Prompts Tell the Model What to Make. LoRA Changes How It Sees.
Both prompts and LoRA training shape your AI outputs — but they work at completely different levels. Understanding the difference helps you know when each one is worth using.
Once you’ve built a moodboard, RandomSeed gives you two ways to use it: prompt fragments and skills, or a trained LoRA. Both shape your outputs. But they do it differently — and understanding that difference helps you know which one to reach for.
What Prompts Actually Do
When you add a prompt fragment — “warm terracotta tones, soft diffused light, linen texture” — you’re giving the model language and asking it to interpret that language into pixels. The model already knows what “terracotta” looks like from its training data. You’re steering it toward that knowledge.
This works well for concepts the model understands deeply. Common aesthetics, recognizable styles, named color families — the model has seen thousands of examples of these and can render them reliably. Prompt fragments are fast, free, and easy to adjust. You change a word, you change the output.
The limitation shows up at the edges. If your aesthetic is unusual — a specific combination of qualities that doesn’t have a name, or a visual feeling that’s hard to put into words — prompts can only approximate it. You’re always translating from visual intuition into language, and something gets lost in that translation.
What LoRA Training Actually Does
A LoRA skips the translation step entirely.
Instead of describing your aesthetic, you show it — through the images on your moodboard. The training process adjusts the model’s internal weights based on those images, teaching it to recognize and reproduce the visual patterns they share. The specific color relationships. The texture quality. The way light behaves in your references. The compositional tendencies you keep returning to.
These patterns become part of how the model generates, not just suggestions it tries to follow. When you generate with your LoRA active, those qualities show up without you asking for them — because they’re baked into how the model is looking at the problem.
Does It Actually Make a Difference?
Yes — but how much depends on what you’re going for.
For common aesthetics, the gap is small. If your board is built around clean minimalism or moody dark tones, prompts can get you most of the way there. The base model already knows these styles deeply. A LoRA adds refinement, not transformation.
For distinctive aesthetics, the gap is real. If your visual world is specific — a particular film stock quality, an unusual color relationship, a recurring compositional choice that doesn’t have a name — prompts will keep approximating. They’ll get adjacent to what you want without quite landing. A LoRA trained on your actual references can capture that specificity because it learned it directly.
Consistency is the other difference. Prompts drift. Change the wording slightly, use a different seed, and the output shifts. A LoRA applies the same learned aesthetic every time. If you need ten generations that feel like they came from the same creative direction, a LoRA makes that easier.
A Useful Analogy
Think of it like briefing a photographer. Prompts are like giving verbal direction before the shoot: “warm light, natural textures, give it a quiet feeling.” A LoRA is like showing them your entire reference folder, letting them study it, and then asking them to shoot in that spirit. Both are valid. The first is faster. The second gets closer to the thing in your head.
They Work Best Together
The most consistent outputs come from combining both. The LoRA handles the aesthetic baseline — the feel and look that’s hard to describe. Prompt fragments and skills handle the specific content and direction of each generation. One sets the visual world; the other navigates within it.
In RandomSeed skills, you can configure both at once: enable LoRA for a board, set the scale (how strongly it influences the generation), and layer prompt fragments on top. The LoRA pulls toward your aesthetic; the fragments guide what you’re actually making.
When to Use Each
Start with prompts and skills if your aesthetic is relatively common, if you’re still exploring your visual direction, or if you want to iterate quickly. Prompt fragments are instant and costless — refine them as you generate.
Train a LoRA when your aesthetic is specific enough that prompts keep missing the mark, when you need consistent results across many generations, or when you’ve identified a quality in your references that you can’t quite put into words. That unnameable quality is exactly what LoRA is designed to capture.
A board needs at least 10 analyzed images to train a LoRA — open your moodboard and check the Brief tab once you’re there. If you’re not at 10 yet, keep adding references. The stronger the board, the better the model learns.
Build Your Moodboard
Open Moodboards, upload references that define your visual world, and generate a style brief. When you’re ready to go deeper, train a LoRA — and see what happens when the model learns your aesthetic directly.
Frequently Asked Questions
What is a LoRA in RandomSeed?
A LoRA (Low-Rank Adaptation) is a fine-tuned model trained on your moodboard's images. It learns the visual patterns in your references directly — color relationships, texture qualities, compositional tendencies — and applies them to every generation, without you needing to describe them in a prompt.
How many images do I need to train a LoRA?
You need at least 10 analyzed images on your moodboard to train a LoRA. More images give the model a richer picture of your aesthetic, but quality matters more than quantity — choose references that clearly represent the style you want.
Does training a LoRA replace prompts?
No — they work best together. The LoRA handles the hard-to-describe aesthetic baseline (the feel, the color language, the texture quality). Prompts and skills control the specific content and direction of each generation. Combining both gives you the most consistent and on-brand results.
When is a LoRA worth training?
LoRA is most valuable when your aesthetic is distinctive or difficult to describe in words, when you need consistent outputs across many generations, or when prompt variations keep drifting your results. If prompts are already getting you close, LoRA is a refinement — not a requirement.
Can I use LoRA and prompt fragments at the same time?
Yes, and that's the recommended approach. RandomSeed skills let you configure both: enable the LoRA for your board and add prompt fragments that layer on top. The LoRA provides the aesthetic foundation; the fragments refine the specific generation.