Building the Ultimate AI Disc Golf Photographer (Currently 2% Complete)

Progress update: 2% complete.

Not exactly the headline everyone hopes for when starting a new AI training run, but if there's one thing machine learning teaches you, it's patience.

This latest project is focused entirely on FLUX LoRA training for disc golf, and the goal isn't just to generate good-looking images. The goal is to create what I believe will become the most realistic AI-generated disc golf photography available anywhere.

Why Start Over?

My previous Stable Diffusion workflow produced some impressive images, but after thousands of generations it became obvious where the weaknesses were.

Hands.

Grip.

Release angles.

Anyone who actually plays disc golf can spot an incorrect power grip immediately. The disc sits wrong. Fingers wrap unnaturally. The wrist angle looks impossible.

Those tiny details separate "cool AI art" from an image that could pass as a professional tournament photograph.

So instead of accepting "good enough," I decided to retrain the model with a much stronger dataset focused specifically on authentic disc golf technique.

More Grip Photos = Better Results

One decision became obvious pretty quickly:

More grip photos.

It sounds simple, but it's probably the single biggest improvement I can make.

Every additional example teaches the LoRA how discs should actually be held during drives, approaches, forehands, and putts. The more accurate those references become, the more convincing every generated image becomes.

Sometimes the biggest breakthroughs come from the smallest adjustments.

Experimenting with Weighted Blending

This version also introduces something I've been excited to test: weighted blending.

Rather than relying on a single training approach, I'm blending multiple strengths together so the final LoRA learns exactly what matters most.

The objective is straightforward:

• Better anatomy
• Better grips
• Better disc positioning
• More realistic motion
• Professional sports photography quality

If everything works the way I expect, this should produce the most authentic AI disc golf imagery I've created so far.

The Waiting Game

Training FLUX on Apple Silicon is surprisingly uneventful.

No roaring GPU fans.

No dramatic spikes in system noise.

Just... silence.

That's because FLUX uses Apple's unified memory architecture, sharing memory efficiently between the CPU and GPU instead of hammering a dedicated graphics card. Quiet doesn't mean idle—it simply means the hardware is doing its job differently.

Of course, that doesn't stop me from checking progress every few minutes.

Right now, the model has only reached 2%, which means there's still a long road ahead.

Why I'm Not Running Stable Diffusion at the Same Time

I considered continuing to generate Stable Diffusion images while FLUX trained in the background.

After looking at the memory requirements, it became clear that wasn't the smartest move.

FLUX training can consume roughly 30 GB of unified memory, while Stable Diffusion image generation can require another 8–12 GB.

On a machine with 32 GB total, that turns into a battle over memory rather than productive work.

Instead, I'm letting FLUX have the machine to itself.

The website already has plenty of images generated from previous runs, so taking a short pause in creating new content is a worthwhile tradeoff if it results in dramatically better image quality.

Sometimes the best optimization isn't doing more—it's letting one process finish without competing for resources.

What's Next?

Once training completes, the plan is simple:

1. Test the new Version 2 LoRA.
2. See whether the grip improvements are as significant as expected.
3. Restart the image-generation pipeline.
4. Eventually replace the old Stable Diffusion workflow with a full FLUX-based pipeline.

If everything comes together, the next generation of images should look less like AI art and more like photographs taken by a professional sports photographer standing on the tee pad.

We're only 2% into training.

The exciting part is imagining what the remaining 98% is about to teach the model.