QWE AI Academy Picture a single frame: a lone figure standing on a desert ridge, twin moons overhead, every line clean and confident, the color palette flat but glowing. It reads instantly as Moebius. Not “sci-fi illustration.” Not “French comic.” Moebius.
That’s the goal here – getting AI art in specific artist styles that’s recognizable, not just vaguely inspired. There are two real ways to do it in 2026, and one is much better than the other for most people. Here’s the short version, then the long one.
The takeaway, upfront
If you want a one-off image in a famous style: use Midjourney’s --sref with a reference image or style code. Five minutes, no training.
If you need a consistent, deeply specific style across dozens of images: train a LoRA on Stable Diffusion. Two hours of setup, but you own the result.
Prompts that just say “in the style of [artist name]” – the move every tutorial recommends – are the weakest option. Style-name prompting is basically a guessing game: which model version are you on? Which encoder was it trained with? The artist’s name might not even be in the text encoder anymore. The other two methods sidestep this entirely.
Why “in the style of” stopped being enough
Early Stable Diffusion learned artist names because the training captions literally contained those names. Type “by Greg Rutkowski” into SD 1.5 and you’d get convincing oil-painted fantasy. Then things changed.
Stable Diffusion 2.x switched to OpenCLIP as its text encoder – and OpenCLIP’s training data had far fewer artists and celebrities than the original CLIP used in SD 1.5. The result, as AssemblyAI documented: many canonical prompting methods from SD 1.5 are essentially obsolete in SD 2.x. That’s why a subset of users still prefers 1.5 or SDXL over version 2 to this day (as of early 2026).
Rutkowski himself opted out of the Stable Diffusion training set. The community responded by training a LoRA that mimics his style and posting it on Civitai – which tells you something about how these systems work. Strip an artist out; someone trains them back in.
Method A: Midjourney style references (the fast path)
Feed Midjourney an image whose style you want to borrow, or use one of its internal style codes. The basic syntax: --sref at the end of your prompt, followed by an image URL or a numeric code.
Style weight --sw controls how strongly the reference influences the result – according to Midjourney’s official docs, it accepts any value from 0 to 1000, with a default of 100 (as of early 2026).
a lone figure on a desert ridge, twin moons --sref https://your-image-url.jpg --sw 250Styles can be blended, too. The syntax: one --sref parameter followed by multiple codes, each weighted with :: and a number – higher numbers mean more influence. A common technique is pairing one sref for texture with another for color palette (source: sref-midjourney.com tutorial).
When you reference an image, don’t pile on style words in the text prompt too. Keep the text describing what you want to see – not how to render it. Two competing style signals usually produce mud.
No reference image? Midjourney’s Style Creator (on midjourney.com only) lets you build a custom sref code by picking from grids of preview images. Most styles stabilize after 5-10 rounds; past round 15, changes are subtle. The output is a reusable code you can drop into any future prompt.
Method B: Training your own LoRA (the deep path)
LoRA – Low-Rank Adaptation – bakes a style into Stable Diffusion permanently. The technique was introduced in a 2021 paper by Hu et al. and works by modifying only the cross-attention layers of the base model – the part where image and prompt signals meet. That’s why LoRA files stay small: typically 2-200 MB versus several gigabytes for a full checkpoint.
The workflow, condensed:
- Gather images. The minimum for SD 1.5 style training: 15 images, each at least 512×512, PNG or JPEG (source: stable-diffusion-art.com, as of early 2026). Variety matters more than count – different subjects, same visual language. In practice, more images covering a broader range of subjects gives a more flexible LoRA; all portraits and it’ll fight you when you prompt for landscapes.
- Caption with a rare token. Use a string like “skw” that the model doesn’t already associate with anything. Captions look like “drawing in skw style.” The token position matters – more on that below.
- Train. Easiest path: a Google Colab notebook or a managed service. On Replicate (as of early 2026): upload a zip, set the task dropdown to “style,” training runs in roughly six minutes and outputs a .safetensors file.
