraints in its output.

Adobe built Firefly to be legally traceable. Every image in its training pool was compensated, and every contributor can, in theory, be traced to a legitimate license. That choice costs something on the craft side. Midjourney made the opposite call: train on scale, optimise for output, and let the terms of service put legal risk on the user.

Why it matters for working creatives and agencies

The gap between these two tools is not just a quality debate. Supple frames it as a livelihood issue: if a retoucher or illustrator is competing for briefs against someone using a scraping-trained model, they are not on a level field. The model may have trained partly on unconsented work from other professionals in the same industry.

The uncomfortable arithmetic: the “ethical” tool currently produces weaker work, and the market rewards results, not ethics.

What agencies are actually doing

According to Supple’s observations across agencies, the practical split looks like this:

• Mood boards and creative exploration: Midjourney, because enforcement has not caught up with usage.
• Final deliverables on large or enterprise campaigns: Firefly or similarly indemnified tools, because legal and compliance teams do not want unresolved IP exposure attached to high-visibility work.

Supple describes this as “a bet on obscurity, not a legal green light.” The risk math is driving the decision, not any resolution of the underlying ethics.

Is Firefly actually clean?

Supple is careful not to hand Firefly a clean bill of health. In 2024 it came out that some Firefly training images were AI-generated, with a portion originating from Midjourney outputs that contributors had uploaded to Adobe Stock. Adobe has since tightened its submission standards, but the “purely licensed” story carries an asterisk. No one in this space, according to Supple, gets to claim total moral high ground yet.

The legal clock is running

Multiple lawsuits against scraping-trained models are already moving through courts. None has produced a landmark ruling yet. But Supple notes that legal consequences, when they do land, often land retroactively. A serious ruling against a major scraping-trained model could expose campaigns already built on that output, not just future work.

This connects to a broader pattern we have been tracking in AI tool adoption across agencies, where usage is running ahead of legal clarity on almost every front.

Our take

Supple’s framing is honest in a way that most AI tool commentary is not. The quality gap he describes is real. Midjourney’s advantage comes from training at a scale that licensed datasets cannot currently match, and pretending otherwise does not help anyone plan their production workflow.

The practical question for any agency or brand is whether the campaign is large enough to attract a lawsuit. For a small business running a social ad, the exposure is low. For a national brand campaign, a Midjourney-generated hero image is a genuine liability sitting in your asset library waiting for a court to catch up.

If you are advising clients on AI integration for creative production, this is the conversation to have now, before the legal system forces it. Adobe’s indemnification offer is worth something. It is also, as Supple points out, priced into a weaker creative output for the moment. That trade-off may narrow as licensed datasets grow, but it has not narrowed yet.

For smaller or exploratory projects, the risk calculus is different. But “nobody has sued us yet” is not a compliance strategy. Document which tool generated what, keep your prompts, and make sure clients understand what they are approving.

What to do about it

1. Audit your current AI-generated assets and note which tool produced each one.
2. Flag any Midjourney-generated work that has been used or is planned for high-visibility, high-budget campaigns.
3. Check whether your Adobe plan includes the enterprise IP indemnification before treating Firefly as a legal safe harbour.
4. Brief clients explicitly on the tool choice and its IP implications before final sign-off, not after.
5. Watch the ongoing scraping lawsuits; a landmark ruling could change the risk calculation overnight.

The honest position right now: pick your tool based on the actual risk profile of the project, document everything, and stop pretending the gap between these two tools is purely about preference.

Source: Bing News · Midjourney