Firefly 4 has issues with generating unrealistic or exaggerated images, but not Imagen or GPT Image

[nmycps]

I suppose this is a general bug in the way Firefly 4 interprets prompts that specify exaggerated objects, and I find it almost impossible to use with this type of generative image. For example, generating a guitar with an unusually short neck; a baseball bat as long as a flagpole, or an extremely large sundae dessert compared to the person holding it. I've tried using some help from Chat GPT and Perplexity, for example, in getting suggestions (which are good) but still Firefly 4 fails to render them with any noticeable exaggeration. Almost as if it cannot understand how to generate unrealistic-style images. Rather it conforms to normalcy as much as possible. 
Now with Imagen 4 or GPT Image, I get a generated image that much more closely aligns with the details of my prompt. Both have consistent results. But I think overall they are proving to be more accurate than Firefly 4. I've used up hundreds of credits with Firefly 4, but after just one or a few trials with the others, I am getting much better results.

[J E L]

Hi @nmycps, interesting. I will experiment with this and let you know how it goes. It sounds like you have good prompts, but can you share some exact (full) prompts? I'll see if I can produce something better using these types of words: 

“Surrealist,” “cartoonish,” “abstract,” “fantasy,” “expressionist,” or “pop-art” as these should be signals for Firefox to depart from realism. For example, “rendered in a surrealist style” or “use a cartoonish aesthetic” might push Firefly toward non-photorealistic outputs.

When it comes to exaggeration, be as specific as possible when it comes to proportions, for example, “a house ten times larger than normal.”

You can also try adding surreal descriptors: “dreamlike,” “otherworldly,” “psychedelic,” or “fantastical.” For example, “a sea with floating islands and glowing waves.” 

“Generate a still image of an oversized, exaggerated banana split ice cream sundae, towering comically large compared to a person holding it, with vibrant, surreal colors like neon pink whipped cream and glowing green bananas, rendered in a whimsical, cartoonish style. The person appears small in scale, standing in awe, with the dessert dominating the scene.”

[nmycps]

Great suggestions. I’ve already tried many variations in Firefly with some
of those you mentioned. My aim is not to have surreal, fantasy-like,
psychedelic, or other-worldly - style images, but highly realistic
photographs but abnormal in terms of scale. My version of the sundae is:
"a girl is holding a giant sundae dessert in an exquisitely-designed tall
glass container. The sundae has vanilla ice cream, with generous flows of
chocolate syrup and caramel syrup running down the ice cream. Neat the top
of the sundae, there's some decorative whip cream, and there's a cherry on
the very top of the sundae. The girl is dwarfed by the sundae which is
dominating the scene and is almost as large as the girl's body. She is
sitting at a table in a retro diner with shiny candy-apple red seats. Shot
from above."

[J E L]

@nmycps, Okay, thanks for these extra details. Here is what I got using Firefly Image 4 Ultra. What do you think?

[nmycps]

Thanks for the follow-up. A good result with your example. The way in which Firefly interprets prompts in this manner sometimes requires redundancy and over-specificity, I think. But even more challenging that Firefly really struggles, is with generating pure abnormalcy, such as a cat with a very short tail, or an elephant with an abnormally short trunk, or a shark that is taller than it is long. The best results that followed my prompts were using the GPT Image model. But it would be great if Firefly could in the future produce the image more accurately from the prompt. 

[Kartika Rawat]

Thanks so much for your thoughtful feedback! While Firefly does well with many creative prompts, it can sometimes struggle with generating more abstract or intentionally “abnormal” scenarios, especially ones that go against typical visual patterns it’s learned.

We’re continually working to improve how Firefly interprets and renders unique and imaginative prompts, and your examples are incredibly helpful. We’ll share this with our product team as we aim to make prompt interpretation more flexible and nuanced in future updates.

Let us know if you have any questions.
^KR

moved to discussions

[nmycps]

Thanks Kartika. I did have a quick question: J E L said above the use of Firefly 4 Ultra made the image shown above, and it's a model I haven't used. I was running under the assumption that Ultra just boosts the resolution? Or is it a further refined model of Firefly that also interprets prompts better than Firefly 4?

[J E L · Accepted solution]

@nmycps, to step in for Kartika, this is the description of Firefly 4 Ultra from the Adobe blog:

“When your projects demand more detail and realism, Image Model 4 Ultra is your go-to. This model shines in rendering photorealistic scenes, human portraits, and small groups, ensuring they look natural and lifelike. It's designed for highly complex needs, when precision and clarity are paramount.”