Okay, I know I usually talk about AWS bills and Datadog dashboards, but I’ve been deep in the AI image gen space for a project (trying to automate some asset creation, long story). So I gave Luma Dream Machine a solid run this past week.
My hot take: their much-hyped "photorealistic" model isn't ready for human subjects yet, at least not for professional use. It's amazing for objects, landscapes, architecture. But people? It feels like it's hitting an "uncanny valley" checkpoint it hasn't fully passed.
Here’s what I mean. I ran a series of prompts for a simple "photorealistic portrait of a 30-year-old woman in a cafe, natural light." The composition and lighting were often great. But the facial details… inconsistencies kept popping up. Slightly misaligned eyes, weird ear structures, hands that looked fused with the coffee cup. It’s like the model has a great grasp of macro scene composition but stumbles on the micro-anatomy.
I even tried to treat it like tuning a cloud monitoring alert—iterative refinement. I used very detailed, SRE-like step-by-step prompting:
```
A photorealistic portrait. Subject: a human woman, age 30. Scene: a cozy cafe. Lighting: soft window light from frame left. Key details: realistic skin texture including pores, subtle eye reflections, separate fingers holding a ceramic mug, natural hair strands.
```
The result was *better*, but still had this subtle "off" feeling. One eye was slightly larger, and the teeth in her slight smile were a blurry mess. For a hobbyist, it’s cool. But if you need reliable, consistent human generation for anything public-facing, the risk of a weird artifact is still too high.
It reminds me of early days with a new observability tool where the graphs look beautiful but the underlying data resolution isn't fine-grained enough to trust for on-call alerts. The foundation is impressive, but the precision for human features isn't production-ready.
Has anyone else pushed it on human-centric prompts? Curious if you've found a prompt "workaround" or if we're all just waiting for the next model version.
cost first, then scale
Totally get what you mean about the hands and the micro-anatomy. It's funny, I see a parallel with some ML inference pipelines we run on AWS - you can have a model that's 99% accurate on the overall task, but that last 1% of edge cases (like human joints or facial symmetry) is where all the cost and complexity hides. The model might be "photorealistic" in aggregate, but fails on the specific details our brains are wired to scrutinize.
Have you tried mixing it with a dedicated portrait model in a pipeline, or is that breaking the "one tool" simplicity you were going for? The iterative, SRE-like prompting is a great approach though.
cost first, then scale