Skip to content
Notifications
Clear all

My results after trying to generate food photography - mostly inedible.

7 Posts
7 Users
0 Reactions
52 Views
(@chrisf)
Reputable Member
Joined: 4 months ago
Posts: 284
Topic starter   [#28019]

I've been trying to use Firefly for mockups of menu items for a small cafe project. I keep asking for things like "a perfect slice of carrot cake on a rustic plate" or "steaming latte art in a ceramic mug."

The results are... weird. The cake often looks melted or has strange textures. The latte foam sometimes looks like concrete. 😅 Has anyone else tried this for food? Are there specific prompts that work, or is this just not the right tool for it?

I'm comparing a few different image generators for this, but Firefly's integration with other Adobe apps is appealing. Just not if the food looks inedible.

Thanks in advance!


Still learning.


   
Quote
(@bench_beast)
Noble Member
Joined: 5 months ago
Posts: 723
 

The weird textures and melted look are common across image gens for food. Firefly seems to struggle with material coherence - it knows the words but not the physics.

For something like latte art, you'll get better results specifying the lighting and angle. Try "overhead shot of a latte in a white ceramic mug, detailed foam art, soft shadows, steam wisps, on a wooden table."

But honestly, for menu mockups I'd still comp the shot in Photoshop with a real base image. Firefly's strength is the integration for editing, not the raw gen quality for this niche.


Benchmarks don't lie.


   
ReplyQuote
(@elenar)
Reputable Member
Joined: 4 months ago
Posts: 293
 

You're right about material coherence being the core limitation. The training data likely contains millions of food images, but the model doesn't internalize properties like viscosity, surface tension, or how steam interacts with light. It's generating plausible 2D pixel arrangements, not simulating a physical object.

This is why specifying lighting and angle helps marginally. It constrains the possible viewpoint, reducing the solution space where those physical inaccuracies manifest most glaringly. An overhead shot simplifies the geometry the model has to fake.

However, I'd slightly push back on the recommendation to comp in Photoshop for all cases. For a small cafe exploring concepts, the speed of generating fifty variations of a "rustic plate" background texture or a "wood grain table" for layout purposes can be valuable. The inedible cake is a failure, but the generated environment around it might be a usable asset. The real workflow is probably iterative generation for elements, then compositing, rather than expecting a single perfect end-to-end generation.


Data doesn't lie, but folks sometimes do.


   
ReplyQuote
(@cloud_cost_hawk_2)
Honorable Member
Joined: 7 months ago
Posts: 472
 

Oh man, the "latte foam sometimes looks like concrete" bit got me. That's the perfect description for AI-generated food texture failure. I tried something similar for a side project, generating "perfectly stacked pancakes" for a diner's website mockup.

Firefly gave me a stack that looked like beige rubber washers. Utterly unappetizing.

You're right to question if it's the right tool. For actual menu items you want people to crave, you're fighting physics the model doesn't understand. The integration is tempting, but you'll spend more time fixing the inedible gloop in Photoshop than you saved.

My hack? I stopped prompting for the food itself. I'd generate a killer rustic plate, perfect woodgrain table, beautiful soft light... then drop in a real photo of the cake I took with my phone. The comp looks cohesive and the food looks real. The AI handles the *environment*, you handle the edible subject. Saves a ton of editing pain.



   
ReplyQuote
(@emilyk99)
Estimable Member
Joined: 4 months ago
Posts: 173
 

I've been running into the exact same issue trying to create images for email campaign headers promoting a bakery's new seasonal items. That "concrete" texture for latte foam is spot on.

I'm also drawn to the Adobe integration, but I wonder if the core problem is in the prompt itself. Have you tried being hyper-specific about the *state* of the food, like "a slice of moist but structured carrot cake with visible carrot shreds and a crumbly top" instead of "perfect"? I've had slightly better, though still not perfect, results when I treat it more like a brief for a food stylist than a simple description.

What are the other image generators you're comparing it to? I'm curious if any of them handle material properties better for this use case.



   
ReplyQuote
(@emmab3)
Reputable Member
Joined: 4 months ago
Posts: 271
 

That concrete foam description is painfully accurate. It's not a prompt problem, it's a fundamental training data and physics simulation gap. These models are trained on labeled images, not on material properties like surface tension or how steam scatters light.

You mention comparing other generators. Have you run a controlled test? Take the exact same prompt you used for Firefly and run it through Midjourney, DALL-E 3, and Stable Diffusion. Compare the outputs side-by-side on the same criteria: texture plausibility, structural integrity of the cake, foam realism. You'll likely find they all fail, just in different aesthetically weird ways. One might give you a structurally sound but plasticky cake, another gives you a realistic texture but the slice is floating.

The integration is the lure, but if the base asset is fundamentally broken, you're just doing cleanup on a bad foundation. For mockups where appetite appeal is secondary to layout, maybe it's fine. For making someone want to buy the carrot cake, it's the wrong tool.


FinOps first, hype last


   
ReplyQuote
(@devops_dad_joke)
Reputable Member
Joined: 8 months ago
Posts: 288
 

Oh yeah, the "concrete foam" is a classic. It's like the model understands "foam" as a visual keyword but not as a physical substance you'd want to ingest.

Honestly, for menu mockups, you're better off using it for everything *around* the food. Generate that perfect rustic plate and soft lighting, then composite in a real photo of the actual cake. You'll save hours of trying to debug why the carrot cake looks like it's sweating.

The Adobe integration is fantastic, but it's a better editing partner than a primary generator for this specific job. It's the difference between a good power tool and the right brush.



   
ReplyQuote