Oh wow, I've been hearing about Nano Banana Pro in random stuff lately, but as a layman the difference is stark. It's the only one that actually looks like a partially eaten burrito at all to me. The others all look like staged marketing fake food, if I'm being generous (only a few actually approach that, most just look wrong).
Hunyuan V3 is the only other one that plausibly has a bite taken. The weirdness of the fillings being decoratively sprinkled on top of it does rather count against it, though.
The NBP looks like a mock of food to me - the unwrapped burrito on a single piece of intact tinfoil, a table where the grain goes all wonky, an almost pastry looking tortilla, hyperrealistic beans and there's something wrong with the focal plane.
It's just not as plasticy and oversaturated as the others.
Yeah, hyperrealistic beans. They don't look real at all. The inside of an actual burrito is messy after you bite into it (and usually before). That burrito has a couple of nearly dry, yet for some reason speckled, beans that look more like they're floating on top of the burrito rather than actually in it.
And yeah, the focal plane is wonky. If you try to draw a box around what's in focus, you end up with something that does not make sense given where the "camera" is - like the focal plane runs at a diagonal - so you have the salsa all in perfect focus, but for some reason one of the beans which appears to be the exact same distance away, it subtly out of focus.
I mean, it's not bad, but it doesn't actually look like a real burrito either.
This shows some gaps in the "same prompt to every model" approach to benchmarking models.
I get that it's allows ensuring you're testing the model capabilities vs prompts, but most models are being post-trained with very different formats of prompting.
I use Seedream in production so I was a little suspicious of the gap: I passed Bytedance's official prompting guide, OPs prompt, and your feedback to Claude Opus 4.5 and got this prompt to create a new image:
> A partially eaten chicken burrito with a bite taken out, revealing the fillings inside: shredded cheese, sour cream, guacamole, shredded lettuce, salsa, and pinto beans all visible in the cross-section of the burrito. Flour tortilla with grill marks. Taken with a cheap Android phone camera under harsh cafeteria lighting. Compostable paper plate, plastic fork, messy table. Casual unedited snapshot, slightly overexposed, flat colors.
Then I generated with n=4 and the 'standard' prompt expansion setting for Seedream 4.0 Text To Image:
They're still not perfect (it's not adhering to the fillings being inside for example) but it's massively better than OP's result
Shows that a) random chance plays a big part, so you want more than 1 sample and b) you don't have to "cheat" by spending massive amounts of time hand-iterating on a single prompt either to get a better result
100%. Between tuning prompt variations depending on the model and allowing a minimum number of re-rolls, this is why it takes a while to publish results from the newest models on my GenAI comparison site.
Including a "total rolls" is a very valuable metric since it helps indicate how steerable the model is.
not adhering to the prompt guide is def a valid strong criticism. resampling i think less so for the demo just because fewer people look at k samples per model, so just taking literally the first one has the fewest of my own biases injected into it
I actually think it's ok to inject your own bias here: if you're deploying these models in production, then you probably test on your own domain other than half eaten burritos lol
But individual users usually iterate/pick, so just sharing a blurb about your preference is probably enough if you choose 1 of n
An interesting American culinary divide is between Scottsdale and Phoenix homemade burritos. The former being close to the Midwest variety, the latter to a Sonoran style.
Even ignoring the Heinz bean outliers, these are all decidedly Scottsdale. With one exception. All hail Nano Banana.
They all just look like generic Mission burritos to me (leaning towards fast food menu photos), except some include lettuce and some have blisters sonoran style. Only Nano Banana really looks like something I'd get at El Farolito or an LA food truck.
One of my tests for new image generation models is professional food photography, particularly in cases where the food has constraints, such as "a peanut butter and jelly sandwich in the shape of a Rubik’s cube" (blog post from 2022 for DALL-E 2: https://minimaxir.com/2022/07/food-photography-ai/ )
For some reason ever since DALL-E 2, all food models seem to generate obviously fake food and/or misinterpret the fun constraints...until Nano Banana. Now I can generate fractal Sierpiński triangle peanut butter and jelly sandwiches.
I can kind of see what you mean in that it went for realism in the aesthetics, but not the object... but that last one would probably fool me if I was scrolling
hrm. yea you're right. the page on fal used to produce it was linked with the image, but maybe i made a mistake and sloppily saved wrong one. ill have to reroll to check
With llms there is a secondary training step to turn a foundational model into a chat bot. Is these something similar going on with these image generation models, that is making them all tend towards making pretty clean images and stopping them making half eaten food even if they have the capabilities?
In terms of prompt adherence, there are two issues with most image generation models, neither of which apply to Nano Banana:
1. The text encoders are primitive (e.g. CLIP) and have difficulty with nuance, such as "partially eaten", and model training can only partially overcome it. It's the same issue with the now-obsolete "half-filled" wine glass test.
2. Most models are diffusion-based, which means it denoises the entire image simultaneously. If it fails to account for the nuance in the first few passes, it can't go back and fix it.
I believe some image generation AIs were RLHFed like chat bot LLMs, but moreso to improve aesthetics rather than prompt adherence.