Hotscreen

Add very hot effects to your screen. · By PerfectFox265

Add to collection

game Community Devlog

Hotscreen community

Mod to generate the description of censored content with an LLM Sticky

A topic by PerfectFox265 created 76 days ago Views: 4,003 Replies: 11

Viewing posts 1 to 8

PerfectFox265Developer76 days ago (9 edits) (+2)

I've create a custom box scene that uses a LLM loaded in LM Studio to generate a description of the hidden content.

Video showcase using images by MIAOKA on civitai :

How to use it

1) Download the file "llm_description.box.tscn" here or here and put it inside your CUSTOM_DATA folder.

2) Start LM Studio. (You can install it from here : https://lmstudio.ai)

Install and load a vision model. (lmf2-vl is fast. qwen2.5-vl-nsfw can allow spicy descriptions.)

Start the server. (*)

3) Start Hotscreen.

Add a filter -> Mods -> Box scene

Edit filter -> Select custom scene -> Open the file "llm_description.box.tscn"

The "mod" is very simple. It's only a text file loaded like a box. It doesn't reference any Hotscreen scripts/variables, making it easy to modify in Godot (or by an AI chatbot). There are variables at the top to change the prompt, color, etc.

I think this mod illustrates the power of Hotscreen's modding capabilities well. I hope you will find it cool too. 😊

(*) Here are my settings in LM Studio which should be the default ones :

PerfectFox265Developer unlisted this topic 76 days ago

PerfectFox265Developer relisted this topic 76 days ago

HotpointUser76 days ago (1 edit) (+1)

That is very cool. I will try it. how can we modify it so it describes it in certain ways. Such as mentioning breast size?

PerfectFox265Developer76 days ago

You can modify the prompt at the top of the file.

HotpointUser76 days ago (3 edits)

That works well and a very nice feature. Anyway to increase the font size? I tried changing var custom_font_size: int = 30 but it didn't increase the font size.

Does it see the whole image and only describe what's in the box or does it only see what's in the box?

Would it help to add a prompt or system setting in the LM Studio program to assist w/ the description or response it writes?

PerfectFox265Developer75 days ago

You have to restart Hotscreen after you modify a .tscn file to refresh the changes. So your font size should be bigger next time.

The image send to LM Studio is obtain in the function get_screen_image_under_box() (line 84). You can easily modify it.

The system prompt ("You are a helpful assistant.") and settings (temperature : 0.8, max_tokens : 500) is sent with your prompt (line 117). I think this overrides the settings of LM Studio.

HotpointUser75 days ago (1 edit)

I changed the Var Prompt to "Response should consist of only a description of breasts visible. Describe the shape and size of the breasts in as much detail as possible. Maintain anatomical correctness and accuracy. Try to accurately state the bra size of the breasts. Describe the color and size of the nipples. Don't use the words [style, image, close-up]." I also created a system prompt in LM Studio to say the exact same thing. It still gives varied descriptions of things in the image besides the breasts but it's a cool feature and kind of works.
Doesn't work so much on videos because of the flashing censor detecting undetecting and not having time to load the description.

I'm going to experiment with different LLM engines. (EDIT: I don't know how to specifically search for models that can do image detection, little yellow eyeball)

PerfectFox265Developer74 days ago

The vision model usually have VL in their name. Qwen3-VL should be available in LM Studio this week.

It would be also possible to finetune a model with this : https://github.com/unslothai/unsloth using their pre-made google colab. We just have to create a dataset of example of images with their "censored description".

kuiluic75 days ago(+1)

good idea😊

psi-conaut75 days ago(+1)

The software crashes when I turn on the mod.

mementum73 days ago

I tried both of the vl models you suggested, but with the nsfw one I still get censored text results. Tried diffrent prompts but couldn't do it

Generic Gio7 days ago

This has been very interesting to observe. I have used this decently. My only complaints are that the Qwen 3 model seems to be rather restricted in it's usage. It gives me very bland and disappointing responses. You ask it to call you one name, and it goes on a superman rant about peace and justice (harmlessness etc). Do you have any suggestions? It is a very good feature integration, I just need a better model it seems. My goal stems not from the description, but from getting implicational second person statements towards the reader (me).

PerfectFox265Developer5 days ago

For now, the best would be to finetune an uncensored Qwen3-vl model. There are pre-made google collab to easily do this, but it is required to create a dataset of images with their expected spicy description.

itch.io

Hotscreen

Mod to generate the description of censored content with an LLM Sticky

How to use it