You are a god.

Will it not work with the lightning lora?

oooooh. you're the Mystixx guy. I was about to ask whats good about this Lora but now I dont need to. Thanks.

thank you for your service, will test now!

This can be used to edit an existing source image correct? I'm actually using your Mystixxx LoRa and the SNOFS LoRa with Qwen Edit with decent success. Does this have better prompt adherence or is there a specific prompt style that should be used? Your example image prompts look like a standard Qwen Image prompt and doesn't seem like they are for Edit. Thanks again!

Does a nice job of creating NSFW scenes. Weewees aren't deformed nightmare creatures, vajoogos are realistic. But unfortunately it does a meh job of retaining the original subject's likeness which is pretty much the point of 2509 IMO. Still useful for a lot of things tho.

I am unable to maintain a characters likeness using this lora. Face, body proportions way off from the original character image ref. Anyone else having better luck?

If it cant maintain a likeness then it is not useful as a image editing tool. Also I am getting a lot of extra arms, missing fingers. Is it a me problem?

Thank you for your time and work.

can confirm it works with nunchaku lightning version. Quality is degraded but still acceptable!

... and along comes our savior Mr. alcaitiff and touches a nsfw lacking model with his healing hands... You're the one! Thank you so much! Every generation improves with this one. First of all the skin texture is way better than before. It's strange, but it also improves the overall sharpness and reduces artifacts in my used cases. Can't wait until you released a lora for Z-Image-Turbo with your wonderful dataset. plzplzplz ♥

I don't understand why none of the example prompts are qwen-edit style prompts: they are all t2i style prompts. I thought this was a qwen-edit lora? How should it prompt for i2i editing?

I've tested 18 latents with Nunchaku 8-step, abliterated Caption-it in the base reference-free CLIP Encode node and it actually produces less nudity than without the LoRA. It also doesn't nudify characters in reference images. CFG 1 may be the reason.

I later got the non turbo version at the risk of getting charged of getting bandwidth overages so I can use higher CFG. 20 steps euler simple. It took maybe 3 minutes to generate the images reference-free and though they were all nude it suffers from the same problem as NSFW LoRAs generally have: The characters appear much heavier and older than they would on the base models. The base models produce very attractive 20-25 year old usually Asian women, these women look like 30-35. It looks like those nude maternity shoots without the belly baggage.

Wow, CFG 4 is slow. I got about halfway with 3 references and it's clearly not going to nudify them. This is probably going to take 15 minutes. Screw this.

It's still way better to inpaint on an existing nude image with denoise ~0.5 to gradually morph a character onto the existing nude images. On the base models. Takes a few tries and like 5 minutes overall. Works with Kontext, Edit/2509 and Flux UMO (the best one) and you don't even need nude reference images, as long as you use the abliterated text encoders. If you do have nude reference images you can get away with a higher denoise but above 0.7 they will all start to put clothes on them.

(If you want to make nude reference images, use inpainting with Qwen snofs or any number of other NSFW LoRAs with a tight DepthAnything v2 or v3 map and a tight mask. Even Stable Diffusion can do that!)

Or if you need high quality original nude characters that is trivial with Flux Krea or Kreamania and Pussy Diffusion. PD and snofs are the only ones I know of that don't make the characters gain weight.

Hey could you maybe upload a workflow and example edit or smth along the lines?

I cant seem to get it to work for qwen edit 2509. Anything slightly nsfw just results in no changes to the image, aside from some color deviation from the vae

so after testing with nunchaku r128 here are the best settings I've found so far:

- edit 2509 lightning 8 step, 1.0 strength

- 8-12 steps

- CFG = 3.0

- euler/beta

- unchained = 1.0 strength

- 1104x1472 and other standard qwen image resolutions

If you go no lightning and CFG 4 + 25 steps as default the results are improved, but I found this to be a better balance for speed. Also everyone says you need to keep resolution <1MP but I think the higher resolutions look great! (at least up to 3MP then you get dupes)

Amazing LORA, but it is supposed to alter the face quite a bit from the original reference image or am I using it wrong?

"It can make the image closer to the intent of the prompt, but its consistency retention capability is poor, not as good as EVA-V3. It is recommended to cover the face before training; the effect should be better. Facial expressions are too easy to affect face consistency."

Qwedit 2511 Unchained, when? :D

What are the odds of you making more sliders for edits, like breast sliders and what not

This lora looks amazing. Unfortunately it's a pain to use it for editing, because it tends to change pretty much everything in the image - the character, the scene, the angles. It's a shame, because I don't think there is any model or lora that can make NSFW edits without changing the person's face or body. I would love to get an improved version of this with improved consistency.

I like the idea of this lora, but it's overtrained.

I love this so much, excellent job!

please make a updated version for 2511 that uses Steps <=8 & CFG 1, CFG=4 is too much

are these useable with Qwen edit workflow too?

Thanks a lot for this, it works quite well even with Edit 2511

When i tried it with Qwen Edit 2511 it didnt work. Is there something I am missing ?

This is probably my favorite lora right now