Midjourney vs Adobe Firefly, Six Months Later: Same Fight, Thicker Lawsuit Stack

Six Months in AI Years

Six months ago, Firefly still had that “responsible corporate citizen” glow, and Midjourney was the brilliant troublemaker I secretly loved working with. I wrote a piece framing them as two very different answers to the same question: how do we let a machine hallucinate images without ending up in court?

Fast-forward half a year, and the answer is: we still don’t know, but the lawyers are definitely getting paid.

In the meantime, Midjourney has shipped Version 7, added video, and doubled down on “insane quality, we’ll figure out the law later.” Firefly has marched through Image Models 3, 4, and now 5, improved technically, but is still the kid who shows up late to the party with a PowerPoint while Midjourney is already dancing on the table. 

And the legal landscape? The artist class action survived, the big studios have entered the chat, and Adobe’s “ethical AI” branding now comes with an asterisk the size of a billboard.

Time for a rematch: not Firefly vs Midjourney as art toys, but as legal liabilities and production tools in late 2025 – with my own experience as a power user watching Firefly still glue weird backgrounds onto otherwise decent people.

Rewind - The Training Data Line in the Sand (Spoiler: It Got Smudged)

The original story was simple enough.

Firefly was marketed as the responsible one: trained on Adobe Stock, public-domain material, and other “licensed content,” designed from day one to be “commercially safe” and wrapped in a corporate indemnity bow for enterprise customers.  

Midjourney was the opposite: a spectacular model trained on whatever its crawlers could inhale from the open web. No detailed public dataset list, no comfort blanket of “we licensed everything,” just a Terms of Service that basically says, “You can use what you create, but if anyone sues, that’s your problem.”

That contrast hasn’t gone away. But the neat moral framing absolutely has.

In April 2024, Bloomberg revealed that roughly five percent of Firefly’s training images were AI-generated – including works from rival systems like Midjourney – because Adobe Stock had accepted synthetic images from contributors and thrown them into the Firefly training pool. 

Ethically, that moved Firefly from “organic, farm-to-table training data” to “mostly organic, with a side of synthetic sausage from the neighbor’s kitchen.” Legally, it didn’t automatically break Adobe’s promise – those images were still licensed through Stock – but the narrative of “we’re nothing like those scrapers” took a serious hit.

Six months later, that detail matters more, not less, because of what’s happening to Midjourney in court. If the source model gets declared unlawful, its fingerprints on Adobe’s supposedly clean training set suddenly look a lot less theoretical.

Midjourney’s Legal Situation - From Artist Complaints to Studio-Level Smackdown

When I wrote the original piece, the big Midjourney story was the Andersen v. Stability AI class action: artists arguing that scraping millions of their images to train models like Stable Diffusion and Midjourney was copyright infringement. In August 2024, a federal judge allowed the core copyright claims to proceed, rejecting attempts by Stability and Midjourney to throw the case out at the motion-to-dismiss stage. 

In plain language, the court did not accept “training is obviously fair use, go away” as an answer. The case moved into discovery, which means Midjourney’s training practices are now under a microscope instead of safely behind PR talking points.

Since then, things escalated dramatically. In mid-2025, Disney and Universal filed separate suits accusing Midjourney of training on copyrighted characters like Yoda, Bart Simpson, Iron Man, Shrek, and friends without authorization, and of enabling users to generate derivative, infringing imagery on demand.  

Warner Bros. Discovery followed with a similar complaint, showing side-by-side comparisons of Midjourney outputs and its own characters, and arguing that the model’s ability to mimic them is evidence of unlawful training. 

On top of that, you now have a separate lawsuit over the outputs of Midjourney: Jason Allen’s suit against the U.S. Copyright Office, after it refused to register his award-winning Midjourney-assisted artwork unless he carved out the AI-generated parts. He’s arguing that human direction and iteration should be enough to claim authorship even when the pixels come from an AI. 

Put all of this together, and Midjourney is now exposed on both sides:

Training risk: “You ingested our copyrighted works without permission” is no longer just an artist complaint – it’s now a multi-studio offense.

Output risk: “Purely or predominantly AI-generated images don’t get copyright” is now being actively challenged in court, but so far, regulators have not moved to bless AI-only works as protectable.

Does this mean you can’t safely use Midjourney in a campaign? Not yet.

No judge has ordered the model to be turned off, and no brand has been sued just for using Midjourney images. But the direction of travel is clear: the more high-profile the model gets, the more attractive it becomes as a legal target.

And there is precisely zero Midjourney indemnity standing between those lawsuits and you.

Firefly’s Halo Cracks - Midjourney in the Training Set and the Adobe Stock Mess

Firefly doesn’t have a big headline lawsuit with its name on it, but it does have a trust problem.

First came that Bloomberg piece: Firefly trained in part on AI-generated images from other tools, including Midjourney, via Adobe Stock. Adobe’s response was, essentially, “Yes, but it’s less than five percent, and every image goes through moderation; we still own or license the lot.” 

