Please consider buying my two albums Sisters of the Wicked and Danger Ahead, my book Dracula and also please check out my YouTube channel. Civitai won't support creators, so any purchase is a real help. Thanks.
DRACULA https://www.amazon.com/dp/B0FM4DXYPL
Sisters of the Wicked https://amazon.com/music/player/albums/B0DSJY3ZHY?marketplaceId=ATVPDKIKX0DER&musicTerritory=US&ref=dm_sh_CdyeuxIBNBDOjZ1hpB1PPJxT0
Danger Ahead https://www.amazon.com/music/player/albums/B0DB2DV354
Music In The Machine https://www.youtube.com/@musicinthemachine/videos
This is the simple merging workflow I use in creating my models. Just load the base checkpoint in the first node, load a Lora in the second and set a strength for it. The third is the node that saves the new resulting checkpoint. If you are merging the 8 Step Lora, the setting is 0.125 for that one.
Description
Simple workflow to merge Lora with checkpoints.
FAQ
Comments (12)
woha is really that dead simple to merge a lora into a checkpoint? wtf really. nice
The merge itself is simple. The hardest part is getting the strength right.
@Seeker70 thank you for this. Do you have any recommendations on how play with the strengths?
@Perrito It's hard to say because different types of Loras need different strengths. For instance the realism lora that was released by Black Forest Labs themselves, I mix in at the value of 1. Other loras like boob loras, if you mix at 1 it becomes nearly impossible to get an image that you don't get nipples on the outside of clothing, so any of those I use are usually more like 0.25 or 0.3. The hyper 8 step lora is only mixed in at 0.125. Unfortunately it is mostly trial and error until you find what works for what you want to accomplish.
and how do you turn back if do a mistake? let's say i have merged 3 Loras succesfully into my checkpoint, but i mess the strength in the 4th Lora. Is there any way to "unmerge" the 4th one?
@alexdimou22 No unmerging. The way to do it is to record what you are doing at each step and keep each one until you get it to the final and then delete the ones that are not the finished product. Or if you have space you can keep each one for more testing later.
Hi, I see in the screenshot it's a pruned fp8 Flux model that's merged with the LoRA. I remember reading somewhere that we should be using the original uncompressed fp32 version to do the merging, and then quantize the result. Is it so?
@jtabox It just depends on what you want. I almost always use the fp8 because the full model is massive and mostly very hard to use unless you have a massive amount of VRAM. Personally, I don't find the full model to be better or at least not enough better to be worth the trouble of trying to deal with it. My computer has 16GB of VRAM and using the full model at least triples or quadruples the time it takes to generate an image and the quality is so close to the same that it is never worth it to me. All but one of my Flux models is based on the fp8 version and they still seem to some of the best if not the best models available.
Hi there, i think the title is wrong, since this is not a lora merger, this is a checkpoint plus lora merger. thanks
Yes, I suppose you're right. Updated the name.
@Seeker70 by the way...do you know if its possible to merge 2 or more flux loras? this what i was looking for when i discover your workflow. To be honest i'm surprised by how easy is to merge a checkpoint and apparently how difficult maybe to merge a few loras, since there is close to no info or workflows on the net....thanks in advance, appreciate your answer.
@tupu I don't know if Loras can be merged. I have also never seen any workflow to do it. I have never really looked for one though either. Most of the Lora model I make are characters and I would not want to merge those, so it has never really come up.