- Use it. In AUTOMATIC1111 or ComfyUI, load the .safetensors and trigger it with the keyword plus a strength multiplier between 0 and 1.
Why this beats prompting: a well-trained LoRA captures things you can’t put into words – line weight, how shadows fall, the artist’s idiosyncratic shorthand for clouds. It also stops drifting between generations.
The version trap
Both methods have a quiet failure mode that wrecks reproducibility. Most tutorials don’t mention this.
On Midjourney: the Style Reference feature was updated for V7, and old style codes saved before mid-2025 may not produce the same styles anymore. To use those codes reliably, add --sv 4 to your prompt or switch to V6 (per Midjourney’s Style Reference documentation). That perfect code you archived? It might quietly render differently today unless you pin the version.
The catch: Midjourney runs two parallel sref systems that share version numbers but behave differently. Random sref codes only work with --sv 4 (legacy) or --sv 6 (newer) – try versions 1, 2, 3, or 5 and you’ll get an error. Upload your own image as a reference, though, and all six versions work (source: geekycuriosity comparison testing, 2025).
For LoRAs, the equivalent problem is base-model mismatch. A LoRA trained on SD 1.5 won’t work properly on SDXL or FLUX – those models have different architectures, and the cross-attention layers the LoRA modifies don’t exist in the same shape.
Caption order: the detail nobody writes about
Where you put the rare token inside a caption changes what the model learns to associate with it.
“Photo of skw man wearing a suit” binds skw to the man. “Photo of a man wearing an skw suit” binds skw to the suit. Same token, completely different association – documented in ViewComfy’s LoRA training guide.
For a style LoRA, your captions should put the trigger token in a position that’s clearly modifying the image overall, not a specific object in it. “Skw style, a landscape with mountains” is safer than “a landscape in skw style with mountains” – in the second version the model might decide “skw” is a property of the mountains specifically.
The ethics question
Copying a living artist’s style without consent is legally murky and ethically uncomfortable. Microsoft Copilot’s official prompting guide explicitly asks users not to reference living artists in prompts. Separately, Spawning – the group behind haveibeentrained.com – gave artists a formal opt-out mechanism before Stability AI began training Stable Diffusion 3 (reported by MIT Technology Review, 2022).
Personal experimentation in the style of long-dead Old Masters? Fine. Building a commercial product on the back of a living illustrator’s work? That’s a different conversation, and courts are starting to weigh in.
FAQ
Can I get a specific artist’s style on DALL-E or Copilot?
Not for living artists – those tools filter explicit artist names. For movements (“Art Nouveau,” “Bauhaus”) or long-dead artists, yes.
How many images do I really need for a style LoRA?
15 is the floor, not the target. The real problem with exactly 15 images isn’t quantity – it’s that most people grab 15 of the same type of subject. Say you pick 15 portraits. The LoRA learns “portrait + style” as one concept. Then you prompt for a landscape and get something that feels off, because the model never saw the style applied to anything else. Mix subjects: portraits, objects, environments, abstract compositions. Keep the style constant. That variety is what makes a LoRA actually flexible, and it matters more than hitting a higher image count.
Why does my output look like the reference image instead of just borrowing its style?
There’s a common misconception here: people assume --sw is purely about “how much style” – but at high values it also pulls subject composition from the reference. Your reference image’s subject is bleeding through, not just its aesthetic. Drop --sw to somewhere in the 50-100 range and check if the composition loosens up. If it’s still happening, the reference itself is the issue – the subject in it is too dominant. Pick a reference where the style is clearly the main feature: abstract art, texture studies, or a piece with a very generic subject.
Try this now
Open Midjourney, pick an image you love from an out-of-copyright collection (Wikimedia Commons works), grab its URL, and run one prompt with --sref [url] --sw 100 and the same prompt with --sw 400. The difference between those two outputs will teach you more about style transfer than another tutorial will.