using ComfyUI and qwen_image, need help please!

[Drac Morbis]

Okay, so I have a local setup now (thanks again to @TailsWin ). I installed ComfyUI and qwen_image. I've done about 30 renders, using an "image reference" node. The blood looks good, and what not. But what sort of prompt do I use for like a girl being stabbed, or a throat slit in process? Stuff like that? I also used this reference image:



And tried stuff like "full body, naked, with small breasts and a shaved vagina". And it fucking did this:



I want her as Deita and Drac's next victim. What am I doing wrong??

@Otto Von Herunterhängen @xj900uk @Lynortis @MCDesigns @Nyghtfall3D @TailsWin @jh snuff

 

[TailsWin]

Download some existing image made with Qwen-Image and drag it into the ComfyUI window. Use that as a template for how nodes are set up. It looks like your VAE might be out of wack (unless that is the default rendering, but I kinda doubt it) so maybe your whole setup is wrong in some way.

To see if nudity works at all, use a simple starter prompt: 'nude female body; lying on ground; outdoors; warm lighting; photorealistic' and see what you get. Or try something medical to see if you get nudity. If not, either the text encoder is censored, or the diffusion model (image generator) isn't trained for that, so you might need to download a LORA of some nsfw sort to add that capability. I've not used Qwen-Image so I'm not sure which it is.

 

[TailsWin]

Wait, this is your prompt for the second image?

full body shot, completely nude, and she looks happy to please you with her naked flesh

Ok I don't have comfyui on this computer so I can't load to see what your setup is. I guess you were trying to use editing mode to insert that reference. I'm not sure how that works, but first try to create a new nudie from scratch and see if that works.

Second, this isn't chatGPT that reinterprets your prompt. You need to be descriptive and precise, you're talking to the raw model that's more like an algorithm that you need to program with words.

Here's one of my older simplest prompts:

front view of girl, age 21; coarse skin texture; her skin and hair are dirty and dusty; messy wavy auburn hair; fully nude; 34C breasts; down on all fours; face to ground, hands flat, looking forward, intense stare; clenched vampire teeth; to her left stands a shadowy ninja fully in clean black, black gloves, holding a dirty old chainsaw just behind her head; rusty abandoned warehouse; diffuse warm light; strong film grain

That was for Z-Image Turbo I think, but that also uses the Qwen 3 text encoder so I guess Qwen-Image will probably have similar prompt structure. Try that and then go further to editing.

 

[Otto Von Herunterhängen]

AIRE uses Qwen on Comfy.ui so it can do nudity, I've done it. It does sometimes add clothes, and it really hates doing a male penis, I usually have to do several takes with those. Since I only use it via AIRE I'm afraid I can't be much help. AIRE does add six components to the mix so my set up will be different from yours.

 

[Brian The Snail]

shouldn't the underwear be purple? (Sorry, old DAZ joke)

 

[Otto Von Herunterhängen]

Ah, you got the clothing from your reference image, that's the interpretation of AI of that black shoulder strap. It also got the necklace from there. You also have to have guidance set high enough for it to use your prompt, in AIRE it would be 1.5 and up. 1.0 will not do prompts dependably.