LUSTY-5B (T2V, I2V, TI2V)

LUSTY-5B is a video model designed to produce NSFW adult material.

It is a trigger-free LoRA model trained on Wan2.2-TI2V-5B-Diffusers

It is not distilled and not quantized.
Checkout the 2nd trailer here: https://civarchive.com/posts/29179560

🚨 UPDATE for ComfyUI users 🚨

Sorry folks I do not use ComfyUI (please post in the comments if you can help!), my answer and understanding is gonna be very limited, but I believe you can try to load it with this node maybe? ->
https://www.runcomfy.com/comfyui-nodes/ComfyUI-Diffusers

Suggested Parameters

• Width: 1280

• Height: 720 (or 704 depending on your system)

• Inference Steps: 25 (20 for fast preview, 30 for better quality)

• Guidance Scale: 3.0

• Flow Shift: 4.0

• Max frames: 121 (quality will degrade violently beyond that)

• FPS: 30 (and put "30 FPS" in your prompt)

• Prompt template: <your prompt eg '25 FPS, anal sex, pov, caucasian...'>, High quality, 4K, pro quality, intricate details, high definition, consistent, coherent, harmonious, well framed, good design. Skin grain, skin pores. Photography, cinematic, central framing

• Negative prompt template: <your exclusion eg 'asian'>, cropped, out of frame, creepy, weird, strange, piercing, piercings, tattoo, tattoos, shaky camera, jerky, glitchy, focus shift, micro jitter, pixelated, bad quality, overexposed, overexposure, highlights, saturated, animation, manga, comic, watermark, logo, branding, low quality, blurry, distorted, subtitle, subtitles, dubbed

Recommended Usage

• LUSTY-5B is optimized for landscape T2V generation at 720p.

• LUSTY-5B can be prompted naturally (try "underwater blowjob", "anal pov"..)

• Although the model tries to generalize, current version works best for blowjobs videos.

• Tip: Use "POV", "close-up" or "medium shot" to avoid low quality faces

• Tip: Use "focused foreground, unfocused background" to avoid low-quality decors

Disclaimer

• LUSTY-5B is trained on +18 adult actors exclusively, but is not trained to reproduce the likelihood of any person in particular. Nevertheless, please use LUSTY-5B responsibly.

• While we only tested T2V, in theory other transformers pipelines (I2V, TI2V) might work.

• Our current version is not trained on portrait videos and has a lack of diversity in skin tones and fetishes. We apologize for that and would like to improve our model in the future.

Example prompts

• POV view, high-angle shot, medium close-up. 28yo Caucasian female, kneeling in front of the camera, closing her eyes, then looking at the cameraman. She passionately deepthroats the cock. Near a window. Focused foreground, unfocused background

• Getting nude. All cute female quickly removing all clothes to show their naked breasts and nude body. Looking at the camera. Static cam. Quick undressing

• Deepthroat. Profile view. Young woman sucking the cock of a man, choking on the cock. Seen from aside.

• Cum facial, POV. Close-up. Young Caucasian female. The cameraman ejaculates on her face, she receives cum on her forehead, nose, and chin, in her mouth, she spits out the semen.

• Anal POV, medium shot, a latina woman is bouncing on the bed, the cameraman's cock is fucking her ass

• Vaginal sex. High angle. POV Medium close-up. 20yo asian female, being fucked in the pussy in doggy. Outdoor near a river, near the forest. Focused foreground, unfocused background

• Vaginal sex. naked caucasian couple is having sex on a bed, the woman is sitting on top of the male, legs spread open. Shaved vulva, she is being penetrated by the man's big cock, the man is moving up and down to fuck her, rapid fucking.

• Low angle. Close-up on a caucasian female's butt. She spreads her ass. Detailed view of her butthole and tight pussy. In a forest.

• POV, Underwater blowjob, a woman swims underwater to sucks the cameraman's dick

Example code snippet:

from pathlib import Path
import torch
from diffusers import AutoencoderKLWan, WanPipeline
from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
from diffusers.utils import export_to_video

MODEL_ID = "Wan-AI/Wan2.2-TI2V-5B-Diffusers"
LORA_PATH = Path("lusty-5b.safetensors")
OUTPUT_PATH = Path("demo.mp4")

PROMPT = "25 FPS, POV view, high-angle shot, medium close-up. 28yo Caucasian female, kneeling in front of the camera, closing her eyes, then looking at the cameraman. She passionately deepthroats the cock. Near a window. Focused foreground, unfocused background"
PROMPT_SUFFIX = (
    "Focused foreground. Unfocused background. Natural lighting. High quality, "
    "4K, pro quality, intricate details, high definition, consistent, coherent, "
    "harmonious, well framed, good design. Skin grain, skin pores. Photography, "
    "cinematic, central framing."
)
NEGATIVE_PROMPT = (
    "cropped, out of frame, creepy, weird, strange, piercing, piercings, tattoo, "
    "tattoos, shaky camera, jerky, glitchy, focus shift, micro jitter, pixelated, "
    "bad quality, overexposed, overexposure, highlights, saturated, asian, "
    "chinese, animation, manga, comic, watermark, logo, branding, low quality, "
    "blurry, distorted"
)

HEIGHT = 480
WIDTH = 800
NUM_FRAMES = 73  # Must be 4n+1.
NUM_INFERENCE_STEPS = 25
GUIDANCE_SCALE = 3.0
FLOW_SHIFT = 4.0
FPS = 25
SEED = 42
LORA_WEIGHT = 1.0


def main():
    if not torch.cuda.is_available():
        raise RuntimeError("This example expects a CUDA GPU.")
    if not LORA_PATH.exists():
        raise FileNotFoundError(f"LoRA file not found: {LORA_PATH}")

    vae = AutoencoderKLWan.from_pretrained(
        MODEL_ID,
        subfolder="vae",
        torch_dtype=torch.float32,
        low_cpu_mem_usage=True,
    )
    pipe = WanPipeline.from_pretrained(
        MODEL_ID,
        vae=vae,
        torch_dtype=torch.bfloat16,
        low_cpu_mem_usage=True,
    )
    pipe.scheduler = UniPCMultistepScheduler.from_config(
        pipe.scheduler.config,
        flow_shift=FLOW_SHIFT,
    )
    pipe.to("cuda")

    pipe.load_lora_weights(str(LORA_PATH), adapter_name="custom_lora")
    pipe.set_adapters(["custom_lora"], adapter_weights=[LORA_WEIGHT])
    pipe.fuse_lora()

    generator = torch.Generator(device="cuda").manual_seed(SEED)
    frames = pipe(
        prompt=f"{PROMPT} {PROMPT_SUFFIX}",
        negative_prompt=NEGATIVE_PROMPT,
        height=HEIGHT,
        width=WIDTH,
        num_frames=NUM_FRAMES,
        num_inference_steps=NUM_INFERENCE_STEPS,
        guidance_scale=GUIDANCE_SCALE,
        generator=generator,
    ).frames[0]

    OUTPUT_PATH.parent.mkdir(parents=True, exist_ok=True)
    export_to_video(frames, str(OUTPUT_PATH), fps=FPS, quality=8)
    print(f"Wrote {OUTPUT_PATH}")


if __name__ == "__main__":
    main()