Firefly generated images are awful

I’m glad to hear that it’s not just ME !!!!!!!!

I’m pretty frustrated with it myself, especially today as I’m trying to find tune a project. Something I’ve noticed that helps control the hallucinations is, first, generate the base image elsewhere (ChatGPT excels at this). Then use it as both the first and last frame of the clip you hope to generate. In order to stay true from the beginning to end, Firefly and its other models are less likely to add elements you didn’t ask for. You can then easily loop the clip to make it longer. Reverse it to add variety. You can even cut it, then copy it, add it to the back of the cut clip, reverse it, and then it creates a slightly different but seamless action without any jump cuts. But, man, I’ve wasted lots of tokens this morning just trying to get a figure to hit drums with drumsticks instead of beating dead air. We’re not there yet.

A lot has changed since this thread started: Firefly now runs on Image Model 5 (October 28, 2025), with meaningful gains in realism, anatomy, and prompt adherence, plus new partnered models to choose from. If you last tried Firefly a while ago, it's worth generating a fresh result to compare. 

We know the earlier experience fell short for many of you here, and that this reflects real time and money invested — we don't take that lightly. Two things that help today: the Firefly AI Assistant for building stronger prompts, and the thumbs-down button on any miss, which sends specifics straight to the team. 

I tried in vain to have it generate a simple cosine function.  No matter how I prompted asking for the image to be simple, I got a crazy fantasy image. A tool that cannot plot a simple cosine wave is useless. It's like the program has a set of horrible stock images that it just has to use.  Commenting because I am ready scream. Any tool that causes so much stress is not worth it.

I tried to generate an image "people in grocery store fighting over toilet paper graphics novel style"

and it got flagged as not appropriate, your limitations on this make me not want to pay for it anymore.

I agree! The videos especially are terrible quality especially when faces or living beings are involved. Even something as simple as a horse running in a field looks terrible and unlifelike. 

Agree strongly, appalling results and not even close to the detailed and specific prompt.
Each generation is worse than the previous!

I agree. I've been working with it, and it has degenerated. Now it takes more and more effort to get it to even mimic a past example. It is so painful to work with. This is from someone who has worked in Photoshop and Indesign for 30 years.

Humans is ok, as long as you don´t try to create to much at the same time. Vehicles on the other hand looks terrible... Try create ford mustang 1969 for example =)))))))

I completely agree. The results are useless. I don't understand how Adobe can release this product. It makes them look ridiculous. Compared to Midjourney, it's like a square wheel. A lot of work needs to be done on it. I only use it to expand images, otherwise for nothing at all.