Then, contributors noticed a more uncomfortable detail in Adobe’s own forums: an official response acknowledging that contributors can take any Adobe Stock image, feed it into Midjourney as a prompt, generate a derivative AI image, and then sell that derivative back on Adobe Stock as long as they claim to hold the rights. 

So, we now have a neat little loop:

1. Human contributor uploads a copyrighted photo to Adobe Stock.

2. Someone uses that image as seed material in Midjourney and creates an AI derivative.

3. The derivative gets sold as “Generative AI – Adobe Stock.”

4. Firefly is later trained on Adobe Stock, including these AI derivatives.

From Adobe’s contract perspective, all of this is “licensed.” From an artist’s perspective, it looks suspiciously like laundering scraped or derivative AI content back into a supposedly clean pool.

Meanwhile, Adobe Stock has quietly become an AI-heavy library: by mid-2025, nearly half of all images on the platform were reported to be AI-generated. 

When Adobe reassures enterprise customers that Firefly is trained on “content we own or licensed,” that statement is technically true but increasingly synthetic. You’re no longer just licensing from human photographers; you’re licensing from a stack of prior models that may themselves sit on contested data.

No one has filed a Firefly-specific training lawsuit yet. Instead, the pressure has come from backlash and policy scrutiny.

Enterprise marketers and lawyers are openly worrying that Firefly’s indirect use of Midjourney imagery could expose them to the same kind of artist claims they were trying to avoid in the first place. 

Adobe has reacted by doubling down on messaging: publishing legal FAQs, reiterating that enterprise plans can include contractual IP indemnification for Firefly imagery, and codifying a pledge not to train on customer project data after a separate terms-of-service scandal blew up in mid-2024. 

Firefly is still framed as “the safest option,” but the optics have changed. The moral high ground is more of a muddy hill. The legal shield is still there, but everyone is now looking very closely at the bolts.

Outputs and Ownership - 2025’s Version of “Do I Actually Own This Image?”

On the ownership side, not much has improved – but the stakes are clearer.

Midjourney still tells paying subscribers they “own” the assets they create and can use them commercially, while free users are limited to non-commercial Creative Commons–style use. That is a contractual promise, not magic copyright dust. If the law says a purely machine-generated work isn’t copyrightable, Midjourney can’t conjure a copyright out of thin air; it’s really promising not to assert its rights against you, not guaranteeing that courts will treat you as the author. 

Allen v. Perlmutter is the test case here. If he wins, courts might start recognizing heavily AI-assisted images as copyrightable when the human contribution is substantial enough. If he loses, the message is: use Midjourney for creativity, but don’t count on IP protection as if you’d hired a photographer.

On the Firefly side, Adobe keeps repeating the same pitch: Firefly outputs are safe for commercial use, and enterprise customers can purchase indemnity so Adobe will step in if anyone sues over infringement tied directly to Firefly’s training set. 

Two caveats matter in 2025.

Indemnity is a contract, not a magic forcefield. It usually applies only to certain Firefly features, only if you are on specific enterprise plans, and only as long as you don’t do something obviously risky with the output. And its real value will only be proven when someone actually sues a Firefly customer, and Adobe has to show up with the checkbook.

Suppose courts ultimately decide that scraping-based training is infringing. In that case, plaintiffs will very likely argue that derivatives of that training, even if “licensed” through Adobe Stock, are tainted by the original misuse. We don’t know yet whether that argument will work, but in a world where Disney is suing Midjourney over character mimicry, you can bet it will be tried. 

So yes, Firefly still feels safer than Midjourney, especially for big brands. But the gap is no longer “safe vs. reckless.” It’s more like “vendor backed with a legal fund vs. model flying without a parachute.”

Tech Rematch - Midjourney V7 vs Firefly 2025, Seen from the Prompt Trenches

Legally, things got messier. Technically, the gap you described six months ago is still very real, just at a higher resolution.

Midjourney V7 became the default in June 2025. It brought exactly the things I feel when I  prompt it every day: far better prompt comprehension, cleaner anatomy, more stable object layouts, and fewer weird micro-artifacts. Documentation and third-party comparisons emphasize richer textures, more coherent hands and faces, and a much more consistent interpretation of complex prompts. 

Crucially for power users, Midjourney still lets you switch versions with --v and keep your older styles alive. You can stay in V6 if a client loves the look of that era, or deliberately mix V6 and V7 to get slightly different reads on the same prompt. Style Reference, Personalization, and the growing “moodboard” ecosystem all work across versions, which makes “my V6 look” effectively a brand asset you can protect by choice, if not by law. 

On top of that, Midjourney’s new video feature turns individual images into short animated clips, giving you another creative dimension – and another vector for IP risk when those clips look suspiciously like studio characters. 

Firefly, to be fair, has not stood still. Image Model 3 and 4 were big jumps in lighting, positioning, and facial realism; Image Model 5 at MAX 2025 finally feels like a serious attempt to play in the same photorealistic league, with native 4 MP outputs and better prompt-based editing. 

