When the Better Model Gets Shut Down: What Firefly Image 3 Teaches About AI Tools in Production

This month, Firefly Image 3 disappears from Adobe's tools for good – after an earlier stage back in April, when the model was already removed from Photoshop's model picker. Generative Fill and Generative Expand now run exclusively on newer models, and Adobe recommends switching to "Firefly Fill & Expand." So far, an ordinary product cycle. What's interesting is a detail that rarely comes up: in Adobe's community, numerous users are asking for the old model to be kept as a paid legacy option – because it delivered better results on hair, textures, and unguided retouching than its official successor.

The narrative that doesn't hold up here

The usual story with AI models goes: the new version is better in every respect, and hesitating just means missing out on progress. With Firefly Image 3, practice contradicts exactly this narrative, and precisely in places no benchmark captures well – how a model resolves fine strands of hair at an edge, how it interprets material surfaces without explicit instruction, how it responds when a retouch is deliberately left vague because you don't yet know exactly what you want. Those are exactly the moments in which a tool proves trustworthy in the studio, or doesn't – and a new model can be more capable overall while performing worse in these specific situations, because training shifts style along with capability.

The actual difference from physical tools

This is exactly where a dependency emerges that a camera or a lens doesn't have – the same distinction between physical and licensed tools I already described regarding product photography. A flash head that worked well in 2014 still works exactly the same in 2026 – nobody can switch it off remotely. A generative model, by contrast, is licensed infrastructure: it runs as long as the provider operates it, and it disappears on the provider's schedule, not the user's. Anyone who's built a retouching routine, a preset, or a specific image look around a model's particular quirks loses that foundation through no fault of their own – not because the tool breaks, but because the manufacturer discontinues it.

Why "better" is never one-dimensional with AI image models

The community's request for a paid legacy option hits a sore spot model providers rarely acknowledge: a newer model usually optimizes for a broader average of use cases, not for every specific niche a predecessor happened to excel at. For a studio that's spent years developing a particular visual language, that niche is often exactly where the value lies – and "better on average" is no guarantee it holds true for your own work. To Adobe's credit, documents already created remain unchanged; the retirement only affects future use. That's at least more honest than silent retroactive changes, but it doesn't solve the actual problem: whoever wants to keep working tomorrow works with a different tool, whether they choose to or not.

What this means for production decisions

For us as an agency building AI-assisted image production into real workflows – most recently put through its paces with a comparable tool stack in a solo campaign – the lesson isn't against generative tools, but for deliberate distance from any single model version. A preset, a workflow, a promise to a client should never be tied to one specific model that can disappear at any time – it should be tied to a result that can, if necessary, be reproduced with a different tool. In practice, that means: regularly testing whether switching models changes your own look before a provider forces the switch, and defining your own visual signature so it stays recognizable independent of whichever tool happens to be in use.

Conclusion

Firefly Image 3 disappears this month, and part of the community is saying openly: for certain tasks, it was the better tool. That's not a nostalgia complaint – it's a sober reminder that generative models are infrastructure, not tools in the classic sense; they run on someone else's schedule. Anyone working with them professionally is better off building their signature around results than around a single model that can be switched off at any time.