dev/app3.py

import gradio as gr
import torch
from diffusers import StableDiffusionPipeline
import os
import uuid
import animation_logic as anim
import video_utils as vid

# --- Model Config (SDXS Optimized) ---
device = "cpu"
model_id = "IDKiro/sdxs-512-dreamshaper"
# Using float32 for CPU stability
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float32)
pipe.to(device)

def run_deforum(
    prompt_list_str, neg_prompt, max_frames, 
    zoom_str, angle_str, tx_str, ty_str,
    cadence, fps, color_match
):
    width, height = 256, 256
    try:
        # Expected format: {0: "prompt", 10: "prompt"}
        prompts = eval(prompt_list_str)
    except Exception as e:
        return None, None, f"Error parsing prompts: {str(e)}"

    # Parse Schedules
    zoom_s = anim.parse_keyframe_string(zoom_str, int(max_frames))
    angle_s = anim.parse_keyframe_string(angle_str, int(max_frames))
    tx_s = anim.parse_keyframe_string(tx_str, int(max_frames))
    ty_s = anim.parse_keyframe_string(ty_str, int(max_frames))
    
    all_frames = []
    prev_gen_frame = None
    first_frame = None
    
    for f in range(int(max_frames)):
        if f % cadence == 0:
            current_prompt = prompts[max(k for k in prompts.keys() if k <= f)]
            
            if prev_gen_frame is not None:
                # Warp
                init_image = anim.anim_frame_warp(prev_gen_frame, angle_s[f], zoom_s[f], tx_s[f], ty_s[f])
                # SDXS 1-step Inference
                new_frame = pipe(
                    current_prompt, 
                    image=init_image, 
                    negative_prompt=neg_prompt,
                    num_inference_steps=1,
                    guidance_scale=0.0,
                    width=width, height=height
                ).images[0]
                
                if color_match and first_frame is not None:
                    new_frame = anim.maintain_colors(first_frame, new_frame)
            else:
                # First frame base generation
                new_frame = pipe(
                    current_prompt, 
                    negative_prompt=neg_prompt,
                    num_inference_steps=1,
                    guidance_scale=0.0,
                    width=width, height=height
                ).images[0]
                first_frame = new_frame

            # Cadence Interpolation
            if cadence > 1 and prev_gen_frame is not None:
                for i in range(1, cadence):
                    alpha = i / cadence
                    interp_frame = anim.lerp_frames(prev_gen_frame, new_frame, alpha)
                    all_frames.append(interp_frame)
            
            all_frames.append(new_frame)
            prev_gen_frame = new_frame
            yield new_frame, None, None
            
    # Post-Processing
    unique_id = uuid.uuid4().hex[:6]
    video_file = vid.frames_to_video(all_frames, f"deforum_{unique_id}.mp4", fps)
    zip_file = vid.export_to_zip(all_frames, f"frames_{unique_id}.zip")
    
    yield all_frames[-1], video_file, zip_file

# --- UI Setup ---
with gr.Blocks() as demo:
    gr.Markdown("# 🚀 Deforum SOON® Animator")
    
    with gr.Row():
        with gr.Column():
            p_input = gr.Textbox(label="Prompt Map", value='{0: "hyperrealistic forest", 10: "burning forest"}')
            n_input = gr.Textbox(label="Negative Prompt", value="blurry, text, watermark")
            
            with gr.Row():
                f_count = gr.Slider(5, 100, value=20, step=1, label="Frames")
                c_val = gr.Slider(1, 5, value=1, step=1, label="Cadence")
                fps_val = gr.Number(label="FPS", value=10)
                color_check = gr.Checkbox(label="Color Match", value=True)

            with gr.Accordion("2D Motion (Keyframes)", open=False):
                z_in = gr.Textbox(label="Zoom", value="0:(1.03)")
                a_in = gr.Textbox(label="Angle", value="0:(0)")
                tx_in = gr.Textbox(label="TX", value="0:(0)")
                ty_in = gr.Textbox(label="TY", value="0:(0)")
            
            run_btn = gr.Button("Generate", variant="primary")
            
        with gr.Column():
            preview = gr.Image(label="Live Stream")
            video_out = gr.Video(label="Final Render")
            file_out = gr.File(label="Batch PNGs")

    run_btn.click(
        fn=run_deforum,
        inputs=[p_input, n_input, f_count, z_in, a_in, tx_in, ty_in, c_val, fps_val, color_check],
        outputs=[preview, video_out, file_out]
    )

demo.launch(theme='SebastianBravo/simci_css')