|
|
|
|
|
import gradio as gr |
|
|
import numpy as np |
|
|
import random |
|
|
import torch |
|
|
import spaces |
|
|
|
|
|
from PIL import Image |
|
|
from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig |
|
|
|
|
|
import os |
|
|
|
|
|
from huggingface_hub import hf_hub_download |
|
|
|
|
|
|
|
|
|
|
|
pipe = QwenImagePipeline.from_pretrained( |
|
|
torch_dtype=torch.bfloat16, |
|
|
device="cuda", |
|
|
model_configs=[ |
|
|
ModelConfig(model_id="Qwen/Qwen-Image-Edit-2509", |
|
|
download_source='huggingface', |
|
|
origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors"), |
|
|
ModelConfig(model_id="Qwen/Qwen-Image-Edit-2509", |
|
|
download_source='huggingface',origin_file_pattern="text_encoder/model*.safetensors"), |
|
|
ModelConfig(model_id="Qwen/Qwen-Image-Edit-2509", |
|
|
download_source='huggingface',origin_file_pattern="vae/diffusion_pytorch_model.safetensors"), |
|
|
], |
|
|
tokenizer_config=None, |
|
|
processor_config=ModelConfig(model_id="Qwen/Qwen-Image-Edit-2509", |
|
|
download_source='huggingface',origin_file_pattern="processor/"), |
|
|
) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
speedup = hf_hub_download(repo_id="Tele-AI/TeleStyle", filename="weights/diffsynth_Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors") |
|
|
telestyle= hf_hub_download(repo_id="Tele-AI/TeleStyle", filename="weights/diffsynth_Qwen-Image-Edit-2509-telestyle.safetensors") |
|
|
|
|
|
|
|
|
pipe.load_lora(pipe.dit, telestyle) |
|
|
pipe.load_lora(pipe.dit,speedup) |
|
|
|
|
|
|
|
|
|
|
|
dtype = torch.bfloat16 |
|
|
device = "cuda" if torch.cuda.is_available() else "cpu" |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
MAX_SEED = np.iinfo(np.int32).max |
|
|
|
|
|
|
|
|
@spaces.GPU |
|
|
def infer( |
|
|
content_ref, |
|
|
style_ref, |
|
|
prompt, |
|
|
seed=123, |
|
|
randomize_seed=False, |
|
|
true_guidance_scale=1.0, |
|
|
num_inference_steps=4, |
|
|
minedge=1024, |
|
|
progress=gr.Progress(track_tqdm=True), |
|
|
|
|
|
): |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
content_ref=Image.fromarray(content_ref) |
|
|
style_ref=Image.fromarray(style_ref) |
|
|
|
|
|
if randomize_seed: |
|
|
seed = random.randint(0, MAX_SEED) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
w,h=content_ref.size |
|
|
|
|
|
|
|
|
|
|
|
minedge=minedge-minedge%16 |
|
|
if w>h: |
|
|
r=w/h |
|
|
h=minedge |
|
|
w=int(h*r)-int(h*r)%16 |
|
|
|
|
|
else: |
|
|
r=h/w |
|
|
w=minedge |
|
|
h=int(w*r)-int(w*r)%16 |
|
|
|
|
|
|
|
|
|
|
|
print(f"Calling pipeline with prompt: '{prompt}'") |
|
|
|
|
|
print(f"Seed: {seed}, Steps: {num_inference_steps}, Guidance: {true_guidance_scale}, Size: {w}x{h}") |
|
|
|
|
|
images = [ |
|
|
content_ref.resize((w, h)), |
|
|
style_ref.resize((minedge, minedge)) , |
|
|
] |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
image = pipe(prompt, edit_image=images, seed=seed, num_inference_steps=num_inference_steps, height=h, width=w,edit_image_auto_resize=False,cfg_scale=true_guidance_scale) |
|
|
|
|
|
|
|
|
return image, seed |
|
|
|
|
|
|
|
|
examples = [] |
|
|
|
|
|
|
|
|
|
|
|
_HEADER_ = ''' |
|
|
<div style="text-align: center; max-width: 650px; margin: 0 auto;"> |
|
|
<h1 style="font-size: 2.5rem; font-weight: 700; margin-bottom: 1rem; display: contents;">TeleStyle</h1> |
|
|
|
|
|
</div> |
|
|
|
|
|
|
|
|
<p style="font-size: 1rem; margin-bottom: 1.5rem;">Paper: <a href='https://arxiv.org/abs/2601.20175' target='_blank'>TeleStyle: Content-Preserving Style Transfer in Images and Videos</a> | Codes: <a href='https://github.com/Tele-AI/TeleStyle/' target='_blank'>GitHub</a></p> |
|
|
<p style="font-size: 1rem; margin-bottom: 1.5rem;">If you encounter an Error with this demo, the most possible reason is ZeroGPU out-of-memory and the solution is to decrease the Min Edge of the generated image from 1024 to a lower value. This is because ZeroGPU has a memory limit of 70GB, while all the examples are tested with 80GB H100 GPUs. </p> |
|
|
''' |
|
|
|
|
|
with gr.Blocks() as demo: |
|
|
|
|
|
with gr.Column(elem_id="col-container"): |
|
|
|
|
|
gr.Markdown(_HEADER_) |
|
|
