Can anyone share if this works, and if does, how to set it up in ConfyUI along with flux2-klein-9b. I could not make it work with standard CLIP loader, makes no difference what so ever.
The text encoder and the LoRA solve two different problems.
Uncensored text encoder = removes the blocked communication path LoRA = adds the missing visual knowledge to the DiT
Without the uncensored text encoder, the prompt conditioning may still be distorted or blocked before reaching the DiT. Without the LoRA, the DiT cannot draw concepts it was never trained on. They are complementary, not redundant.
What you said is exactly my thinking before but since I've never used an uncensored text encoder but all the nsfw loras still work, never thought an uncensored text encoder would do anything.
any idea of setting up workflow on ComfyUI to specifically generate the realistic and NSFW content. Mine outputs look like AI-ish style and incomplete ( like not really following promts )of picture with both Normal and NSFW prompts. Here is the current workflow btw. for both image generation and editing Thank you in advance.
you just use the comfyui templates for the model. They do the kinda complicated conditioning path for you. And if you need to add a lora, which you will need) you put it a load lora node or more between diffusion model loader and whatever comes next. But as said above that text encoder will enable very little if anything by itself.
Talking of actual uncensored. try Chroma1hd, it's a classic. Which is Flux1 based. But a very promising successor is the fairly new Kroma0.1 lora for Krea 2. (same basic thing from the same guy who made Chroma). It completely uncensors Krea2 (without needing a uncensored clip) - except for prompt enhancing.
