without comfy? why? even with comfy its not easy. the loras should have identical technical specs like rank, alpha and dtype or it wont work well. But without comfy? no idea...

thanks nphSi, I dont use comfyui to generate images, my question is about your Loras , I suppose they have same technical specs.

Do you want to merge two loras to have one .safetensors file at the end or do you just want to use two Loras together when generating an image?

same, or better way, I have tried to generate image with 2 your Loras, with a prompt character 1 and character 2 , but seems two twins, I tried with kohya script merging the 2 Loras, but the result its not good.

i see, this will not work very well because both loras share the same base weights and mix together. All you can do is to be really exact at prompting describing each character individually. describe as much as possible and use only the vrtl trigger to name them, no real names. see how i did it with bud and terence. Let the sum of the lora weights not go over 1.4. Or use inpainting.

can you write an example ? I want to generate image with Claudia Schiffer and Adriana Lima together, maybe something like " the photo describe (vrtlClaudiaSchiffer) and (vrtlAdrianaLima) together in a party .
or something better ? Are you talking about one merged Lora or two Loras ?

single loras or merges does not matter. here is the bud and terence prompt:

https://huggingface.co/nphSi/Z-Image-Lora/discussions/17

about Bud Spencer and Terence Hill example, this is one lora training 2 character , ZI Bud and Terence (vrtlBudSpencer,vrtlTerenceHill).safetensors , I wish to use 2 different Loras with one unique character.

can you post one image of your Loras with 2 character using 2 Loras , and prompt example ?

It behaves the same, important is to tell the models that the tokens are individual and not to be mixed.
I am currently on a training run, ill try to put a sample online when finished.

thanks @nphSii I will wait your example.

As said these models are not really made to be used this way. Inpainting or Edit-Models are the better solution. ZI-Edit is still to come.

• This are 2 Loras loaded at 0.5 each:

• This is a combo lora i quickly trained from both datasets loaded at 0.9. It works a bit better. Download link below.

Prompt for both:

• a scene with two individual woman sitting on a bench in a flower garden with spring at a summer evening at sun set.
  the woman look different and are not related to each other.
  the left woman is Adriana Lima (vrtlAdrianaLima), a tanned beauty with long black hair wearing a light summer dress and red sandals.
  the right woman is Claudia Schiffer (vrtlClaudiaSchiffer), an older light skinned blonde beauty with short wavy hair in her 50s, she is wearing a blue jeans and a brown blouse and white sneakers.
  they both are looking at the camera. vrtlAdrianaLima is making a smartphone selfie.

• Combo Lora: https://huggingface.co/nphSi/Z-Image-Lora/tree/main/temp

I am not really experienced with this, maybe someone with better skills can add some tips and tricks...

hi, thanks for the example, I tried to merge your 2 Loras with Koyha sd script , this is the result, how you can see its not good, your Combo lora its better, anyway maybe someone can explain better some tips to have better results , here my image with your prompt .

Well the next logical step would be inpainting and replacing the faces. Multiple characters/loras is really limited in standard models out of the box.

maybe its impossible, I dont know, anyone has some tips or tricks ?