Hey folks, hope everyone's doing well. I've been deep-diving into Adobe Firefly for a creative project, and I've hit a weird snag that feels a bit like a security misconfiguration, but in the creative space. 😄
Whenever I generate a portrait—regardless of the prompt variations like "elderly asian woman with kind eyes," "young freckled redhead laughing," or "middle-aged man with a beard"—the output images all have the *same* underlying facial structure. It's like the model is locked onto a single "base" face template and just tweaks the superficial features. It's frustrating because I'm aiming for diversity in characters.
Here's what I've tried so far, thinking in terms of "parameter tuning":
* **Prompt Engineering:** Adding specific descriptors (e.g., "oval face," "wide-set eyes," "distinctive nose"). It listens, but the core face ghost is still there.
* **Style Variations:** Cycling through all the different artistic styles. The style changes, but the facial geometry remains eerily consistent.
* **Using Reference Images:** This sometimes works better, but it defeats the purpose of pure generation for this project.
Has anyone else encountered this "face lock" issue? It feels analogous to an IAM policy that's too restrictive—maybe there's an implicit bias or a limitation in the training data subset Firefly is using for portraits? I'm curious if there's a known workaround, or if we just have to wait for model updates.
From a DevSecOps perspective, I wonder if there's a way to "break" the default inheritance, similar to how you'd use a custom policy to override a too-restrictive SCP. Not sure what the equivalent would be here—maybe a specific negative prompt?
security by default