Related notes
How to evaluate a trained LoRA
Test new prompts, compare a baseline and document useful failures.
Write the test before judging the result
Decide what success means for the concept. A character test may focus on identity across viewpoints. A product test may focus on silhouette and materials. A style test may focus on its visual language across unrelated subjects. Avoid judging every concept only by surface attractiveness.
Compare against the base model
Use the same prompt and comparable settings for baseline and adapter generations. Keep a shared seed where the generator supports it, while recognizing that matching a seed is not a guarantee of identical structure. Record adapter strength, base version and the exact prompt.
Test something absent from the dataset
Change the background, composition or pose. For a style, use a new category of object. For a product, try a new surface and viewpoint. If the result only repeats the reference image, the concept may not be flexible enough for the intended use.
Record failures, not just favorites
Save examples that lose the identity, ignore the prompt, repeat one composition or introduce artifacts. Group the problems by cause rather than selecting only the best-looking outputs. For video, examine the action over time instead of judging one attractive frame.
Plan one useful next change
Use failure notes to choose the next experiment: add meaningful variation, revise captions or adjust a supported parameter. Keep versions to explain changes. Once Krea 2 training completes, compare outputs in Generate with the same trigger, seed and aspect ratio. Other models remain draft and test-planning tools.