This morning I was skimming through 20,675 images on my Nitro 5 and slowly realizing what I was looking at.

They're crops. Every enemy, every UI element, every frame of Darwin the crab that appeared in 904 frames of Everything Is Crab gameplay. OWLv2 found them, cropped them, saved them. I was the one who ran the process. I wasn't the one who did the work.

The quality surprised me. That's not the word I would have used before I saw the results. I expected noise — blurry half-frames, clipped edges, garbage detections. Instead I got clean isolated objects, consistent enough to train on. The final run produced something I didn't expect to hold up.

It cost $9 to get there.

That number matters because of what it includes. Not just the final run — the wrong approaches, the failed attempts, the session where I learned what I actually wanted before I knew how to ask for it. The $9 covers all of it. Another $10 gets me a reusable model trained on those crops. That model is also portfolio material. That model is also the kind of thing a developer community finds interesting.

The path to those 20,675 images wasn't obvious. The idea existed before the method did. I knew I wanted to automatically understand what was happening in gameplay footage — to replace the part of my ContentPipeline where I manually scout timestamps worth clipping. But "understand gameplay footage" is too far to reach for. You can't build toward something that vague.

What made it possible was finding the four models: CLIP for scoring visual interest, BLIP for generating descriptions, OWLv2 for detecting and cropping objects, YOLO for fast inference once trained. That stack made the idea concrete. It gave me four specific problems to solve instead of one impossible one.

The second wall was hardware. OWLv2 on a laptop CPU is slow enough to matter. I'd run into this ceiling before on other ideas — the interesting thing is right there, the compute to reach it is not. RunPod answered that. Rent a GPU for an evening, run the process, download the result, close the pod. The RTX PRO 4500 handled 904 frames overnight. Total bill: $9.

RunPod did the same thing for me that Ollama did when I first ran a local model, and OpenRouter did when I realized I didn't have to host anything to access capable APIs. It moved a wall. The wall was cost. The solution wasn't cheaper hardware — it was rented hardware at the moment I needed it. That pattern keeps showing up: find the idea, find the challenge behind it, find the cost behind the challenge, accidentally find a cost-effective alternative later and circle back.

I waited for the download to finish overnight. Closed out the pod in the morning. Sat with the folder of 20,675 crops and understood that I'd solved one problem and found a larger one directly behind it.

The ContentPipeline I've been running since May produces YouTube Shorts automatically. I play a game, record it, the pipeline clips it, narrates it, schedules it. The human step that remains is scouting timestamps — deciding which moments in the footage are worth clipping. I do that manually right now. It takes attention I'd rather spend elsewhere.

SpriteHarvester is the other side of that pipeline. Not the output side — the intelligence side. A model trained on those 20,675 crops knows what Darwin looks like. It knows what a boss health bar looks like. It knows what an evolution card looks like. Point that model at new footage and it finds the timestamps automatically. A boss appeared at 4:23. Darwin's health dropped to critical at 7:51. The evolution UI triggered at 11:09. Those are the clips.

The pipeline already exists. The intelligence layer is what was missing. $9 of GPU time and one overnight session built the dataset that trains it.

If this works the way I think it will, the method shifts. I play Everything Is Crab. Shorts come out the other end. No manual review, no timestamp scouting, no production decisions.

I don't know exactly what else RunPod unlocks. I know it solved a cost problem I'd been working around. That's usually how it goes — the door opens before you know what's on the other side.

If you're building something similar or working through your own pipeline problems, the intake page is there.