app.py

try:
    import spaces
except ImportError:
    class spaces:
        @staticmethod
        def GPU(fn):
            return fn
        
import gradio as gr
import numpy as np
import imageio

import torch
from diffusers import AutoencoderKLWan
from vsfwan.pipeline import WanPipeline
from vsfwan.processor import WanAttnProcessor2_0
from diffusers import WanVACEPipeline
from diffusers.utils import export_to_video
import uuid

import sys
import os
model_id = "stabilityai/stable-diffusion-3.5-large-turbo"
from src.sd3_pipeline import VSFStableDiffusion3Pipeline
pipe = VSFStableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-large-turbo",
    torch_dtype=torch.bfloat16,
    hf_token=os.environ.get("HF_TOKEN", None)
) 

from nag import NAGStableDiffusion3Pipeline
nag_pipe = NAGStableDiffusion3Pipeline.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    token="hf_token",
)
# pipe = pipe.to("cuda")
# nag_pipe = nag_pipe.to("cuda")
import os
@spaces.GPU
def generate_video(positive_prompt, negative_prompt, guidance_scale, bias, step, seed, nag_guidance, nag_alpha, nag_tau, nag_step, progress=gr.Progress(track_tqdm=False)):
    global pipe, nag_pipe
    lambda total: progress.tqdm(range(total))

    print(f"Generating image with params: {positive_prompt}, {negative_prompt}, {guidance_scale}, {bias}, {step}")

    output = pipe(
        prompt=positive_prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=step,
        guidance_scale=0.0, 
        generator=torch.Generator(device="cuda").manual_seed(seed),
    ).images[0]
    os.makedirs("images", exist_ok=True)
    path = f"images/{uuid.uuid4().hex}.png"
    output.save(path)
    output_path = path
    print(f"Image saved to {output_path}")

    output_nag = nag_pipe(
        prompt=positive_prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=nag_step,
        nag_scale=nag_guidance,
        nag_alpha=nag_alpha,
        nag_tau=nag_tau,
        guidance_scale=0.0,
    ).images[0]
    nag_path = f"images/{uuid.uuid4().hex}_nag.png"
    output_nag.save(nag_path)
    print(f"NAG Image saved to {nag_path}")

    return output_path, nag_path

import json
with open("sample_prompts.json", "r") as f:
    sample_prompts = json.load(f)

def load_sample():
    sample = np.random.choice(sample_prompts)
    return sample['prompt'], sample['missing_element']

with open("anti_aesthetics.json", "r") as f:
    anti_aesthetics_prompts = json.load(f)

def load_anti_aesthetics_sample():
    sample = np.random.choice(anti_aesthetics_prompts)
    return sample['prompt'], sample['missing_element']

nouns = ["cat", "dog", "car", "bicycle", "tree", "house", "computer", "phone", "book", "chair", "table", "lamp", "flower", "mountain", "river", "ocean", "cloud", "bird", "fish", "butterfly"]
methods = ["painting", "sketch", "drawing"]
def load_abstract_prompt():
    noun = np.random.choice(nouns)
    method = np.random.choice(methods)
    prompt = f"An abstract {method} of a {noun}."
    negative = f"{noun}"
    return prompt, negative
# def rephrase_prompt(pos_prompt, neg_prompt):
#     completion = client.chat.completions.create(
#     extra_headers={
#         "HTTP-Referer": "<YOUR_SITE_URL>", # Optional. Site URL for rankings on openrouter.ai.
#         "X-Title": "<YOUR_SITE_NAME>", # Optional. Site title for rankings on openrouter.ai.
#     },
#     extra_body={},
#     model="qwen/qwen3-vl-235b-a22b-instruct",
#     messages=[
#         {
#         "role": "user",
#         "content": "Repahrase the following prompt to one sentence for positive prompt and a few words for negative prompt.\n\nOriginal Prompt: {}\n\nNegative Element: {}. \n make sure the generated prompt follows the positive-negative prompt pair, do not mention the negative prompt in positive one".format(pos_prompt, neg_prompt)
#         }
#     ]
#     )

