wavymulder/CollageXL

CollageXL SAFETENSOR DOWNLOAD LINK (24.7 MB) - This is an SDXL LoRA trained on collage images.

Use collage style, I recommend at the start of the prompt.

SDXL is much better at making collages than 1.5 was. This LoRA helps push SDXL towards making them a bit more analog.

Compared to Collage Diffusion for 1.5, this LoRA is much more responsive to prompting. One drawback is that sometimes the LoRA will just give an image, without much "collage style". I find best results when my prompt includes two distinct visual elements. In general, the more creative/weird your prompt, the better the result.

Can increase the "hand-made" effect with terms like "arts and craft", "cut and paste", "scrapbook". Can use these terms in negative for opposite effect, though this can also break the collage style.

"sepia" in negative helps brighten up the colors and remove the vintage look.

Please see this document where I share the parameters (prompt, sampler, seed, etc.) used for all example images. All samples are direct outputs from the base model + LoRA, no refiner pass or hi-res fix or anything else.