''' I almost gave up on synthetic media.
Two years ago, I looked at the first images I generated and thought, "This is just a toy." The faces were distorted, the physics were wrong, and the results felt like a cheap Photoshop filter. My first attempts at video were even worse—a flickering, incoherent mess. I was ready to dismiss the whole thing as another overhyped tech trend.
But I have a rule. Before I write something off, I go deep. I mean, really deep. I’ve been lucky enough to have a couple of successful exits, one of which was MovieLaLa, a company deep in the world of visual media. That experience taught me that the next frontier would be creating media, not just curating it. So, I spent the last two years and a significant amount of money not just dabbling, but trying to master this space. I’ve invested in some of the foundational companies in this area, like Anthropic and Scale AI, so I had a front-row seat to the underlying technology. I didn't just want to understand it; I wanted to be able to use it to create professional-grade content.
It was a grind. I made a ton of mistakes. But along the way, I uncovered a few truths that nobody was talking about. These aren't the obvious tips you find in a "beginner's guide to Midjourney" blog post. These are the hard-won lessons that can save you months of frustration and thousands of dollars in wasted effort.
Lesson 1: Your Prompt Is Only 20% of the Final Image
Everyone obsesses over the perfect prompt. They buy prompt packs, study "prompt engineering," and spend hours tweaking a sentence, hoping for magic. That’s a mistake.
The reality is, the initial prompt is just the starting point. It’s the first guess. The real work, the work that separates amateur content from professional results, happens after the first generation. The best creators I know don’t just write a prompt and pray. They iterate. Relentlessly.
My process looks more like a conversation. I’ll start with a simple prompt for, say, a "futuristic cityscape at sunset." I’ll get four variations back. Three are garbage. One has a glimmer of potential—maybe the lighting on one of the buildings is interesting. I take that one image and start refining. I’ll use inpainting to fix a weirdly shaped car, outpainting to expand the sky, and then run it through another model with a low denoising strength to add texture. I might even take it into a tool like Magnific to upscale and add detail, then bring it back for more adjustments.
This is where the art is. It’s not in the words; it’s in the eye. It’s in seeing the potential in a flawed image and knowing which tools to use to bring it to life. I’ve probably spent over 1000 hours just in Stable Diffusion, and I can tell you that the most valuable skill I’ve learned is not how to write a better prompt, but how to see a path from a 60% good image to a 95% great one. Stop searching for the magic words. Start practicing your eye for iteration.
Lesson 2: Video Is a Different Beast—And It’s All About Consistency
If you think generating a static image is hard, just wait until you try to make a video. The biggest challenge isn't just creating a single good frame; it's making 100 frames in a row that all look like they belong in the same universe. This is the problem of temporal consistency, and it will drive you insane.
When I first started, my videos were a nightmare. A character’s face would morph from one frame to the next. A shirt would change color. An object in the background would pop in and out of existence. It was unusable.
I wasted a lot of time trying to solve this with complex prompts. I thought if I just described the character in enough detail—"a man with brown hair, a blue shirt, and a small scar over his left eye"—the AI would get it. It didn’t. The models, especially the early ones, have no memory from one frame to the next.
So, what works? Two things. First, you need a strong reference. This could be a source video (video-to-video) or a very detailed set of character reference images. For one project, I generated over 50 images of the main character from different angles just to have a consistent base to work from. Second, you have to use tools that are specifically designed for consistency. This means using control nets like OpenPose to lock down character movement, or using workflows that re-feed the previous frame’s output as an input for the next one. It’s a technical, painstaking process. There is no "make video" button that works for anything serious. You are not directing a movie; you are wrestling with a machine, frame by frame, to force it into submission.
Lesson 3: The "Best" Model Changes Every Month. Build a Flexible Workflow.
I see so many people go all-in on a single tool. They become a "Midjourney artist" or a "Stable Diffusion expert." That’s a massive strategic error.
Two years ago, DALL-E 2 was the king. Then Midjourney took over for photorealism. Then Stable Diffusion offered unparalleled control and an open-source community. Now we have new models like Ideogram that are better at generating text. For video, Runway was the leader, then Pika, and now Kling is showing incredible results in China. The pace of change is staggering.
If you tie your identity and your workflow to a single model, you’re building your house on sand. The moment a better model comes out, your skills become obsolete.
I learned this the hard way. I had built a complex workflow around a specific version of Stable Diffusion. I had custom models, scripts, the whole nine yards. Then a new architecture came out that was twice as fast and produced better results, but it wasn’t compatible with my old setup. I had to rebuild everything. It took weeks.
