A workflow that lets you use 8 reference images (seems to be the max), it is basically the standard demo workflow, with the added images, and with permanent random seed.
Prompt Guide
The prompt in the workflow lets you combine objects and concepts but the 6 image version is almost like an instant Lora training, let me give you the prompt:
Identify the same character that appears consistently across image 1 and image 2 and image 3 and image 4 and image 5 and image 6 and image 7 and image 8.
Treat image 1 and image 2 and image 3 and image 4 and image 5 and image 6 and image 7 and image 8 as different views of the same character, not separate subjects.
Infer the character’s overall form, proportions, surface details, and material qualities based on what is consistent across those views.
The character is named Bob.
place Bob in a forest holding a cup of teaThe last 2 lines are interesting, you can nickname the characters in the different images and Flux.2 remember. The last line is where you use your normal prompting skills.
note that Image 1 will determine the aspect ratio
Transfer Cloths
Treat image 1 as the original base image.
Treat image 2 and image 3 and image 4 and image 5 and image 6 and image 7 and image 8 as different views of the same costume, not separate objects.
Extract only the shared costume inferred from those references, preserving its geometry, seggmentation, surface detail, and material qualities as they appear consistently across the images.
Apply the costume form onto the character in image 1, aligning it to the target body’s proportions and contours.
Adjust scale, curvature, and orientation so the form conforms naturally to the body, wrapping, bending, or separating where appropriate.
Keep everything else from image 1 unchanged: identity, anatomy, pose, surface material, camera angle, lighting, shadows, and background.
Do not transfer anatomy, pose, or rendering style from image 2 or image 3 or image 4 or image 5 or image 6 or image 7 or image 8 Use those images only as visual references for the shared costume design.
The word costume pulls the weight since if you say clothing Flux thinks textile
In the 6 and 8 image version you can set the output resolution depending on your amount of vRAM. Just like classic SDXL high resolution might mess up the image, but i have had success wit 2 megapixel
Description
8 ref images + Lora support. legend has it this is the ultimate workflow!
be sure to actually add a lora
FAQ
Comments (13)
Thanks so much for this! I've tried your first prompt in Flux Klein, it's just like magic! Before that I was only uploading different pictures of the same character and I was prompting like "Make the character do this etc.". But now with your prompt, everything is clearer, more logical, and much more efficient!
Do you have a similar prompt for character swapping? I've tried several ones, sometimes it does the job, and sometimes not, it gets easily confused too
I'm glad you liked it. I'm amazed that flux can take so much logic into the statements. but yes sometimes it does get confused, you can add the statement that it should "produce a single image" it helps a bit.
What exactly do you mean by Swapping, as in swapping cloths or swapping positions or gender?
@Torque (Sorry for the late reply). By swapping, I mean replacing a character with another one in a picture, so that the new character has the same pose as the replaced one. To be frank this is something that seems completely inconsistent currently with Klein. It frequently ends up mixing the picture, or just replacing the head, or mixing the clothes.
Ive had success with that kind of swapping, Ive had success posing a ugly old 3d character and replacing it with a photo of a person. yeah it is a bit inconstant, but keep the near misses and use those iteratively. it is not controlnet but i would say pretty close
CLIPLoader
'nvfp4'
PLZ make for 50 series~!!!!!!
This is a bit beyond the workflow itself but here is what i did to get nvfp4 support.
go to:
https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-nvfp4
agree to the terms, you need to log in
Download this file to ComfyUI\models\diffusion_models:
https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-nvfp4/blob/main/flux-2-klein-9b-nvfp4.safetensors
press Ctrl+R to refresh, then select the file in unet_name.
make sure comfyui is up to date
I haven't been able to get Qwen-3 NVFP4 text encoder to work, if you can figure it out please post it
cheers
tbh, it looks to me like all it did was combine image 1 and 2, then ignored the rest. it might be helpful to upload some outputs that show that an image can actually be generated in a variety of poses, or angles at the very least.
i was trying to figure out a similar setup myself and was hoping your workflow had the answer to my issue, but after downloading this workflow it looks like you set yours up the same way i had mine setup. the issue i'm having is when i load image refs, it tends to favor some more than the others. it's strange too because lets say image 4 is very prominent across multiple generations, and then i switch it to be image 1, or 5, or whatever, it will still continue to be the most prominent image.
as for pointing out "do this to image 1 and do that to image 2", i really dont see/understand how it could possibly know what image that actually is. i cant imagine it knows what you labeled the node as. it's gotta have something to do with the order it's linked in, right? but even then, would it really recognize that using the phrase "image 1" or "image 2"? i think it makes more sense that it can analyze whats actually in the image. so when prompting, you should call those things out specifically. like with a 3 image example, "put the tshirt on the character and then place the character in the desert scene". if you're saying image 1, 2, and 3... it's just gonna be kinda random maybe, i dunno. ive been trying to wrap my head around this for a few hours now. lol sorry
My experience is that it does in fact use all 8 images, and that 8 images is the limit, officially it is only 4 images, but yeah it does tend to pick one out if you don't specify a pose. Try asking it to "make Bob lie on a bed".
I've had quite a lot of success with taking characters from their original in game renders and turn them very photo realistic. flux edit seems to be excellent at processing images for later lora training.
I'm pretty sure it is the loading order that defines the different images and makes it know which image is which, image 1 is special though since the workflow makes it define the aspect ratio.
It seems that once we tell flux it is the same character from different angles it is even ok with having multiple views in a single image.
But yeah, I'm on a discovery path just like you and would love to hear your findings.
is there 2 image +lora version ?
no, but you should be able to make it yourself quite easily, click the edit button (the one with👇) and use use Ctrl+B to disable the nodes image scale and reference conditioner.. (the ones repeated in the bottom left) do this for all but the two last input of the images. then just load images into image 7 and 8 and call them image 1 and image 2 in the prompt
Nice workflow - thanks for your work ! I have just one request - can you please modify the workflow if it possible so you can disable the image slots that you wont use more easy with buttons ? Thank you !
I'll see what i can do..
hey could you make a version thats just the default klein workflow, but with a lora?