with gr.Blocks(title="Value Sign Flip SD3.5 Demo") as demo:
    gr.Markdown("# Value Sign Flip SD3.5 Demo \n\n This demo is based on SD3.5-L-Turbo model and uses Value Sign Flip technique to generate videos with different guidance scales and biases. More on [GitHub](https://github.com/weathon/VSF/blob/main/wan.md)\n\nPositive prompt should be at least 1 sentence long or the results will be weird.")
    # gr.Markdown("# Value Sign Flip Wan 2.1 Demo \n\n This demo is based on Wan 2.1 T2V model and uses Value Sign Flip technique to generate videos with different guidance scales and biases. More on [GitHub](https://github.com/weathon/VSF/blob/main/wan.md)\n\nPositive prompt should be at least 2 sentence long or the results will be weird.")

    with gr.Row():
        pos = gr.Textbox(label="Positive Prompt", value="A polished bicycle frame leans against a weathered brick wall under soft morning light.")
        neg = gr.Textbox(label="Negative Prompt", value="wheels")
        
        # rephase = gr.Button("Rephrase Prompt") 
        with gr.Column():
            sample = gr.Button("Load A Sample Prompt")
            sample.click(fn=load_sample, inputs=[], outputs=[pos, neg])
            anti_aesthetic_sample = gr.Button("Load An Anti-Aesthetic Sample Prompt")
            anti_aesthetic_sample.click(fn=load_anti_aesthetics_sample, inputs=[], outputs=[pos, neg])
            abstract_sample = gr.Button("Load An Abstract Prompt")
            abstract_sample.click(fn=load_abstract_prompt, inputs=[], outputs=[pos, neg])
    with gr.Row():
        gr.Markdown("## VSF Generation Parameters")
        guidance = gr.Slider(0, 5, step=0.1, label="Guidance Scale", value=3.0)
        bias = gr.Slider(0, 0.5, step=0.01, label="Bias", value=0.1)
        step = gr.Slider(4, 15, step=1, label="Step", value=8)
        seed = gr.Number(label="Seed", value=0, precision=0)
        set_strong_vsf = gr.Button("Set to VSF Strong Settings")
        set_strong_vsf.click(fn=lambda : (3.8, 0.2), inputs=[], outputs=[guidance, bias])
        set_mild_vsf = gr.Button("Set to VSF Quality Settings")
        set_mild_vsf.click(fn=lambda : (3.3, 0.2), inputs=[], outputs=[guidance, bias])

    with gr.Row():
        gr.Markdown("## NAG Generation Parameters")
        nag_guidance = gr.Slider(1, 10, step=0.1, label="Guidance Scale", value=5)
        nag_alpha = gr.Slider(0.1, 1.0, step=0.01, label="Alpha", value=0.25)
        nag_tau = gr.Slider(1, 10, step=0.01, label="Tau", value=3.0)
        nag_step = gr.Slider(4, 15, step=1, label="Step", value=8)
        set_strong = gr.Button("Set to NAG Strong Settings")
        set_strong.click(fn=lambda : (11, 0.5, 5.0), inputs=[], outputs=[nag_guidance, nag_alpha, nag_tau])
        set_mild = gr.Button("Set to NAG Quality Settings")
        set_mild.click(fn=lambda : (4, 0.125, 2.5), inputs=[], outputs=[nag_guidance, nag_alpha, nag_tau])

        

    with gr.Row():
        vsf_out = gr.Image(label="VSF Generated Image")
        nag_out = gr.Image(label="NAG Generated Image")

    btn = gr.Button("Generate")
    btn.click(fn=generate_video, inputs=[pos, neg, guidance, bias, step, seed, nag_guidance, nag_alpha, nag_tau, nag_step], outputs=[vsf_out, nag_out])

demo.launch(share=True)