app.py

import gradio as gr
import torch
from diffusers import AnimateDiffPipeline, DDIMScheduler, MotionAdapter
from diffusers.utils import export_to_gif
from diffusers.utils import export_to_video
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import uuid
import spaces

# Available adapters (replace with your actual adapter names)
adapter_options = {
    "zoom-out":"guoyww/animatediff-motion-lora-zoom-out",
    "zoom-in":"guoyww/animatediff-motion-lora-zoom-in",
    "pan-left":"guoyww/animatediff-motion-lora-pan-left",
    "pan-right":"guoyww/animatediff-motion-lora-pan-right",
    "roll-clockwise":"guoyww/animatediff-motion-lora-rolling-clockwise",
    "roll-anticlockwise":"guoyww/animatediff-motion-lora-rolling-anticlockwise",
    "tilt-up":"guoyww/animatediff-motion-lora-tilt-up",
    "tilt-down":"guoyww/animatediff-motion-lora-tilt-down"
}

device = "cuda"
adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-2", torch_dtype=torch.float16)
model_id = "SG161222/Realistic_Vision_V5.1_noVAE"

model_llm = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-128k-instruct",
    device_map="cuda",
    torch_dtype="auto",
    trust_remote_code=True,
    device="cuda"
)
tokenizer_llm = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-128k-instruct",device="cuda")

pipe = AnimateDiffPipeline.from_pretrained(model_id, motion_adapter=adapter, torch_dtype=torch.float16).to(device)
scheduler = DDIMScheduler.from_pretrained(
    model_id,
    subfolder="scheduler",
    clip_sample=False,
    timestep_spacing="linspace",
    beta_schedule="linear",
    steps_offset=1,
)
pipe.scheduler = scheduler

@spaces.GPU
def generate_video(prompt,negative_prompt, guidance_scale, num_inference_steps, adapter_choices):

    pipe.to(device)

    messages = [
        {"role": "user", "content": "You have to complete my given prompt into a complete description. The description should be heavily detailed. Feel free to add your own fillers if need. The purpose of this description is to descibe a video generation. My Prompt: " + prompt},
    ]

    pipe_llm = pipeline(
        "text-generation",
        model=model_llm,
        tokenizer=tokenizer_llm,
        device="cuda"
    )

    generation_args = {
        "max_new_tokens": 500,
        "return_full_text": False,
        "temperature": 1,
        "do_sample": False,
    }

    output = pipe_llm(messages, **generation_args)
    print(output[0]['generated_text'])

    # Set adapters based on user selection
    if adapter_choices:
        for i in range(len(adapter_choices)):
            adapter_name = adapter_choices[i]
            pipe.load_lora_weights(
                adapter_options[adapter_name], adapter_name=adapter_name,
            )
        pipe.set_adapters(adapter_choices, adapter_weights=[1.0] * len(adapter_choices))
        print(adapter_choices)

    output = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_frames=16,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
    )
    name = str(uuid.uuid4()).replace("-", "")
    path = f"/tmp/{name}.mp4"
    export_to_video(output.frames[0], path, fps=10)
    return path



iface = gr.Interface(
    theme=gr.themes.Soft(primary_hue="red", secondary_hue="pink"),
    fn=generate_video,
    inputs=[
        gr.Textbox(label="Enter your prompt"),
        gr.Textbox(label="Negative Prompt"),
        gr.Slider(minimum=0.5, maximum=10, value=7.5, label="Guidance Scale"),
        gr.Slider(minimum=4, maximum=24, step=4, value=4, label="Inference Steps"),
        gr.CheckboxGroup(adapter_options.keys(), label="Adapter Choice",type='value'),
        # Updated for multiple selections
    ],
    outputs=gr.Video(label="Generated Video"),
)

iface.launch()