NSFW Loras used
Hi There - first of all congrats on a very good work with the Mega AIO NSFW and SFW checkpoints and Comfyui workflow. Thumbs Up.
Out of curiosity - what NSFW Loras were used on Mega v11? Just to get "deeper" understanding of what prompts to use for best results.
Thanks in advance - Keep up the GOOD Work.
Just find out what NSFW Loras were used from another post.
This is the list of LORAs used in v11 NSFW merge. I usually add things to improve a certain area and then they get mixed in at much lower strengths than shown (it is more a relative strength number). If anyone is aware of a particular LORA that messes with faces, let me know:
lora:WAN\NSFW\Wan2.2-Low-MysticXXX-v2.safetensors:0.5:0
lora:WAN\NSFW\dreamlay-t2v-wan-low-v2.safetensors:0.5:0
lora:WAN\NSFW\DR34ML4Y_AIO_PREVIEW_V1.safetensors:0.5:0
cum/cumshots
<lora:WAN\NSFW\56Low noise-Cumshot Aesthetics.safetensors:0.6:0>
<lora:WAN\NSFW\Wan2.2 v2 - T2V - Body Shots - LOW 14B.safetensors:0.4:0>
lora:WAN\NSFW\wan-thiccum-v3.safetensors:0.3:0
lora:WAN\NSFW\wan22-f4c3spl4sh-154epoc-low-k3nk.safetensors:0.4:0
lora:WAN\NSFW\Wan22_Cum_low_noise_1.V1.safetensors:0.3:0
lora:WAN\NSFW\wan2.2_lownoise_cumshot_v1.0.safetensors:0.3:0
sexual motions
lora:WAN\NSFW\doggyPOV_v1_1.safetensors:0.25:0
<lora:WAN\NSFW\Wan2.2 - T2V - POV Titty Fuck - LOW 14B.safetensors:0.4:0>
lora:WAN\NSFW\Sensual_fingering_v1_low_noise.safetensors:0.25:0
lora:WAN\NSFW\jfj-deepthroat-W22-T2V-LN-v1.safetensors:0.5:0
lora:WAN\NSFW\wan_dr34mj0b_t2v_HD.safetensors:0.25:0
<lora:WAN\NSFW\Wan2.2 v2 - T2V - Cowgirl - LOW 14B.safetensors:0.5:0>
lora:WAN\NSFW\wan_cowgirl_v1.3.safetensors:0.25:0
lora:WAN\NSFW\wan2.2_t2v_lownoise_pov_missionary_v1.0.safetensors:0.5:0
lora:WAN\NSFW\wan22-pov-handjob-low.safetensors:0.25:0
lora:WAN\NSFW\Handjobs_low_noise.safetensors:0.25:0
What is awesome is that Phr00t has the workflow in the metadata area of the file details. So if you navigate to https://huggingface.co/Phr00t/WAN2.2-14B-Rapid-AllInOne/blob/main/Mega-v11/wan2.2-rapid-mega-aio-nsfw-v11.safetensors you'll see the json in the metadata area. Copy that out to ComfyUI and you can see how the whole thing is made including the list above. Or if there are tweaks of that list, you'll see it.
This is where I learned how to use "extra-state-dict" to force a WAN model to become VACE compliant. Although I can't get it to apply to the model that I really want. I cook up the model + VACE weights, but it does not respect the input images for extended videos. Jib Mix WAN is an amazing model that is meant for WAN T2I only, but if you combine it with MystiXXX High + Low LoRAs (2x KSampler Advance) it becomes an amazing T2V model. When I add the VACE dict to it, it doesn't respect any of the VACE image inputs to make it I2V or extended video. (It kept being T2V.) I need to figure this out. I tried some model merges, but just got horror.
Another thing I wanted to do was cook Mega with High and Low models (separately). Without the High, action will always be slow and you also miss out on so many subtle details. At the same time, I want to remove Dr34ml4y LoRAs as I believe they are the reason why all extended videos become cartoony mush. If you look at the Dr34ml4y examples on civitai (especially preview), that's the look extended videos turn out to be. I also want to experiment with the different LoRAs. Phr00t likes to use I2V with FFLF mode so he doesn't see the T2V problems. I have to gather the correct evidence to show possible improvements.
Dreamlay covers many different concepts, so removing it means filling in the gaps with other LORAs at higher strengths (perhaps multiple). MysticXXX seems like a similar "all in one", so I do add it too. However, I think pure Wan 2.2 LORAs move some of their motion into the "high" noise model that I don't use, hence needing to use a pinch of Wan 2.1 models to compensate. It is a balance and if you find a better mix, let me know!