I'm in the same boat, it doesn't want to run in Wan2GP. I think the issue is that this is a lokr, not a lora.

Yup it doesn't work in WanGP sadly.

there are numbers you can call in an emergency.

@crombobular which one 666?

Looking at the ComfyUI logs, the reason this doesn't work with Nunchaku is because it doesn't currently accept Lycoris files.

Would you mind sharing the LoRA weight you used for 2511?

@Saxlive Oh, I just used 1.0, it's been working well for me. Solo-subjects work really amazingly, I have been having some issues with using it on multiple people, it doesn't always get the positions right, but still way better than base

but image edit model not released yet for 2512

@Ashen3 Somewhat? It works better than 2512 models! Still the king!

I really recommend this guys: https://civitai.com/models/1981116/dasiwa-wan-22-i2v-14b-lightspeed-v7-v1-or-fp8-safetensors I've tried some loras in the past but usually the results are subpar - getting the right strength to work without it going to mush is...interesting lol

i ve used this lora with qwen image edit and the results were satisfactory but not excellent, who knows if it works for you. But u may give it a try

Yes it works to an extent

Can you give me some tips? Are you using it in image video? Or text video. I tried to image the video and I couldn't get it to do anything. Is your prompts what makes it work? I'm just more used to z image turbo. So any advice would be great ?

Qwen Rapid AIO already includes this lora (and others). By using it as well, you will further increase its influence.

I experience this the most when using high aspect ratios; i believe the base was trained on 3:4. What is your output size set to?

google bro

When a lora asks for specific words or key words. It means these are the words that will activate the lora based on who trained it. If you don't use these words you may see some effect of the lora but it will be a lot less. If you use no lora and rely on your prompt then the model eg qwen zimage sdxl has to know what those words mean. Some models are sanitized of NSFW terms. Using certain loras adds these things back so you can generate your concept, or a better version of it.

@ai_fanatik_dude That wasn't very helpful. If you google the word bro, pictures of jerks show up. If you can't help someone when they ask a question other than telling them the obvious, don't bother.

@Ashen3 omfg it's xmas again. thank you

@Ashen3 will there be one for version 4b?

Nice, good job!

Didn't think to combine these. Which rapid aio version are you using and at what strength do you use snofs?

@milkfarm I more less use it for the positions. qwen rapid AIO nsfw v23. I just inpaint any bad gens if need be with something like biglust. I just left the strengh at 1 and edit men into images n junk. Going from a nuetral image to a nude woman on top of a man in 4 steps with decent results most times. Euler A/Beta. I am sure someone else can get better results than me though as I am not the best at this stuff.

Qwen AIO already includes this lora (and others). By using it as well, you will further increase its influence. The list of loras used by AIO is available in the metadata.

@oqe599 wdym by in metadata? are there trigger words that I am missing? some things will trigger for me but some won't. I for instance can't get a doggystyle position triggered without the lora, theres always a dog somewhere in the image. and stuff like "sex from behind" is just kind of hit or miss. I love AIO and if I can use it more effectively I would love to.

I didn't say that adding snofs to AIO was useless. Snofs and the others are not normally designed for qwen edit, so it still works more or less well despite all the weight adjustments made to the lora in the various versions of AIO. I was just pointing out that AIO's capabilities were largely due to a collection of Lora, mainly snofs. At the beginning of the AIO version 23 file, we find the following data:

"_meta": {"title": "TextEncodeQwenImageEditPlus"}}, "30": {"inputs": {"text": "# overall decent, if a bit wonky composition 4/5\n<lora:ImageEdit\\gnass_2512_qwen.safetensors:0.9:0>\n\n# honestly perfect, 5/5\n<lora:ImageEdit\\Qwen_Snofs_1_3.safetensors:1:0>\n\n# plastic, kinda cartoony, strange penis stuff, but followed prompt nicely 3.5/5\n<lora:ImageEdit\\yky_000032000.safetensors:0.66:0>\n\n# very clear, poor prompt following though, forcing sex, 3.5/5\n<lora:ImageEdit\\Meta4.safetensors:0.66:0>\n\n# very strange, but not absolutely terrible, poor penis, strange composition 2.5/5\n<lora:ImageEdit\\Qwen Sex-tacular V1.safetensors:0.33:0>\n\n# ehhhhhh not great, not terrible 3/5\n<lora:ImageEdit\\qwen_MCNL_v1.0.safetensors:0.5:0>\n\n# very odd issues, pretty bad, but grid issue seems gone? 1.5/5\n<lora:ImageEdit\\Qwen4Play_v2.safetensors:0:0>\n\n# finger horror, oddities, pretty bad (penis and vagina wasnt half bad tho) 1.5/5\n<lora:ImageEdit\\YARNv0.2.safetensors:0:0>\n\n# very good, 4/5\n<lora:ImageEdit\\Unchained-XXX.safetensors:1:0>", "normalize_weight": 1.1, "model": ["41", 0], "clip": ["41", 1]}, "class_type": "LoraTagLoader", "_meta": {"title": "NSFW Blended LORAs"}}

Snofs is described as “honestly perfect, 5/5"

Yes. It's not compatible with Nunchaku. Would appreciate it if someone converts it to a LoRA.

@Ashen3 Subpar to you is probably still better than any of the other things out there. Still have it lying around?

Also, I've actually gotten some skills myself though, and might release some missionary position stuff soon. If it only does one thing it probably won't have any problems on Nunchaku. Is there anything special about your dataset that causes it not to depart from Qwen's style?