What's your prompt formula for photorealistic food shots that don't look fake?

Another day, another thread chasing the shiny AI-generated object. I see the usual galleries of impossibly glossy, surgically perfect burgers and cakes that look like they were extruded from a plastic mold. Everyone's so focused on the aesthetic output they forget they're just feeding a stochastic parrot that has never tasted, smelled, or understood the physics of food.

So you want "photorealistic"? That's a cost-benefit analysis of prompt engineering versus the fundamental limitations of the model. It's not about a magic formula. It's about specifying constraints to counteract the model's inherent tendencies to over-saturate, over-sharpen, and create surfaces with zero porosity or correct subsurface scattering.

First, you must aggressively counter the Midjourney defaults. You need to think like a lighting director and a food stylist, because the model certainly isn't one. Generic terms like "photorealistic" or "delicious" are useless. You must define the *failure modes* you want to avoid and explicitly negate them. Then, you must define the material properties and environmental context with obsessive specificity.

Here's a starting point for a structured prompt. Think of it as a configuration block. You're not writing poetry; you're parameterizing a render.

```
[subject description], prepared on [surface type], [meal context]
**Material Properties:** [texture 1], [texture 2], slight irregularity, [specific moisture/oil level]
**Lighting:** [lighting type] from [direction], soft shadows, [light temperature], subtle fill light
**Camera:** [lens type], [focal length], [aperture equivalent], shallow depth of field, viewpoint: [angle]
**Style & Refs:** photorealistic, food photography, style of [photographer name], 8k
**Negative Prompt:** shiny, glossy, plastic, slimy, CGI, 3d render, cartoon, painting, oversaturated, hyperdetailed, perfect symmetry, studio lighting
```

Example applied:
`A grilled portobello mushroom burger with melted swiss cheese and garlic aioli, on a toasted brioche bun with a leaf of butter lettuce, prepared on a weathered wooden cutting board, casual lunch setting. Material Properties: charred grill marks, porous bun texture, melting cheese viscosity, slight irregularity, light oil sheen. Lighting: late afternoon window light from frame left, soft shadows, warm natural light, subtle fill light from right. Camera: 85mm prime lens, f/2.8, shallow depth of field, viewpoint: overhead angle. Style & Refs: photorealistic, food photography, style of Jonathan Gregson. Negative Prompt: shiny, glossy, plastic, slimy, CGI, 3d render, cartoon, painting, oversaturated, hyperdetailed, perfect symmetry, studio lighting.`

Notice the effort? You're fighting the model's bias at every step. The "negative prompt" is your most critical line of defense—it's your error budget. "Slight irregularity" counters the unnatural uniformity. Specifying a real photographer gives it a corpus of actual photography to lean on, not just other AI images.

Ultimately, you'll burn through a lot of GPU credits and time tuning this. Is the ROI there compared to just... using a stock photo or, dare I say, an actual camera? Probably not for most use cases. But if you're determined to proceed, this is how you implement guardrails.

-- cynical ops