Suggested tags: Blend the <object> into the image seamlessly
Suggested strength: 0.8-1.0
Suggested usage: Inspired by the Image Fusion model for Qwen Image Edit.
The purpose is to run a single image that already has a roughly cutout image of objects, people, cars, etc and run it through the model to make it blend seamlessly into backgrounds, onto people, all kinds of things. The more simple the task the better it does. See the before/after examples for what it can do currently. The after images should still have their workflow if you need a workflow.
Tried something complex like a UFO floating in a city and it created a dropshadow in the sky unfortunately, but that may be for a V2.
Description
Suggested tags: Blend the <object> into the image seamlessly
Suggested strength: 0.8-1.0
Suggested usage: See Description and example images
FAQ
Comments (17)
What kind of LORA klein_snofs_v1_3.safetensors?
This is amazing, at full strength it does change faces slightly, but very good still! Love it!
Need V2 pls!!! Best lora ever!
I'd be happy to make a V2 if people would tell me what the V1 isn't doing correctly lol. Any input would be helpful for making my V2 Dataset
im here to check for V2, Awesome lora! ♥
Anything you think it could improve upon?
@cranpeach to be honest you did amazing work and it works perfectly so far till now! ,, and looking for more from you, i will let you know if i face anything wrong with it! ♥
I believe you are planning to work on version 2 for this? If so, great and go ahead. I have toyed with lots of generations using this LoRA. You can look at the gallery below.
This lora is already great. But I have a few thoughts about areas that you might consider improving.
1) For realistic humans: increase training samples for blending between subjects & settings from drastically different conditions (day/night, different light condition, etc)
2) Nice to be able to define relative z-ordering of cutout more reliably against other elements in the canvas image, to make cutout actually exist, say, next to a chair and not a little behind, in final render. (See an example with Pikachu and grass in the gallery below. relative z-ordering allows you to put Pikachu in front or grass or among the grass. v2 should be able to do so much more reliably)
I think number 1 is the more important. it should improve human faces very consistent across different context.
question: not sure if i'm using the workflow right. should i add the object before like in photoshop? or should be the backgroudn in 1 input and the object i want to mix in the other image input?
overlay on top
@aungkhant0911 thanks! but in the same workflow with both inputs or externally in photoshop? :O how did you do it?
@Thrill123_ No clue what you are talking about with "same workflow". There is only one input image which is supposed to be cutout laid on top of whatever image. Description clearly mentioned it, plus you can see for yourself in Gallery below.
@aungkhant0911 theres another input node that is bypassed. thought but got it thank you.
Yep you'll need to add the object before like in Photoshop. There is before pictures above for examples
BEST OF THE BEST LORA EVER ON FLUX, just in case anyone wondering .... you can do like photoshop stuff adding another image to image you want, you can use pixaroma image composer node, search it up, I wonder when will you work on v2, i really like it, i have had hard time using qwen edit in this type of workflow tho, it's really helpful!!
hey how many images did u train this lora on i wanna train a similar lora that can blend their action to the image for example if the other person is talking they talk basically interact with the environment and did u trained it in pairs how many pairs by the awesome lora