This LoRA was trained with a very specific phrase, and tested with Flux outpainting to turn everyday photos into what looks like Photoshopped works of art. I think there still needs to be a v2 for it to work great with outpainting but it does okay with normal generations for now.
An example prompt would be
This pair of images transitions a professional stock photo into a whimsical goofy accident; The [ORIGINAL] image shows a professional stock photo of a woman with red hair facing the viewer looking emotionless, while the [ALTERED] image shows the same woman but instead of her face is a scary hole in her face, which shows millions of tiny teeth inside.
You'll notice a few things:
1) We introduce that there are two images in the composite.
2) We give a small description of what will happen.
3) We mention the [ORIGINAL] image
3) We mention the [ALTERED] image
Grab the workflow from this photo
Description
FAQ
Comments (9)
I'd love to learn more about your training process for this, could you share?
I did not train it the way you would train the other in context loras, which I need to study and learn to do because then I think inpainting would work better, however, the current training was simply 30+ images composed of BEFORE and AFTER shots. I went on YouTube, online Photoshop tutorials, etc, and went into Photoshop and took their original photo and the finished product and just made 30 of them. Then I wrote the captions on all of them the same way.
Can you share the recommended resolution for this LoRA?
I do not have a recommended resolution since the training data has many different resolutions. I would try some of my examples however.
Just to be clear, this is a LORA for Flux Dev, not for Flux Fill?
It's kinda confusing now, having the official inpainting flux model.
Could you please provide a ComfyUI workflow?
Sorry for the late reply. I trained this Lora with Flux Dev. Here is a photo with working workflow. https://civitai.com/images/41699120
@Ugleh No worries, thanks!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.