This is a personal project of mine, thought I'd share.

At the most basic level, this workflow allows you to press a button, and get an image, a video based on that image, and an upscaled version of the video.

It Utilizes an editable list of parameters that get fed into an llm, which spits out a krea2 prompt based on those parameters. For stability purposes, i specifically tailored the LLM system prompt and some of the parameter pickers to make stable blowjob images. This can all be changed by editing the system prompt and changing the available parameter choices,

This takes quite a bit of Horsepower to run. I'm sure there are some optimizations that can be made, but this version works pretty consistently for me. I have 24gb VRAM and 32gb RAM.

To lower VRAM and RAM Usage: You're probably gonna want to use lower precision Krea2 models and LTX models. You can use lower precision or parameter LLMS, but for the LTX prompt, you need to have an VL capable model in this setup. I would also completely omit the upscale portion of this workflow. MAYBE RTX super resolution.

I highly recommend a nsfw oriented Krea2 tune for this, as well as an uncensored LLM (the system prompts were specifically designed for Gemma), and some Loras for more variety. I like these models:

LUSTIFY! [NSFW checkpoint] - v10 (Krea 2) | Krea 2 Checkpoint | Civitai

LTX2.3 10Eros - v1.4 | LTX Video Checkpoint | Civitai

These give me good results like 90% of the time

randomized Lora strengths are supported; it can add a lot of variety to the starting image. Haven't decided if it's worth it for the LTX loras, but its implemented.

This workflow has 2 pause points. 1. when the image generates. 2. When the final Ltx2.3 finishes and the video generates. The first allows you to iterate pretty quickly to get an image you like. the second stops the workflow if you don't wanna upscale.

Upscaling is pretty time consuming and heavy (especially seedvr2 and optical realism). A good lightweight upscale is simply using RTX super resolution, interpolation, and maybe LensFX for some film grain. that generates a pretty good output. I actually don't even bother with interpolation if I want a more cinematic feel.

There is literally no prompt you have to input at all. if you want it to run without you even looking at it, disable the pause nodes and it'll do it all.

You could also lock in the prompt and test out different lora strengths on the same seed. there's lots of little things you can use this for if you tinker with it

I am probably going to rewrite this after I get some sleep, but I just want to shout out the two workflow creators i took the majority of the actual sampling pipeline from. @lonecatone23 and @MrXin. the sampling pipelines are basically just carved from their workflows, the main additions from me is just the automation and connection of the two.

Have fun! I hope to see the video feed flooded with blowjob videos

it can also be refactored for SFW content, I guess.

PLANNED ADDITIONS

In the next version, I plan to add some instructions in the workflow itself. This first version i cobbled together in one night, so it's a little messy.

Template Randomizer: I am going to add another layer of randomization that allows seamless switching between different actions. For example, after this change, you apply a list of action types besides oral, and the workflow will apply those randomly (say i apply: oral, missionary, dancing. Now each run will choose to generate one of those three actions, besides just a blowjob every time).

Low VRAM and RAM optimizations.

Wan2.2 and Anima support

Probably much more.