gr.Markdown("This is a demo of TeleStyle-Image, enabling Content-Preserving Style Transfer capability to Qwen-Image-Edit-2509.") |
|
|
with gr.Row(): |
|
|
with gr.Column(): |
|
|
with gr.Row(): |
|
|
content_ref = gr.Image(label="content ref", type="numpy", ) |
|
|
style_ref = gr.Image(label="style ref", type="numpy", ) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
result = gr.Image(label="Result", show_label=True, type="pil") |
|
|
|
|
|
with gr.Row(): |
|
|
prompt = gr.Text( |
|
|
label="Prompt", |
|
|
value='Style Transfer the style of Figure 2 to Figure 1, and keep the content and characteristics of Figure 1.', |
|
|
show_label=True, |
|
|
placeholder='Style Transfer the style of Figure 2 to Figure 1, and keep the content and characteristics of Figure 1.', |
|
|
container=True, |
|
|
) |
|
|
run_button = gr.Button("Edit!", variant="primary") |
|
|
|
|
|
with gr.Accordion("Advanced Settings", open=True): |
|
|
|
|
|
|
|
|
seed = gr.Slider( |
|
|
label="Seed", |
|
|
minimum=0, |
|
|
maximum=MAX_SEED, |
|
|
step=1, |
|
|
value=123, |
|
|
) |
|
|
|
|
|
randomize_seed = gr.Checkbox(label="Randomize seed", value=False) |
|
|
|
|
|
with gr.Row(): |
|
|
|
|
|
true_guidance_scale = gr.Slider( |
|
|
label="CFG should be 1.0", |
|
|
minimum=0, |
|
|
maximum=10.0, |
|
|
step=0.1, |
|
|
value=1.0 |
|
|
) |
|
|
|
|
|
num_inference_steps = gr.Slider( |
|
|
label="Number of inference steps should be 4", |
|
|
minimum=1, |
|
|
maximum=50, |
|
|
step=1, |
|
|
value=4, |
|
|
) |
|
|
|
|
|
minedge = gr.Slider( |
|
|
label="Min Edge of the generated image", |
|
|
minimum=256, |
|
|
maximum=2048, |
|
|
step=8, |
|
|
value=1024, |
|
|
) |
|
|
|
|
|
|
|
|
with gr.Row(), gr.Column(): |
|
|
gr.Markdown("## Examples") |
|
|
gr.Markdown("changing the minedge could lead to different style similarity.") |
|
|
default_prompt='Style Transfer the style of Figure 2 to Figure 1, and keep the content and characteristics of Figure 1.' |
|
|
gr.Examples(examples=[ |
|
|
['./qwenstyleref/pulpfiction_2.jpg','./qwenstyleref/styleref=6_style_ref.png',default_prompt,123,False,1.0,4,832], |
|
|
['./qwenstyleref/styleref=0_content_ref.png','./qwenstyleref/110.png',default_prompt,123,False,1.0,4,832], |
|
|
['./qwenstyleref/romanholiday_1.jpg','./qwenstyleref/s0099____1113_01_query_1_img_000146_1682705733350_08158389675901344.jpg.jpg',default_prompt,123,False,1.0,4,800], |
|
|
['./qwenstyleref/styleref=0_content_ref.png','./qwenstyleref/125.png',default_prompt,123,False,1.0,4,832], |
|
|
['./qwenstyleref/fallenangle.jpg','./qwenstyleref/styleref=s0038.png',default_prompt,123,False,1.0,4,832], |
|
|
['./qwenstyleref/styleref=0_content_ref.png','./qwenstyleref/styleref=s0572.png',default_prompt,123,False,1.0,4,832], |
|
|
['./qwenstyleref/startrooper1.jpg','./qwenstyleref/david-face-760x985.jpg','Style Transfer Figure 1 into marble material.',123,False,1.0,4,1024], |
|
|
['./qwenstyleref/startrooper1.jpg','./qwenstyleref/125.png',default_prompt, 123,False,1.0,4,1024], |
|
|
['./qwenstyleref/possession.png','./qwenstyleref/s0026____0907_01_query_0_img_000194_1682674358294_041656249089406583.jpeg.jpg',default_prompt,123,False,1.0,4,832], |
|
|
['./qwenstyleref/styleref=0_content_ref.png','./qwenstyleref/Jotarokujo.webp',default_prompt,123,False,1.0,4,832], |
|
|
['./qwenstyleref/wallstreet1.jpg','./qwenstyleref/034.png',default_prompt,123,False,1.0,4,1024], |
|
|
['./qwenstyleref/bird.jpeg','./qwenstyleref/styleref=s0539.png',default_prompt,123,False,1.0,4,832], |
|
|
|
|
|
|
|
|
|
|
|
], |
|
|
inputs=[content_ref, |
|
|
style_ref, |
|
|
prompt, |
|
|
seed, |
|
|
randomize_seed, |
|
|
true_guidance_scale, |
|
|
num_inference_steps, |
|
|
minedge,], |
|
|
|
|
|
outputs=[result, seed], |
|
|
fn=infer, |
|
|
cache_examples=False |
|
|
) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
gr.on( |
|
|
triggers=[run_button.click], |
|
|
fn=infer, |
|
|
inputs=[ |
|
|
content_ref, |
|
|
style_ref, |
|
|
prompt, |
|
|
seed, |
|
|
randomize_seed, |
|
|
true_guidance_scale, |
|
|
num_inference_steps, |
|
|
minedge, |
|
|
|
|
|
], |
|
|
outputs=[result, seed], |
|
|
) |
|
|
|
|
|
if __name__ == "__main__": |
|
|
demo.launch(server_name='0.0.0.0') |
