Skip to main content
Rodrigo36141290yrjz
Participant
August 27, 2026

Firefly Image generation performance is really poor, maybe a harness issue

  • August 27, 2026
  • 3 replies
  • 34 views

Image generation performance is really poor; even with perfect prompts generated by another LLM and vetted by an engineer with 20 years of experience, the LLM fails to maintain the thread of the conversation. It is likely an integration issue—specifically, the way you filter content before it reaches Gemini (for instance). I tried it directly in Gemini and it works perfectly, so the problem lies with your software; this is one of the reasons people are migrating to other tools.

3 replies

Rodrigo36141290yrjzAuthor
Participant
August 27, 2026

Hi Matt, the web versions of Firefly and Photoshop are acting a bit dumber than usual...
There are several examples; one involves generating a landscape but making it 360-degree. You use a short prompt that even Gemini understands perfectly—"make it a seamless 360-degree panoramic landscape"—yet the application copies the same landscape onto itself—a glaring and absurd error. It’s not consistently reproducible; it’s one of those clear cases where the harness consumes part of the semantics and chews it up before passing it on to the final LLM.

 

Community Manager
Matt DelargeCommunity Manager
August 31, 2026

Hi Rodrigo36141290yrjz,

Thanks for the example and the image, really helpful.

I've tested your exact prompt ("make it a seamless 360-degree panoramic landscape") on two different models on my end, and both times the result was a clean, continuous panorama, no mirroring or duplication like what you're seeing. So the prompt itself works, at least in my setup, which means we need to narrow down what's different about yours (you can see the images generated at the bottom).

A few things that would help pin it down:

  • Which model were you using when this happened?
  • Was this a fresh text-to-image generation, or did it involve a reference image, editing, extending an existing image, or "generate similar"?
  • Are you using the web app, Photoshop, or another integration?
  • What aspect ratio or size did you have set?
  • Roughly how often does it happen, every time now, or occasional (1 in a few, 1 in 10)?

Getting these specifics will help isolate the actual variable and put together a solid case to escalate.

Thanks

Matt

 

Community Manager
Matt DelargeCommunity Manager
August 27, 2026

Hi Rodrigo36141290yrjz,

Thank you for taking the time to share this feedback.

While I can't confirm the specific backend integration theories regarding how the prompt is routed or parsed, I completely understand your frustration.

To help us to investigate this and see where the breakdown is happening, we need to reproduce the environment. Would you be willing to share:

  • An exact example of a prompt that failed in Firefly but worked well elsewhere?
  • Are you using the Firefly web application, or generating within a specific app like Photoshop?

Getting these specifics will allow me to escalate a concrete use case to the product team so they can review the prompt parsing and generation performance.

Thanks

Matt

Terms & ConditionsAccessibility statement
Using the community Terms of use Privacy Cookie preferences Do not sell or share my personal information ad choicesAdChoices