But people’s criticism from the trenches still lands: Firefly struggles badly with backgrounds and integrated lighting when you push it beyond “pleasant catalog photo.” You ask for cinematic depth and coherent environmental storytelling; it gives you something that looks like a kindergarten collage, with subject and background emotionally divorced from each other.

That mismatch is not just an aesthetic annoyance – it’s a production cost. Midjourney V7 can give you a near-final visual concept in a few iterations. Firefly often gives you something you still have to fix in Photoshop: rebuild the background, massage the lighting, remove the uncanny seams. If your workflow is “get it 80% there with the model and polish in post,” Midjourney is 80% and Firefly still feels like 40–50% on complex scenes.

Firefly’s real strengths live where you’d expect from Adobe: It is tightly integrated with tools creatives already use – Photoshop’s Generative Fill, Illustrator’s vector generation, and bulk-automation APIs for removing backgrounds or resizing thousands of images in one pass. 

It is built for boring but essential production tasks at scale, where “consistent, not spectacular” beats “breathtaking but unstable.”

Midjourney, by contrast, still feels like a creative lab with a thin layer of production. For someone like me, who is comfortable living in that lab and then rebuilding the result in other tools, it’s fantastic. For a conservative brand team that wants a controllable pipeline, it’s still a bit of a fever dream.

Commercial Safety Nets in 2025 - How Much Comfort Is Left?

From a brand or agency perspective, the story today looks something like this.

Midjourney is more powerful, more flexible, and more visually rewarding. It also has more legal heat on it than ever: artists, film studios, and regulators all have it in their sights, and there is no indemnity layer between their lawsuits and your marketing department. The upside is creative impact; the downside is that if there’s ever a “first brand sued over Midjourney image” headline, you do not want to be the case study. 

Firefly is technically better than it was when I wrote the first article, but still behind in the areas I personally care about most: complex backgrounds, subtle lighting, and that “this looks like it belongs together” feeling. However, it remains the flagship example of a vendor willing to put legal skin in the game. Adobe is still extending IP indemnity for eligible Firefly outputs on enterprise plans and loudly positioning Firefly as the industry blueprint for “legally grounded AI.” 

The Firefly training controversy hasn’t destroyed that story, but it has weakened the “we are the ethical alternative to the scrapers” narrative. The reality is now more nuanced:

Firefly’s contracts and indemnity remain a greater comfort zone than anything the open-web models offer.

Firefly’s data provenance is cleaner than Midjourney’s, but not as pure as Adobe’s early marketing implied, and it now clearly contains indirections back to models built on scraping. 

If you are a risk-averse CMO, you still pick Firefly – but you do it with your eyes open, not with the old “Adobe is the good guy, everyone else is shady” comfort.

So Which Model Would I Trust With a Real Campaign Right Now?

If you strip away the marketing spin, we’re left with a weirdly pragmatic answer.

If your priority is maximum visual quality, flexibility, and style continuity over time, you stay where you already are: Midjourney, with V6 and V7 as your two main instruments. The ability to keep V6 alive while exploring V7 is a real strategic advantage. It means your “house style” in Midjourney is not at the mercy of a single model update; you can keep your version 6 look supporting version 7 improvements instead of watching your brand visuals morph overnight. 

But you also accept that, legally, Midjourney is now the model most likely to be used as Exhibit A in the next big copyright case. That doesn’t mean you stop using it; it means you decide where you let it touch production:

Use Midjourney as your high-end concept engine – pitches, internal boards, “what if” explorations, mood and style development – and then either recreate those concepts with human artists, photography, or Firefly-assisted workflows for final assets.

Use Firefly when you need boring reliability, audit trails, and a corporate vendor willing to include indemnity in your MSA. It may never give you the same shivers as a perfectly dialed Midjourney V7 prompt, but your legal team will sleep a lot better.

For me personally, this is the paradox I’m already living: as a power user, I can consistently get Midjourney to produce results Firefly simply cannot match, especially with complex backgrounds and sophisticated lighting. Firefly might catch up one day – the jump from Image 2 to 5 suggests Adobe is at least racing – but right now, I still see it fall apart exactly where Midjourney shines.

The only thing that truly changed in six months is how honest everyone has to be about that tradeoff. Firefly is still the closest thing to “corporate-friendly AI art,” but the Adobe Stock/Midjourney loop means it is not ethically distinct from the rest of the ecosystem. Midjourney is still the untamed creative engine, but it is no longer just an artist community controversy; it is now a front-line defendant against some of the biggest IP holders on the planet.

So, my updated verdict for late 2025 looks like this:

If your goal is to push creative boundaries and you’re willing to manage legal risk like a grown-up, Midjourney remains the better brush.

If your goal is to push campaigns into the world without having to explain AI training law to your CEO every three weeks, Firefly remains the safer canvas – even with a faint Midjourney watermark under the paint.

And if you actually care about the long-term health of the creative ecosystem, you should probably be slightly uncomfortable with both. That discomfort, more than any single lawsuit, is the real story of the last six months: every “ethical” shortcut has a shadow, and every breathtaking image is now a quiet reminder that the legal fight over who owns the training data is nowhere near over.