Reference images, brand kits, or fine-tuning: how to lock AI to your brand

If your AI images keep coming out almost-right but not quite yours, the fix is not a better prompt. It is choosing how you feed the brand into the model in the first place. There are three real options, and they are not interchangeable: attach reference images at generation time, hand the model a structured brand kit it reads on every run, or fine-tune a model on your own images so the look is baked in.

Here is the short version before the detail. Start with a brand kit plus reference images, because together they cover most of what "on brand" means and neither requires training. Reach for fine-tuning only when you have a signature style or a recurring character that references keep drifting on, and only when you have a clean set of examples to train on. Fine-tuning is the most powerful of the three and the easiest to over-invest in.

What are the three ways to lock AI to a brand?

A reference image is an example you attach to a single generation. The model looks at it and borrows its style, subject, or composition for that one output. Google's Gemini image models take up to 14 reference images in a request, including up to 3 used specifically as style references and up to 5 to hold a character consistent, per the Gemini image generation docs. Fal's Flux 2 processes up to four reference images per generation. No training, no waiting; you supply the examples and the model uses them right then.

A brand kit is the part of your brand you can write down, stored so the model reads it automatically. Colors as exact hex values, fonts, tone rules, do-and-don't notes. It rides along with every prompt instead of living in someone's head. Some models read this precisely: Flux 2 understands hex codes directly, so #FF5733 yields that color rather than a close cousin, according to fal's Flux 2 developer guide. We cover the concept in more depth in what is an AI brand kit.

Fine-tuning teaches a model something it did not know by training a small add-on, usually a LoRA, on a set of your images. The result is a reusable file that applies your learned look across many generations. Fal's guide recommends around 15 to 30 images for a product or subject LoRA, and notes training finishes in minutes on their infrastructure. The catch is that you need those images first, and you retrain when the look changes.

The three methods at a glance

Reference images: fast, flexible, re-supplied every run

Reference images are the quickest way to steer output toward your brand, and for most teams they do more work than any prompt tweak. Instead of describing your lighting or your product angle in words, you show it. The model matches what it sees.

The limits are practical. There is a ceiling on how many references a model accepts, and each generation starts fresh, so the examples have to come along every time. Style also holds steadier than identity. Establishing a consistent palette or illustration style across a set is fairly reliable; keeping one specific face or product identical across many scenes is harder, and holding a mascot on-model is its own problem. For a campaign that needs a coherent look rather than a pixel-identical mascot, references usually get you there.

A brand kit: the spec the model reads every time

A brand kit handles the parts of your identity that are exact. Your primary color is one hex value, not "sort of orange." Reference images are bad at this, because a JPEG of your brand color is at the mercy of the model's interpretation, but a hex code fed to a model that respects it is not. That precision is the whole point: it removes the color-correction loop where you generate, check, adjust, and regenerate.

What a brand kit cannot do is carry the parts of a style you cannot put into words. It will not describe the particular way your photography crops a product or the softness of your lighting. That is reference-image territory. So the kit and the references are complements, not rivals: the kit locks the specifiable attributes, the references carry the feel. Feeding both into the generation step, rather than fixing output afterward, is the core move in our playbook for keeping AI images and video on brand.

Fine-tuning: teaching a model a look it can't get from references

Fine-tuning is the heavy option, and sometimes the right one. When you have a genuinely distinctive style, or a mascot or product that has to stay identical across hundreds of assets, a trained LoRA holds it more tightly than references re-supplied each run. You train once and reuse.

It also has real failure modes, which is worth knowing before you commit a week to it. Research on LoRA-based style transfer documents that these methods "still face significant challenges such as content inconsistency, style misalignment, and content leakage," per the ConsisLoRA paper (arXiv, 2025). Content leakage means bits of your training images bleed into outputs where they do not belong. In practice that means a fine-tune is only as good as its dataset: inconsistent training images produce inconsistent results. You also retrain when the brand shifts, and a LoRA is tied to the base model it was trained on, so a model upgrade can mean redoing the work. Powerful, but not free.

Which should you reach for first?

For most marketing and ecommerce teams, the order is: brand kit, then reference images, then fine-tuning only if the first two leave a gap.

The brand kit comes first because it is a one-time codification that pays off on every generation and stops the most common form of drift, which is off-model color and ignored rules. Reference images come next because they cover the feel that a written spec cannot, and they need no setup. Between them, these two handle the large majority of "make it look like us" without anyone training anything.

Fine-tuning is for a specific situation, not a default. If your brand rides on a signature illustration style that no combination of prompt and reference reproduces, or a recurring character that has to be exactly the same every time, and you have a clean, consistent set of images to learn from, train the LoRA. If you are reaching for fine-tuning to fix generic-looking output, a brand kit and better references will almost always get there faster and cheaper. If you are still diagnosing why the output looks generic in the first place, start with why your AI content looks off-brand, and for the wider framework see the complete guide to on-brand AI content.

How this looks in Orisu

Orisu is built for the first two methods to work together and for the third to plug in when you need it. A brand kit stores your colors, fonts, and rules once and feeds them into every generation node on the canvas, so the specifiable attributes are enforced without anyone pasting a hex code by hand. Reference images attach at the generation step alongside the kit, so the feel and the spec arrive together rather than one being patched on after.

When a project genuinely needs a fine-tune, a trained model slots into the same workflow as any other generation step, so the LoRA runs behind the same brand kit and the same review gate as everything else. The point is not to pick one method forever. It is to use the cheapest one that holds the brand, and to escalate only when the work in front of you demands it.