Train the look once, prompt new subjects

FLORA restored LoRA training and generation for Flux 2 and Krea 2 in September 2026. A LoRA adapts an image model to a particular visual look or concept using example pictures. The point for a designer is to keep an established art direction available when the next brief asks for a new subject. FLORA later added Recraft V4 Styles to the platform's image options. These are existing capabilities, not a new release today.

Think of an illustrator with a series of approved, related ink drawings. Instead of describing the line weight, texture and visual mood afresh for every poster, the illustrator can train a style from their own source images. A later prompt might ask for a train station or a flower in that visual world. This is a proposed application, not a claim that FLORA will reproduce every design detail exactly.

The dataset and model choice matter

Start with images that share the visual properties you want to carry forward, while varying their subjects enough to show the style beyond one composition. Use material you own or are permitted to use for training. FLORA's published example trains a Krea 2 style from ten acupuncture diagrams, then applies it to different subjects. The example shows the intended use, not a universal image-count requirement.

Choose the target model family before training. FLORA's current developer guide treats the training set as a saved Style with a compatible inference model, and directs users to check each trainer's image limits, captioning options and strength range. A style trained for one family is not a universal switch for every image model. The practical comparison is to give each compatible setup the same new subject and judge which preserves the intended look while leaving the subject readable.

Use the saved style, then inspect the result

On a FLORA canvas, add an image node and select a LoRA-capable model. FLORA says the trained style appears in that node's parameter panel. Once training finishes, select the saved style, enter a new prompt and adjust the available influence control for that model. The developer guide describes training as an asynchronous run, so do not assume the style is ready immediately. Training also uses workspace usage; check the current cost shown for your chosen model before starting.

Treat the output as a candidate image. A style can bring the visual language closer while still changing a product's shape, inventing lettering or narrowing the range of compositions. A small comparison across subjects and settings is more informative than one attractive sample. The final brand asset still needs a human check against the approved source work and the actual brief.

Sources & verification

Prepared with AI assistance from FLORA's official materials. The illustration and examples are editorial, not XANTRANS test results.