Qwen GGUF Autoprompting Naughty H3 Instruction

This is nothing new or novel. I'm publishing this because I plan to use automatic prompting in the workflows I post going forward. Simple Qwen custom nodes is what I use. So instead of cluttering up the model description with all the info required to get this running I'm posting it as a template, something you can load by itself to get the prompting module that I'm including in my workflows up and running by itself before trying it in the workflow. Which would be a waste of time if this isn't good to go. Because this was such a pain in the ass to set up, at least for me, I've included a patched .py and some instructions. Please read the instructions on the git page to install the right library.

So, just to be crystal clear, this workflow here is just a template. You can load an image and make prompts and whatnot, but the actual implementation, using my custom instructions, will be in my workflows, and I will point to this page for the specifics on the prompting section.

Using an abliterated instructor is required if you want Qwen to write filthy prompts for you. No way around it. This is what abliterated/heretics are actually for, by the way. Not encoding your naughty prompts, but crafting them.

Also included is a modified presets .json that has all of my custom H3 instructions in it, for various workflows, most unpublished at the moment. If you have a better setup for running an abliterated instructor, by all means use that. The only unique material here is my custom H3 instructions. There are a bunch of presets made for specific scenarios, such as my Infinitizer workflow and my Infinite Iteration workflow. I will update the latter soon to include the autoprompting.

Repo for Models: Abliterated Qwen3-VL-8B Repository

The Quants I Use:

• Instructor Model: Q8_0 GGUF Model File

• Projector File: mmproj Q8_0 File

If Qwen gguf nodes are new to you, you must download the patches I have included. To make the setup work you have to:

• Install the custom nodes (duh)

• Install the CORRECT llama cpp library. There are some notes in the folder, use your go-to LLM for help with this. It will not work if you have the wrong one.

• Overwrite the qwen3vl_run.py with the patch I've included. The default code still has a damn typo in it, even though the author said it was fixed. Clearly he can't spel.

• Overwrite the system_prompts.json with the one I've included. Mine has two changes- first, it sets the paths for the gguf pairs (instructor+tail), second, it adds several of my own special instructions for H3, the relevant one here being 'H3 Infinite Loop.'

• Edit this same .json such that it points to your gguf paths, not mine. My paths will be slightly less than useful for you, considering that they are on my supercomputer, not yours. The models are loaded as pairs, and you can also include any config settings that are in the full config node. If you do not add your paths, you must use the config node and manually enter your paths in the widgets.

Depending on your level of computerization, getting this set up and running with your abliterated instructor and correct library can be anywhere from a breeze to a nightmare. I'll admit that it took me a while to figure it out. A long while. Definitely consult an LLM if you do not understand what is required here.

There are some notes in the workflow on general usage as well as the changes in the recent updates to the Simple Qwen nodes (added a LOT of functionality). There will also be tons of notes in the upcoming workflows about the specifics of each implementation and how the instructions work. They are very different for each use-case.