In the search for Character consistency I created a workflow to create dataset images and maximize the consistency of my character LoRAs. This workflow employs PuLID and eight reference images. I also use my Character LoRAs.
This is NOT a low VRAM workflow.
I am using a Impact WildcardProcessor node for text prompt resolution.
I utilize MultiGPU Support to utilize my 2 NVidia GPUs.
I, as always, implemented LoraManager support, because: If your aren't using LoraManager, what are you even doing?
Instructions for use are in the workflow, so I won't detail them here. All the nodes are available through ComfyUI Manager. Links for the base model FLUX.2-klein-9b-kv-fp8 and the CLIP and VAE models are in the workflow as well.
For best results you will most likely need to experiment a bit with the reference images.
Happy Generating!
Description
FAQ
Comments (4)
Before I discovered this workflow, my life was simple.
The more I use it, the less I understand what it does.
It is for character consistency using reference images. What are you having problems with?
@DarthRidonkulous hi, thanks for the quick answer,
the four images on the left (in the "PuLID Face Reference" group), do not seem to influence anything, the only reference used seems to be the first of the 8 reference images on the right (in the "Flux 2 Eight Image Ref" group). i can put 4 portraits of nelson mandela on the right, and one michele obama in the 'Image Reference 1 - Face" node and the result will be a Michele Obama.
Although i managed to get slightly different face orientation, change of gaze etc. , it is a quite brutal face transfer, almost a copy paste of the image put in "Image Reference 1 - Face" node, if the prompt is kept simple. but that's maybe the way it's supposed to be.
The "🔀 CR Image Input Switch (4 way)" node setting (1,2,3 or 4) doesn't have any influence
In the "Apply PuLID ✦ Flux.2" node there is this mystery parameter of "image_index" that doesn't seem to change anything when set on anything between 0 and 9
I changed the loras in the "Lora Loader (LoraManager)" to 4 loras that work fine with the model in other workflows and only one of them is usable, the other 3 give a "file not found" error although the spelling is right, and the models are in the right place.
I feel a bit stupid to not understand exactly what the workflow does with the references, I see there's a string you use to concat the different prompt parts, it' in a "CLIP Text Encode (Positive Prompt)" above the wildcard node. it says "change the woman's face in reference_image1 with the woman's face in reference_image1". Is that the way the face transfer is done? I can change it or even erase it it doesn't change anything (Poor Michele, she's suffering with my test)
I did not test anything past the "Sampler" group.
Don't worry, i'll keep investigating, and thanks because it made me discover powerful things like the wildcards
@crystle00928 I will try to answer these questions tonight after work.