KreaMAX-Edit is a merge of several custom models I've been working on, both standard and edit variants. One thing to note is that the edit only works with what the dataset was trained on, so overly broad edits will fail. The dataset used was fairly narrow and very targeted so that small changes to images will work well while not disturbing the reference image.
Settings I use are either Euler/Simple or Euler_Ancestral/Beta, CFG 1, 8 steps to just generate an image, 10 steps to edit, with a reference image.
The training data was captioned using a template with this sort of layout, so making your edit prompts, follow this structure for best results.
[1 line of intent, optional]Change [ANCHOR A]'s [attribute] from [old] to [new], keeping [structure/style/light] the same.Change [ANCHOR B]'s [attribute] from [old] to [new], keeping [structure/style/light] the same.Convert [ENVIRONMENT] from [old state] to [new state]: update [sub‑element 1], [sub‑element 2], [sub‑element 3] accordingly; keep [parts of that environment that must stay] unchanged.Do not change anything else — keep [faces/expression], [pose/position], [clothing & props & their colours], [framing/angle/composition/aspect ratio], and [subject/object count] exactly as in the reference.Description
A merge of custom models that include custom LoRA's trained on Civitai, using styling from the LexiVision-Z (Z Image Turbo) model also hosted here.
Comments (3)
Hey I just want to comment:
To use this use it through the exact same workflow you already will use now for the Krea2 Identity Lora.
So just replace the normal checkpoint you use there + the Identity Lora with this new EditModel
I havent done enough A/B testing with it but the editting definitely is working very nicely!
Its hard to say if the results are better than the normal identity Lora though since both seem to work well to maintain identity when you prompt it for very short - 1 beat - changes
I want to give an update to my comment before.
I recommend this model as a Krea2 Edit model.
For very simple basic edits like: 'Replace her skirt with shorts' you are not going to see much difference between this and the normal Turbo + Identity Lora workflows...
But when doing complex actions, like trying to 'get a standing character in an outfit to then change positions and lie on a bed' all while maintaining the facial identity + outfit, you can really see the differences that this model will lock in the characters face a bit better than the normal Identity Lora workflows will...
Thanks very much for trying this out, and for the feedback, really appreciated!