AI models but they generate images of your dogs, and its like ridiculously easy to make. How freaking awesome is that?
The Idea
It all started like 8 months ago when FLUX.1 by Black Forest Labs was blowing up. If you’re unfamiliar, FLUX.1 is this badass open-source image generation model known for its surprisingly good ability to generate legible text and realistic people (including decently accurate images of celebrities). I’ve dabbled in AI image generation before, had some time to kill, and thought to myself — “Hey, what if I make hilarious AI generated images of my dogs and send them to my lovely fiancée?”
I thought I’d need thousands of images to train the model. Nope. Turns out, thanks to LoRa (Low-Rank Adaptation), you only need around 10–15 images. LoRa basically acts as an “adapter” to the existing FLUX.1 model—like DLC, but for neural networks.
Oh and fair warning, you’re going to see lots of AI generated images of my dogs in this post; so here are some real photos of them also for you to compare to:
Doing Things The Hard Way
Thanks to some YouTube videos, I found a website that would handle the training for cheap, like $5. It required manual labeling of images, which wasn’t too bad honestly, and the compute resources at the time were pretty slow. Training Sadie’s LoRa model took about an hour and resulted in a ~175MB model file.
Following the guides, I installed ComfyUI—a node-based workflow tool that’s powerful but tedious to set up—I managed to connect the LoRa model to the really big (like 25+GB) FLUX.1 base model. ComfyUI is like “Blueprints” in Unreal Engine or “Workflows” in HubSpot, but more confusing and you have to install it yourself.
After a good amount of trial and error, and way more time than I’d like to admit, I got it to work! I typed prompts, and boom—there was Sadie, my sweet chihuahua poodle, in all sorts of predicaments. Here are some of my favorites:
Revisiting the Project (The Easy Way)
When I launched this blog, I was brainstorming different things to write about and this one came to mind. I finally got around to it and tried to find the old files and photos I made. After some digging, in my Hugging Face account, I found Sadie’s LoRa model, still ready to go. I also discovered that Hugging Face made things stupid easy—no downloads required. You can literally just generate images directly in the browser now? I don’t know why it didn’t click for me before, but a switch definitely flipped in my head, and I feel like I unlocked the full breadth of what Hugging Face has to offer.
Unfortunately, Stella’s model was MIA. But during a Google deep dive, I rediscovered the original site I’d used. The training feature I used for her model was now deprecated but redirected me to their “fast training” service. With $5 still sitting unused in my account, I gave it a go.
This new method was absurdly easy. I uploaded 12 Stella pictures, chose 4,000 steps of training, clicked run, and spent like $4. Less than twenty minutes later—no exaggeration—I had a fully trained LoRa model, instantly uploaded to Hugging Face via my access token. The only downside is that it seems like it has a hard time putting her in outfits or picturing her with famous people. The Sadie model does not struggle with that.
Here are some of my favorite Stella photos:
Hugging Face Spaces (The Ultimate Shortcut)
My next goal was embedding the AI generation tool directly into this blog post, because who wouldn’t want to generate funny pictures of my dogs? Hugging Face Spaces seemed perfect, except I wasted half an hour trying (and failing) to configure Sadie’s model. AI definitely led me astray here…
Then it hit me, and I feel really stupid for saying this honestly (because it should be so so so apparent for many people) — but why reinvent the wheel? Hugging Face Spaces are open source. I found someone else’s FLUX.1 LoRa adapter, shamelessly copied their app.py, tweaked it to point to Sadie’s model instead, updated the requirements.txt, and voilà—it worked perfectly. In 30 frustrating minutes, I failed spectacularly. In 5 glorious seconds of copy-pasting, success.
JBigs, I’m Jealous and Want a Tutorial
Okay okay, I hear you. It’s actually really easy, just follow these steps:
- Gather like 10-15 images of your subject. Try and have only them in the frame.
- Go here and make an account and add like $5. Upload your images. For Trigger word, put the name of your subject — I used “Sadie the dog” and “Stella the dog”. Expand Advanced Settings and set # of Steps to 1000 – 4000 steps.
- When it’s done, click Run Inference. You can type your prompts straight in there and it will generate images for you of your subject.
That’s the noob friendly tutorial, I’ll spare you the details on connecting to Hugging Face, locally running the model, adding in other models, etc.
The Takeaways
- Open-source is the best. Seriously. Someone else has already figured it out, so leverage their work. Just because you can ask ChatGPT or Claude to code you a script, doesn’t mean its the best approach. I could have saved myself 30 minutes if I just turned off the lazy vibe coding side of my brain, thought critically, and realized I could have just grabbed someone else’s script right away.
- Revisiting old projects doesn’t have to be a painful experience. I was kind of dreading writing this blogpost because its another one of my past projects from a while ago, where I have to dig up old files and try and think through what I did at the time. This was fun because I actually get to see just how much the process and space has improved since I last worked in it — AND I learned something new.
Wanna Try It Out?
I set up Hugging Face Spaces so you can generate images of Sadie and Stella yourself. Knock yourself out. If you make anything cool, please send them to me.
Here is Stella’s image generator:
Here is Sadie’s image generator:
Oh, and P.S.
No neural network could ever capture the raw unhinged energy of Sadie in this moment. Some things are just too powerful for artificial intelligence:
