>>10123
WAN 2.2 does better than 2.1 with holding a style, but still imperfect. I noticed lips too, "pale lips" in prompt may help, but still a dice roll I think.
Where to begin? Long ass post here.
I used SmoothMixWAN GGUF:
https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF/tree/main Using Q8 right now, but used Q4 K_M before & worked fine with less resources. Thing with WAN 2.2 is it uses High & Low noise models & LoRAs. When generating you have half your steps on High noise & other half on Low. WAN 2.2 follows prompts better than Framepack's Hunyuan that's for sure. I can even gen text on a banner so long as it's not a lot of words.
WAN 2.1 VAE:
https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/vae/wan_2.1_vae.safetensors
Text Encoder (Q4 K_M):
https://huggingface.co/city96/umt5-xxl-encoder-gguf/tree/main
Clip Vision:
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/clip_vision
I started with this workflow:
https://civitai.com/models/1829052 which gens multiple sections with different actions per section, but I've modified the workflow a bit for adding extra LoRA per section, setting resolution based on a maximum width/height value, & setting duration in seconds instead of frames. My modified workflow:
https://files.catbox.moe/g11aub.json String Selector node uses one line per section while Precondition & Postcondition apply to the prompt on EVERY section. I didn't really
create any workflows & only merged things from others to expedite usage.
SmoothMixWAN has Lightning LoRA speed-up baked in so it can gen at 4-6 steps natively. SmoothMix also has some nsfw LoRAs baked in so some things like breast jiggle would happen without prompting & in my tests so far it knows deepthroat & "shaved innie pussy" to
some extent, yet still varies. Can use "SmoothMixAnime" in prompt for anime style (their anime style examples have gradient shading though & this tag can be detrimental to more 2D shaded images in some cases) Their other trained tag is "SmoothMixRealism" for realistic junk of course. uni pc + simple sampling gives me the least grief, but euler ancestral + normal is claimed to be decent.
LoRAs I used:
i2v FusionX 0.8 weight (all sections):
https://huggingface.co/vrgamedevgirl84/Wan14BT2VFusioniX/tree/main/FusionX_LoRa
t2v innie pussy LoRA 0.6 weight (all sections):
https://civitai.com/models/1602000
i2v penis LoRA 0.3 weight (all sections):
https://civitai.com/models/1476909?modelVersionId=2284083
i2v facial 0.7 weight (section 2 only):
https://civitai.com/models/1962545/cumfacial-wan22
I use FusionX LoRA since it was cleaning up noise on realism gens, but may not always be necessary. Using both High noise & Low noise LoRA for everything else. Innie pussy LoRA generally wasn't needed, but it was smudging details there when she touched herself so I included the LoRA to prevent that. Lightning LoRA in my workflow is not needed if using a model that has it built in like SmoothMix so for the Ashsuki vid I swapped the Lightning LoRA nodes for penis LoRA at 0.3 weight since my first attempt erased the head of the cock on section 1. WAN 2.1 LoRAs & VAE are compatible with WAN 2.2 & generally I saw others using one WAN 2.1 LoRA for High & Low noise with High at 1.0 & Low at 1.5 weight. I tried doing that 1.0 / 1.5 weight with an anime LoRA to try to make 2D motion cleaner before, but I didn't use that here since gradient shading exists.
Precondition:
SmoothMixAnime. Full motion animated video. Dark room background. A small anime girl named FC with blue eyes, pale lips, fluffy dark hair, innie pussy, sitting on a cushion on the left. A tall man named MC stands on the right.
String Selector:
FC is sitting in front of MC, MC is grabbing her head and pulls her head towards his penis, then MC starts face fucking her, MC thrusts his hips back and forth moving the penis inside her mouth, then FC gives him a deepthroat blowjob, MC is grabbing her head and he pulls her head towards his midsection moving the penis deeper into her mouth, FC gags on the penis but still keeps the whole penis in her mouth, she is gagging. FC continuously rubs her innie pussy with one hand.
FC closes her eyes. MC slips his penis out of her mouth revealing his entire penis. MC removes his right hand from FC's head and then strokes his penis back and forth with his right hand. Then cum shoots out of the penis and lands on FC's face and her mouth and her chest. Big white cum load covers FC's face and chest. FC continuously rubs her innie pussy with one hand.
Postcondition:
The camera is still. Motion starts immediately. High quality animation with smooth motion.
I name characters with simple initials like MC for Male Character so it is quicker to type out prompts & perhaps avoid confusion in the AI. Some use of pronouns is fine, but I think names help a little.
Tip: Use Breast Helper LoRA to undress:
https://civitai.com/models/1419535 with prompt "tiny sized breasts with pale medium areole" to keep chest flat. There's also Small Breasts LoRA:
https://civitai.com/models/1983079 but trigger words seem less well defined.
Dunno all nodes this workflow uses. Tried recreating prompt & everything from notes for screenshot.
Upside of scenario workflow is not only a variety of actions, but also a preview of each section so you can cancel if it sucks. Not as granular as Framepack (which I've also used), but something. Don't recommend setting duration length super low since you can lose details between sections. I usually gen 2-3 sections, but can go up to 5 with this. Workflow looks complicated since ComfyUI over-complicates everything visually, but you only really mess with green nodes until you want LoRAs. Recommend rgthree node for a lot of nice features, but also enable setting in pic #2 to quickly toggle bypass of all nodes in a group. I use it to toggle high-low noise nodes.