You're going to like this one. Can go from speedo to wet speedo with visible penis line or show some side ball cleavage in a jock strap. You can add a basic bulge with trigger word: underboner, but get creative to do some of the swaps of color or wet/dry or clothing changes. It doesn't seem to let you change the camera angle from my tests, so that's beyond the scope of what this lora can do as far as editing. You can always use a real image edit model to change the camera and then use that one.
Image Editing Lora for Krea 2 - READ THIS
Edit bulges or lack thereof in an image, and even change the garment or color in the process. Try saying it is soaking wet, or extremely tight, for different effects. Version 2 trained on more control and target pairs with a focus on realism, some VPL pairs in addition to the full-on bone-zone target images.
See showcase for examples. I like using the int8 convrot of the raw model with the raw-to-turbo Lora and then 12 steps, cfg 1, exp_heun_2_x0_sde / linear_quadratic .
Custom Node Required
Krea 2 isn't an image editing model, but there are ways to make it behave like one. You will need the Ostris Krea2 Edit custom node and the example workflow that is included with it. I can provide mine if requested, but it is not much different than his. This is not the same as the identity edit node or clothing edit node for Krea 2. You just put the node inline with the model connections and turn on the flag saying the lora is an edit Krea 2 trained lora. https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit He used this uncensored text encoder for the workflow. I put a link to one that should also work in the optional files
I would not use this one in the online generator because of this. I have a normal LoRA for Krea 2 here if you are not editing an existing image.
Description
trained with more data focused on realism
FAQ
Comments (3)
This model (v2) is VERY overfit to a mega ripped guy, enough to create a high percentage of abominations. Example: "A professional photograph of a man wearing a swim brief." (weight 1.0) vs. a non LoRA render.
You might want to try masked loss training to pick out the subject matter within a select number of source images at a much lower number of training steps.
Basically, you create a second image for each source image you want that highlights the area you want to train (white for 100%, black to ignore).
Onetrainer has the option if you train locally, some online trainers may also offer masked loss training. I use the hell out of the technique myself for focused object training.
The control and target pairs is how all these image edit model loras are trained. This one is not a real image edit model so if you read my blurb, you 'll see that it REQUIRES a custom node to work in comfyui as a Krea 2 edit model. This one needs to have an image also. I have a Krea 2 lora that is just regular non-edit version that was not trained with pairs and can be used for text to image workflows. I suspect you just didn't use the custom lora.
and did you see how many of my showcase are not 'mega ripped' guys? But I suspect you just over looked the requirement of a custom node and need to use ComfyUI to generate it with an image that has the crotch somewhere in frame for it to work with.