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 from pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline from qwen_vl_utils import process_vision_info import os from huggingface_hub import hf_hub_download def update_textbox(selected_items): # Join the selected list of strings into a comma-separated string return ", ".join(selected_items) pipe = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2509", torch_dtype=torch.bfloat16) print("pipeline loaded") pipe.to('cuda') pipe.set_progress_bar_config(disable=None) ''' 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="witcherderivia/Qwen-Image-Style-Transfer", filename="diffsynth_Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors") qwenstyle= hf_hub_download(repo_id="witcherderivia/Qwen-Image-Style-Transfer", filename="diffusers_Qwen-Image-Edit-2509-Style-Transfer-V1.safetensors") pipe.load_lora_weights( qwenstyle,adapter_name='style' ) pipe.load_lora_weights( speedup,adapter_name='dmd' ) pipe.set_adapters(["style", "dmd",], adapter_weights=[1.0, 1.0]) pipe.fuse_lora(adapter_names=["style", "dmd"], lora_scale=1.0) pipe.unload_lora_weights() dtype = torch.bfloat16 device = "cuda" if torch.cuda.is_available() else "cpu" MAX_SEED = np.iinfo(np.int32).max @spaces.GPU(size="xlarge") 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), checkbox=[], ): content_text_input='describe main objects (fewer than 3) with separated words, each word is separated by comma, the total number of words is strictly fewer than 3' style_text_input='describe only the artistic style, material and stroke in 5 words, not objects.' #pipe.text_encoder.eval() content_prompt='' style_prompt='' if content_ref is not None: content_ref=Image.fromarray(content_ref) content_messages = [ { "role": "user", "content": [ { "type": "image", "image": content_ref, }, {"type": "text", "text": content_text_input}, ], } ] content_text = pipe.processor.apply_chat_template( content_messages, tokenize=False, add_generation_prompt=True ) image_inputs, video_inputs = process_vision_info(content_messages) inputs = pipe.processor( text=[content_text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt", ) inputs = inputs.to(device) # Inference: Generation of the output generated_ids = pipe.text_encoder.generate(**inputs, max_new_tokens=1024) generated_ids_trimmed = [ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids) ] content_prompt = pipe.processor.batch_decode( generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False )[0] print(f"content_prompt={content_prompt}") if style_ref is not None: style_ref=Image.fromarray(style_ref) style_messages = [ { "role": "user", "content": [ { "type": "image", "image": style_ref, }, {"type": "text", "text": style_text_input}, ], } ] style_text = pipe.processor.apply_chat_template( style_messages, tokenize=False, add_generation_prompt=True ) image_inputs, video_inputs = process_vision_info(style_messages) inputs = pipe.processor( text=[style_text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt", ) inputs = inputs.to(device) # Inference: Generation of the output generated_ids = pipe.text_encoder.generate(**inputs, max_new_tokens=1024) generated_ids_trimmed = [ out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids) ] style_prompt = pipe.processor.batch_decode( generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False )[0] print(f"style_prompt={style_prompt}") if randomize_seed: seed = random.randint(0, MAX_SEED) sw,sh,w,h=0,0,0,0 if content_ref: w,h=content_ref.size #minedge=1024 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 if style_ref: sw,sh=style_ref.size if sw>sh: r=sw/sh sh=minedge sw=int(sh*r)-int(sh*r)%16 else: r=sh/sw sw=minedge sh=int(sw*r)-int(sw*r)%16 print(f"Seed: {seed}, Steps: {num_inference_steps}, Guidance: {true_guidance_scale},") if content_ref and style_ref: images = [ content_ref.resize((w, h)), style_ref.resize((sw, sh)) , #style_ref.resize((minedge, minedge)) , ] elif content_ref: images = [ content_ref.resize((w, h)), #style_ref.resize((sw, sh)) , #style_ref.resize((minedge, minedge)) , ] elif style_ref: images = [ #content_ref.resize((w, h)), style_ref.resize((sw, sh)) , #style_ref.resize((minedge, minedge)) , ] if "infer with content prompt" in checkbox and content_prompt not in prompt: prompt=','.join([prompt,content_prompt]) if "infer with style prompt" in checkbox and style_prompt not in prompt: prompt=','.join([prompt,style_prompt]) if "infer with content prompt" not in checkbox and content_prompt in prompt: prompt=prompt.replace(content_prompt.strip(','),'') if "infer with style prompt" not in checkbox and style_prompt in prompt: prompt=prompt.replace(style_prompt.strip(),'') prompt=prompt.strip(',') print(f"Calling pipeline with prompt: '{prompt}'") inputs = { "image": images, "prompt": prompt, "generator": torch.manual_seed(seed), "true_cfg_scale": true_guidance_scale, "negative_prompt": " ", "num_inference_steps": num_inference_steps, "guidance_scale": true_guidance_scale, "num_images_per_prompt": 1, "width": w or sw, "height": h or sh, } with torch.inference_mode(): image = pipe(**inputs) image = image.images[0] return image, seed, content_prompt, style_prompt, prompt # --- Examples and UI Layout --- examples = [] _HEADER_ = '''

QwenStyle V1

Paper: QwenStyle: Content-Preserving Style Transfer with Qwen-Image-Edit | Codes: GitHub

Update: prompt enhancer provided, and the model supports content ref/style ref only input, which means you could use the model as an image editing model and style transfer model at the same time. So you don't have to provide a style reference now, the model also accepts prompt for style transfer, which makes the model more flexible.

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.

''' with gr.Blocks() as demo: with gr.Column(elem_id="col-container"): gr.HTML('Qwen-Image Logo') gr.Markdown(_HEADER_) gr.Markdown("This is a demo of QwenStyle v1, the first Content-Preserving Style Transfer Lora on 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", ) #print(f"type(content_ref)={type(content_ref)}") #input_images = gr.Gallery(label="Input Images", show_label=False, type="pil", interactive=True) result = gr.Image(label="Result", show_label=True, type="pil") #result = gr.Gallery(label="Result", show_label=True, type="pil") with gr.Column(): checkbox=gr.CheckboxGroup(["infer with content prompt", "infer with style prompt"], label="Prompt Enhancer", ) content_prompt=gr.Text( label="Content Reference Prompt", show_label=True, container=True, ) style_prompt=gr.Text( label="Style Reference Prompt", show_label=True, container=True, ) 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): # Negative prompt UI element is removed here 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, content_prompt, style_prompt,prompt], fn=infer, cache_examples=False ) # gr.Examples(examples=examples, inputs=[prompt], 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, checkbox, ], outputs=[result, seed, content_prompt, style_prompt,prompt], ) if __name__ == "__main__": demo.launch(server_name='0.0.0.0')