Wan 2.2 i2v under 18 GB VRAM and 32 GB RAM

wan_example.py

import os
import torch
import gc
from diffusers import WanImageToVideoPipeline, WanTransformer3DModel
from diffusers.utils import export_to_video, load_image
from transformers import BitsAndBytesConfig
from PIL import Image

PROMPT = "A cinematic shot of a white cat surfing on a wave, wearing sunglasses, sunny beach background, high quality, 4k"
NEGATIVE_PROMPT = "low quality, bad hands, distorted, blur, motion artifacts"

INPUT_IMAGE_PATH = "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG"

OUTPUT_PATH = "C:/Tmp/wan_result.mp4"

print("--- Initializing ---")
gc.collect()
torch.cuda.empty_cache()

MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"

nf4_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

print("--- Loading Models (This takes a moment) ---")

transformer_high = WanTransformer3DModel.from_pretrained(
    MODEL_ID,
    subfolder="transformer",
    quantization_config=nf4_config,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True
)

transformer_low = WanTransformer3DModel.from_pretrained(
    MODEL_ID,
    subfolder="transformer_2",
    quantization_config=nf4_config,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True
)

pipe = WanImageToVideoPipeline.from_pretrained(
    MODEL_ID,
    transformer=transformer_high,
    transformer_2=transformer_low,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True
)

pipe.enable_model_cpu_offload()

print("--- Loading LoRAs ---")
try:
    pipe.load_lora_weights(
        "Kijai/WanVideo_comfy",
        weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
        adapter_name="lightx2v"
    )
    pipe.load_lora_weights(
        "Kijai/WanVideo_comfy",
        weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
        adapter_name="lightx2v_2",
        load_into_transformer_2=True
    )
    pipe.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1.0, 1.0])
    print("LoRAs loaded successfully.")
except Exception as e:
    print(f"LoRA Load Failed: {e}")
    print("Continuing without LoRA (Standard Speed)...")

print("--- Preparing Image ---")
if INPUT_IMAGE_PATH.startswith("http"):
    input_image = load_image(INPUT_IMAGE_PATH)
else:
    input_image = Image.open(INPUT_IMAGE_PATH)

def resize_for_wan(image, max_dim=832):
    w, h = image.size
    scale = max_dim / max(w, h)
    new_w = int(w * scale)
    new_h = int(h * scale)
    new_w = (new_w // 16) * 16
    new_h = (new_h // 16) * 16
    return image.resize((new_w, new_h), Image.LANCZOS)

input_image = resize_for_wan(input_image)

print("--- Generating Video ---")
gc.collect()
torch.cuda.empty_cache()

seed = 42
generator = torch.Generator(device="cuda").manual_seed(seed)

output_frames = pipe(
    image=input_image,
    prompt=PROMPT,
    negative_prompt=NEGATIVE_PROMPT,
    height=input_image.height,
    width=input_image.width,
    num_frames=81,
    num_inference_steps=8, #Needs this number or it'll crossfade in the middle.
    guidance_scale=1.0,
    guidance_scale_2=1.0,
    generator=generator
).frames[0]

export_to_video(output_frames, OUTPUT_PATH, fps=16)
print(f"DONE! Video saved to: {OUTPUT_PATH}")