Unlike Midjourney reference images, Style Reference is usually used as "copying this image". '--sref' mainly borrows style, such as color, texture, light, and medium, and is not responsible for strictly reproducing characters, objects, and compositions. If you want the result to be more similar, first distinguish whether you want the style to be consistent, the characters to be consistent, or the composition is close.
Do not mix the three reference diagrams
If you want to unify the temperament of the picture, use Style Reference, which is '--sref'. It is suitable for making a set of posters, avatars, and illustrations in a unified style, but not for requiring "the same person, the same clothes, the same pose".
If you want the subject or composition of the reference image to appear in the picture, use the image prompt and describe the content you want to generate in the text prompt. Instead of writing imperative descriptions like "copy this style and make it a dog", instead describe the target screen directly, such as "a detailed portrait of a dog".
If you want to keep people, personas, or brand elements consistent, you need to use an entry like a character reference or Omni Reference, not just '--sref'. Style image is not the same as identity image.
How to tune it to be more stable
The first step is to reduce conflicting words. The reference image already has a clear oil painting texture, so don't stack "cyberpunk, film, 3D rendering, realistic photography" in your prompt. The more conflicting the style words, the more diluted the reference diagram is.
The second step is to adjust '--sw'. The higher the value, the stronger the influence of style reference; Too low will be like a normal prompt, and too high may sacrifice the subject description. Try it near the default value first, and then increase it in small steps.
The third step is to use a clean reference diagram. Complex collages, text, mixed with multiple people, and images with too small subjects will make the model not know what to learn. It is best to choose a single image with a clear style, stable lighting and no obvious watermark.
The easiest pit to step on
Don't expect '--sref' to remember a face, and don't cram multiple pictures with completely different styles together. Multi-reference diagrams can mix styles, but they can also cancel each other out.
If you're after "same set but not exactly duplicated", '--sref' is appropriate; If you are pursuing "the same character appears consecutively", you should use the character/subject reference instead, and write the costumes, hairstyles, and props more clearly.
