How I Create My LoRAs Using Imagination and SORA
As of 9/31/2025, OpenAI has removed SORA from ChatGPT subscription. I can't find a reason, but did find mention of SORA2, so stay tuned. In the mentime, just use ChatGPT to create the training set, but create the images one at a time. Sometimes it will say "would you like me to create a set of 20 images?" Just say no thanks. ChatGPT crashes.
As you can probably tell, I enjoy making LoRAs. Over time I’ve developed a simple approach that combines imagination with OpenAI’s SORA, and I’d like to share it here. Hopefully, it gives you a few ideas for your own workflow.
Step 1: Start with Imagination
The first step is all about visualizing the character or subject you want to build your LoRA around. Begin by crafting a simple prompt in SORA, something like:
“Photo of a beautiful blonde woman with blue eyes.”
From there, refine your description little by little. For example:
“Photo of a beautiful blonde woman with long hair, blue eyes, a narrow face, and high cheekbones.”
Keep iterating until you find a face that matches your vision. At this stage, focus only on facial features. Don’t worry about clothing or body details yet—that comes later.
Tip: If you already have a photo of someone similar to what you’re imagining, reference it in your prompt. For instance:
“Create a photo of a beautiful blonde woman with long hair, blue eyes, narrow face, and high cheekbones, similar to the woman in the attached image.”
This helps SORA stay aligned with your concept.
Step 2: Generate Training Images with SORA
Once you’ve nailed the look, the next step is to generate your training data. I usually start with prompts like:
“Create various images of the woman with different poses, different outfits, and different locations.”
I prefer using portrait (2:3) formatting for consistency. After generating a few results, I expand the variety:
“Create close-up facial images of the woman for professional portraits, business outfits, in a photographer’s studio.”
“Create images of the woman walking on a street in different styles of clothing.”
The key here is variety—outfits, colors, and settings should all change. If you lock your dataset into one outfit or one location, your LoRA will struggle to break free of those details, no matter what prompt you use later.
Tips I’ve learned along the way:
Avoid focusing too much on explicit body features. Prompts like “large breasts” or even “beach setting” can get rejected by SORA. Instead, keep prompts natural and general.
Create different views such as "side view", "profile view", or "over the shoulder view"
Aim for 50–100 usable images as training material.
Be selective when creating your training data. Many images that SORA creates, I will not use because they don't represent my ideal target image.
Sometimes SORA will generate one image containing multiple poses. This is actually great—you just need to crop or screen capture each pose into its own image.
SORA will sometimes give you an error saying that it violates their TOC. Many times, asking it to recreate the image will create an image, or you can modify your prompt accordingly to remove anything that might be offensive.
See the attached file for a typical sequence of this workflow.
Step 3: Train Your LoRA with CivitAI
With your curated set of images ready, it’s time to head to CivitAI. The platform has tools, guides, and community-shared tutorials (including plenty on YouTube) that will walk you through the LoRA training process step by step.
Final Thoughts
That’s my process—start with imagination, refine with SORA, and finalize with CivitAI. It’s straightforward, repeatable, and gives you a lot of creative control.
If you end up trying this method, I’d love to see what you create. Drop me a message and share your results!
NeroOnFire has some great tips on LoRA creation at: https://civitai.com/articles/18529