Now, my approach is completely different. I think in terms of a "stack," not a single tool. My workflow is modular. I might use Midjourney to brainstorm initial concepts because it’s fast and creative. Then I’ll take the best idea and jump into Automatic1111 (a Stable Diffusion UI) to get fine-grained control over the composition. I might use a specific fine-tuned model for the character, another for the background, and then composite them together. For video, I might use one tool for motion and another for style transfer.
Don’t be a fanboy of one tool. Be a mercenary. Use the best tool for the job at hand, and be ready to drop it the second a better one comes along. Your value is not in knowing how to use a specific piece of software; it’s in your ability to achieve a specific result, regardless of the tool.
Lesson 4: Stop Chasing Photorealism. The Real Opportunity Is in Stylized Content.
Everyone’s first instinct is to try and create a perfect, photorealistic image. It’s a natural benchmark. Can this thing make something that looks real? And yes, it can. But that’s a trap.
First, it’s a crowded space. Everyone is trying to do it. Second, it’s where the "uncanny valley" is most dangerous. The closer you get to reality, the more jarring the small imperfections become. A slightly weird hand or a subtly wrong shadow can ruin the entire image.
But most importantly, photorealism is often boring. It’s a copy of the world we already have. The real magic of synthetic media is in creating worlds that don’t exist. It’s in developing a unique, stylized aesthetic that is instantly recognizable as your own.
Think about the most successful animated studios, like Pixar or Ghibli. You know one of their films the moment you see it. They have a signature style. That is the opportunity we have now as individual creators. We can build our own "style models."
I’ve had the most success, and the most fun, when I’ve leaned into this. I created a custom model trained on a set of vintage sci-fi book covers. The result is a look I couldn’t get anywhere else. It’s not photorealistic, and that’s the point. It’s distinctive. When you see an image made with it, you know it’s mine.
Stop trying to replicate reality. Invent your own. That’s where the real value is, and it’s a lot more defensible than just being good at making a fake photograph.
Lesson 5: The Technical Barrier Is Your Moat
I hear this all the time: "I’ll just wait until the tools are easier to use." That sounds logical, but it’s a losing strategy.
Yes, the tools are getting easier. There are more web UIs, more one-click solutions. But the moment a tool is easy enough for everyone to use, the value you can create with it drops to zero. If anyone can do it, it’s not a skill. It’s a commodity.
The technical complexity is not a bug; it’s a feature. It’s a filter. The fact that you have to install a GitHub repo, troubleshoot dependencies, and learn a bunch of arcane settings is what keeps most people out. That difficulty is your moat. It’s what creates the opportunity for you to provide a service that others can’t.
I’ve made over 200 angel investments, and I look for founders who run towards the hard problems, not away from them. The same applies here. The people who are willing to spend the time learning the difficult, technical, and frustrating parts of this field are the ones who will be in the best position to reap the rewards.
Every hour you spend in a command line interface, every time you debug a Python script, you are building a competitive advantage. Don’t resent the difficulty. Embrace it. The barrier to entry is your friend.
The Real Work Is Just Beginning
After two years, I can confidently say that synthetic media is not a toy. It’s the most powerful creative tool of our generation. But it’s not magic. It’s a craft. It requires patience, technical skill, and a good eye.
Don’t get distracted by the hype or the "get creative quick" schemes. The real work is in the iteration, the technical details, and the development of a unique point of view. If you’re willing to put in that work, you won’t just be able to create amazing things—you’ll be building a skill that is scarce, valuable, and deeply rewarding. '''
Frequently Asked Questions
Are these recommendations still relevant in 2026?
Absolutely. While specific tools and tactics change, the underlying principles remain consistent. I update my thinking regularly based on what I'm seeing in the market and across my portfolio companies.
Can I implement all of these at once?
I'd strongly recommend against it. Pick the 2-3 items that resonate most with your current situation and focus there. Trying to do everything simultaneously is a recipe for doing nothing well.
Which item on this list has the highest impact?
It depends on your stage and context, but in my experience, the items near the top of the list tend to have the broadest applicability. That said, sometimes the less obvious items create the biggest breakthroughs for specific situations.
How were these items selected?
Each item on this list comes from direct experience, either from building my own companies or from patterns I've observed across the 200+ startups I've invested in. I prioritize practical, actionable items over theoretical concepts.