Huggingface demo

[motowntek]

Hello. Since the model is already on Huggingface, can you put up a useable demo on Huggingface? I'm sure HF will give you a Zero Space. Or at minimum a Space that we can duplicate.
Thanks.

[jjmlovesgit]

https://github.com/jjmlovesgit/contentv

videowatch.mp4

import os
import torch
import gradio as gr
from diffusers.utils import export_to_video
from diffusers import AutoencoderKL
from contentv_transformer import SD3Transformer3DModel
from contentv_pipeline import ContentVPipeline

--- NPU Setup (Keep as is if you need it) ---

USE_ASCEND_NPU = int(os.getenv('USE_ASCEND_NPU', '0'))
if USE_ASCEND_NPU:
import torch_npu
from torch_npu.contrib import transfer_to_npu
torch.npu.config.allow_internal_format = False

---------------------------------------------

--- Model Loading ---

print("Loading models... This may take a moment.")
model_id = 'ByteDance/ContentV-8B'
try:
vae = AutoencoderKL.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
except Exception as e:
print(f"Could not load VAE with AutoencoderKL, trying AutoencoderKLWan: {e}")
from diffusers import AutoencoderKLWan
vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)

transformer = SD3Transformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.bfloat16)
pipe = ContentVPipeline.from_pretrained(model_id, vae=vae, transformer=transformer, torch_dtype=torch.bfloat16)
pipe.to("cuda") # Ensure your device is 'cuda' or 'npu' based on your setup
print("Models loaded successfully!")

-----------------------------------------------------------------

def generate_video_gradio(prompt: str, negative_prompt: str, num_frames: int, seed: int, progress=gr.Progress()) -> str:
"""
Generates a video based on the given prompts, number of frames, and a seed,
with progress updates to the Gradio UI.

Args:
    prompt (str): The text prompt for video generation.
    negative_prompt (str): The negative prompt to guide video generation away from certain concepts.
    num_frames (int): The number of frames in the output video.
    seed (int): The random seed for reproducibility.
    progress (gr.Progress): Gradio's progress object for UI updates.

Returns:
    str: The file path to the generated video.
"""
print(f"Generating video with prompt: '{prompt}', negative prompt: '{negative_prompt}', frames: {num_frames}, seed: {seed}")

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

# Define the callback function to update Gradio's progress bar
# Adjusted the signature based on contentv_pipeline.py line 381
def internal_diffusers_callback(pipeline_instance, current_step: int, timestep: int, callback_kwargs):
    # `current_step` is the actual step value we need for progress
    # `pipeline_instance` is the `self` from the pipeline, we don't directly use it for progress here.
    
    # Based on your previous console output (e.g., "4/50"),
    # we'll assume a total of 50 diffusion steps for the primary progress.
    total_diffusion_steps = 50 
    
    if current_step is not None and isinstance(current_step, int):
        current_progress_percentage = (current_step + 1) / total_diffusion_steps
        progress(current_progress_percentage, desc=f"Generating video... (Step {current_step+1}/{total_diffusion_steps})")
    else:
        # This case should ideally no longer be hit if the pipeline consistently passes integers for the second arg
        print(f"Warning: internal_diffusers_callback received invalid 'current_step' value: {current_step}. Skipping progress update.")
    
    # Return an empty dictionary to satisfy ContentVPipeline's expectation
    return {}


video_frames = pipe(
    num_frames=num_frames,
    prompt=prompt,
    negative_prompt=negative_prompt,
    generator=generator,
    callback_on_step_end=internal_diffusers_callback,
).frames[0]

# Ensure the progress bar completes to 100% after video generation
progress(1, desc="Video generation complete!")

output_path = "generated_video.mp4"
export_to_video(video_frames, output_path, fps=24)
print(f"Video saved to {output_path}")
return output_path

--- Gradio Interface ---

iface = gr.Interface(
fn=generate_video_gradio,
inputs=[
gr.Textbox(
label="Prompt",
lines=3,
value="Premium smartwatch rotates elegantly on a glass surface in a minimalist studio. 360-degree camera movement revealing all angles. Clean, bright lighting with subtle reflections. Color palette: all white, monochrome tones, Nikon D850 DSLR 200mm f/1.8 lens, f/2.2 aperture,3D, UHD",
),
gr.Textbox(
label="Negative Prompt",
lines=2,
value="overexposed, low quality, deformation, text, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion",
),
gr.Slider(
minimum=50,
maximum=200,
step=5,
value=125,
label="Number of Frames"
),
gr.Number(
label="Seed (for reproducibility)",
value=42,
minimum=0,
step=1
)
],
outputs=gr.Video(label="Generated Video"),
title="ContentV Video Generator",
description="Generate high-quality videos from text prompts using the ContentV model. Adjust prompts, frame count, and seed for desired results.",
theme=gr.themes.Soft(primary_hue="teal", secondary_hue="teal"),
)

if name == "main":
iface.launch()

[jjmlovesgit]

Done locally on 5090:

[jjmlovesgit]