The aim of this model is to help create nsfw "nude portraits", with the person standing and posing like they would at a photoshoot. It's also supposed to help with the physics and structure of anatomy. The dataset consisted of about 220 captioned videos, with varied faces so that there is minimal influence on your character loras. I tried to minimize the occurrence of jewelry, piercings, tattoos, long nails, and captioned accordingly, so you can mention any of these in the negative prompt to get rid of them.
The training was done at a 2:3 aspect ratio, so you can use these resolutions: 256x384, 384x576, 512x768, 640x960, 768x1152, 896x1344, 1024x1536
You can try this as your starting prompt: "A beautiful nude woman at a photo studio photoshoot, she is posing seductively and then turns around and looks back. She is completely naked. Professional studio lighting, high quality production"
You should be able to control various actions with prompting, like hand placement, body turning etc.
Use the lightning loras with this at 1 strength. https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan22-Lightning/Wan2.2-Lightning_I2V-A14B-4steps-lora_HIGH_fp16.safetensors https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan22-Lightning/Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16.safetensors
They will allow you to get a very good result at only 8 or 10 steps.
In Comfy Native, separate the steps like so:
Notice that for the first step only, you set the CFG to 3.5, then afterwards CFG =1
The High Noise models go into the first two samplers, Low Noise models into last sampler.
Euler beta57 works best.
Info relevant to the Wan 2.1 model:
The workflow is in the static image. Using a native comfyui workflow, along with the Causvid_v2 lora that Kijai made at 1 strength. This will allow you to then use 8-12 steps, at cfg 1. For schedulers/optimizers DPMPP 2M sgm_uniform seems to work very well, otherwise try Unipc/Simple. UPDATE: Instead of Causvid you can use the Lightv2x lora that Kijai made at 1.0 strength, which allows good generations in only 4-6 steps. LCM/beta works well with it