AI Image Generation Through a Photographer's Eyes: What I've Learned After a Year of Experimenting

I resisted AI image generation for longer than I'd like to admit.

After building a career on the precision of real light hitting a real subject — twenty years of advertising campaigns for Coca-Cola, Pepsi, General Motors, and dozens of other brands — the idea of generating images from text prompts felt like a shortcut that missed the point entirely. Photography is about presence. About the moment between the photographer and the subject. How could a machine replicate that?

Then I actually tried it. And my perspective shifted — not in the direction you might expect.

Why I Started: Curiosity, Not Replacement

Let me be clear about something upfront: AI has not replaced my camera. It hasn't replaced anyone's camera. What it has done is add a new layer to my creative process — one that I find genuinely exciting, even as it raises questions I'm still working through.

I started experimenting with Stable Diffusion and SDXL about a year ago, partly out of professional curiosity and partly because I noticed agencies starting to ask about it. My first attempts were terrible. The lighting was flat, the skin tones were wrong, and the compositions felt random — exactly what you'd expect from a tool that has no understanding of how light actually behaves in physical space.

But then something interesting happened. I started feeding the tools better prompts — prompts informed by two decades of understanding how a Rembrandt light falls on a face, how a rim light separates a subject from a background, how color temperature creates mood. And the results improved dramatically.

The realization was obvious in hindsight: AI image generation is only as good as the visual knowledge of the person operating it. A photographer who understands light, composition, and emotional storytelling will get fundamentally better results than a designer who doesn't.

The Models: What Actually Works in Practice

I've now spent significant time with three main approaches, and here's my honest assessment as a working photographer:

Stable Diffusion / SDXL

This is where I do most of my work. The ecosystem around SD/SDXL — ControlNet for pose control, IP Adapter for style reference, LoRAs for fine-tuning — gives you the kind of precision that a commercial photographer actually needs. When a client wants to see three variations of a campaign concept before the shoot, I can generate them with enough accuracy that the art director can make informed decisions.

The consistency is the key selling point. In advertising, you need images that feel like they belong to the same campaign. SD/SDXL delivers that reliability, especially with trained LoRAs.

Flux

Flux produces the most photorealistic single images I've seen from any model. The lighting quality — and I say this as someone who's obsessive about lighting — approaches what you'd get from a well-executed studio setup. For close-up portraits and texture work, it's genuinely impressive.

The tradeoff is unpredictability. When you combine multiple LoRAs or push complex scenarios, Flux can give you wildly different results between generations. For commercial work where consistency matters, that variability is a problem. For artistic exploration, it's actually a feature.

The LoRA Training Experience

This is where my photography background becomes a genuine advantage. I've trained custom LoRAs on NVIDIA H100 hardware — sessions that run 8 to 10 hours but produce results that capture subtle lighting nuances and stylistic details that faster training methods miss.

The process feels surprisingly similar to developing a visual style in traditional photography: you're teaching the model to see light the way you see it, to prioritize the textures and tones that define your work. It's slow, it's expensive, and it's worth it if precision matters to your clients.

Where AI Fails (And Where Photographers Win)

Here's what no AI model can do, and what I don't expect them to do anytime soon:

Presence. When I photograph Marc Anthony or Juanes, there's a human exchange happening — a trust between photographer and subject that produces expressions and moments no prompt can generate. The micro-expression that happens when a musician hears a playback of their own song during a portrait session. The shift in posture when an executive relaxes because they trust the photographer's direction. These are irreplaceable.

Physical light. AI can simulate the look of a Rembrandt setup, but it can't produce the actual physics of light bouncing off a reflector onto skin. There's a dimensionality to real light — what I explore in depth in The Lighting Playbook — that AI approximates but doesn't replicate.

Accountability. When an agency hires me for a Coca-Cola campaign, they're not just hiring my eye — they're hiring my judgment, my production management, my ability to deliver under pressure. AI is a tool in that process, not a replacement for the professional using it.

How I Actually Use AI in My Workflow

My current workflow combines AI with traditional photography in three specific ways:

Pre-visualization for clients. Before a shoot, I generate concept images that give art directors a tangible reference point. This has cut pre-production meetings in half and virtually eliminated the "that's not what I expected" conversation after the shoot.

Mixed media art. At The Dana Art Gallery, I exhibit fine art pieces that blend AI-generated textures with photographs I've shot on analog film. The collision of the two produces something neither could achieve alone — and it's the most creatively exciting work I've done in years.

Texture and atmosphere exploration. Using ComfyUI as my primary interface, I experiment with atmospheric modifications — adjusting backgrounds, testing color palettes, exploring what a scene would feel like in different lighting conditions — before committing to a physical setup.

The Bottom Line for Photographers

If you're a photographer worried that AI is going to take your job: it won't. Not if you're good at what you do. The photographers who will struggle are the ones who were already producing generic, interchangeable work — because AI can produce generic and interchangeable work faster and cheaper.

But if your value is in your eye, your lighting mastery, your ability to connect with a subject and capture something true — AI is a tool that amplifies those strengths, not a threat that replaces them.

I went from skeptic to practitioner in about six months. The technology is moving fast, and photographers who understand light have a natural advantage in directing it — whether the light is physical or computational.

Diego Sanchez Cadavid is an advertising photographer and music video director based in Miami. His work has been recognized by Lürzer's Archive, Hasselblad Masters, and the Effie Awards. He explores AI-photography integration at ASA Creative and exhibits fine art at The Dana Art Gallery. Author of The Lighting Playbook.