Update app.py

app.py

@@ -15,9 +15,6 @@ import os
 from PIL import Image
 import re
 
-# Cohere 모델 초기화
-client = InferenceClient("CohereForAI/c4ai-command-r-plus", token=os.getenv("HF_TOKEN"))
-
 device = "cuda"
 ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors")
 ckpt_IPA_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors-IP-Adapter-Plus")
@@ -28,16 +25,16 @@ vae = AutoencoderKL.from_pretrained(f"{ckpt_dir}/vae", revision=None).half().to(
 scheduler = EulerDiscreteScheduler.from_pretrained(f"{ckpt_dir}/scheduler")
 unet_t2i = UNet2DConditionModel.from_pretrained(f"{ckpt_dir}/unet", revision=None).half().to(device)
 unet_i2i = unet_2d_condition.UNet2DConditionModel.from_pretrained(f"{ckpt_dir}/unet", revision=None).half().to(device)
-image_encoder = CLIPVisionModelWithProjection.from_pretrained(f'{ckpt_IPA_dir}/image_encoder', ignore_mismatched_sizes=True).to(dtype=torch.float16, device=device)
+image_encoder = CLIPVisionModelWithProjection.from_pretrained(f'{ckpt_IPA_dir}/image_encoder',ignore_mismatched_sizes=True).to(dtype=torch.float16, device=device)
 ip_img_size = 336
 clip_image_processor = CLIPImageProcessor(size=ip_img_size, crop_size=ip_img_size)
 
 pipe_t2i = pipeline_stable_diffusion_xl_chatglm_256.StableDiffusionXLPipeline(
     vae=vae,
-    text_encoder=text_encoder,
-    tokenizer=tokenizer,
-    unet=unet_t2i,
-    scheduler=scheduler,
+    text_encoder=text_encoder, 
+    tokenizer=tokenizer, 
+    unet=unet_t2i, 
+    scheduler=scheduler, 
     force_zeros_for_empty_prompt=False
 ).to(device)
 
@@ -60,34 +57,16 @@ pipe_i2i.load_ip_adapter(f'{ckpt_IPA_dir}' , subfolder="", weight_name=["ip_adap
 MAX_SEED = np.iinfo(np.int32).max
 MAX_IMAGE_SIZE = 1024
 
-def call_api(content, system_message, max_tokens=1000, temperature=0.7, top_p=0.95):
-    messages = [{"role": "system", "content": system_message}, {"role": "user", "content": content}]
-    response = client.chat_completion(messages, max_tokens=max_tokens, temperature=temperature, top_p=top_p)
-    return response.choices[0].message['content']
-
-def generate_prompt(korean_prompt):
-    system_message = """
-    Given the following description in Korean, 
-    translate and generate a concise English prompt suitable for a Stable Diffusion model. 
-    The prompt should be focused, descriptive, 
-    and contain specific keywords or phrases that will help guide the image generation process. 
-    Use simple and descriptive language, avoiding unnecessary words. 
-    Ensure the output is in English and follows the format typically used in Stable Diffusion prompts. 
-    The description is: [Insert Korean description here]
-    """
-    optimized_prompt = call_api(korean_prompt, system_message)
-    return optimized_prompt  # 최적화된 프롬프트 반환
-
 @spaces.GPU
-def infer(prompt,
-          ip_adapter_image = None,
-          ip_adapter_scale = 0.5,
-          negative_prompt = "",
-          seed = 0,
-          randomize_seed = False,
-          width = 1024,
-          height = 1024,
-          guidance_scale = 5.0,
+def infer(prompt, 
+          ip_adapter_image = None, 
+          ip_adapter_scale = 0.5, 
+          negative_prompt = "", 
+          seed = 0, 
+          randomize_seed = False, 
+          width = 1024, 
+          height = 1024, 
+          guidance_scale = 5.0, 
           num_inference_steps = 25
           ):
     if randomize_seed:
@@ -97,14 +76,14 @@ def infer(prompt,
     if ip_adapter_image is None:
         pipe_t2i.to(device)
         image = pipe_t2i(
-            prompt = prompt,
+            prompt = prompt, 
             negative_prompt = negative_prompt,
-            guidance_scale = guidance_scale,
-            num_inference_steps = num_inference_steps,
-            width = width,
+            guidance_scale = guidance_scale, 
+            num_inference_steps = num_inference_steps, 
+            width = width, 
             height = height,
             generator = generator
-        ).images[0]
+        ).images[0] 
         image.save("generated_image.jpg")  # 파일 확장자를 .jpg로 변경
         return image, "generated_image.jpg"
     else:
@@ -115,10 +94,10 @@ def infer(prompt,
         image = pipe_i2i(
                 prompt=prompt,
                 ip_adapter_image=[ip_adapter_image],
-                negative_prompt=negative_prompt,
+                negative_prompt=negative_prompt, 
                 height=height,
                 width=width,
-                num_inference_steps=num_inference_steps,
+                num_inference_steps=num_inference_steps, 
                 guidance_scale=guidance_scale,
                 num_images_per_prompt=1,
                 generator=generator
@@ -126,10 +105,26 @@ def infer(prompt,
         image.save("generated_image.jpg")  # 파일 확장자를 .jpg로 변경
         return image, "generated_image.jpg"
 
-# 사진 설명 기능 추가를 위한 참조 코드 통합
+css="""
+#col-left {
+    margin: 0 auto;
+    max-width: 600px;
+}
+#col-right {
+    margin: 0 auto;
+    max-width: 750px;
+}
+#output {
+    height: 500px; 
+    overflow: auto; 
+    border: 1px solid #ccc; 
+}
+"""
+
+# 추가 코드 통합
 from transformers import AutoProcessor, AutoModelForCausalLM
 
-model = AutoModelForCausalLM.from_pretrained('gokaygokay/Florence-2-SD3-Captioner', trust_remote_code=True).to("cuda").eval()
+model = AutoModelForCausalLM.from_pretrained('gokaygokay/Florence-2-SD3-Captioner', trust_remote_code=True).eval()
 processor = AutoProcessor.from_pretrained('gokaygokay/Florence-2-SD3-Captioner', trust_remote_code=True)
 
 def modify_caption(caption: str) -> str:
@@ -145,7 +140,7 @@ def modify_caption(caption: str) -> str:
     return modified_caption if modified_caption != caption else caption
     
 @spaces.GPU
-def describe_image(image):
+def run_example(image):
     image = Image.fromarray(image)
     task_prompt = "<DESCRIPTION>"
     prompt = task_prompt + "Describe this image in great detail."
@@ -153,7 +148,7 @@ def describe_image(image):
     if image.mode != "RGB":
         image = image.convert("RGB")
 
-    inputs = processor(text=prompt, images=image, return_tensors="pt").to("cuda")
+    inputs = processor(text=prompt, images=image, return_tensors="pt")
     generated_ids = model.generate(
         input_ids=inputs["input_ids"],
         pixel_values=inputs["pixel_values"],
@@ -164,124 +159,96 @@ def describe_image(image):
     parsed_answer = processor.post_process_generation(generated_text, task=task_prompt, image_size=(image.width, image.height))
     return modify_caption(parsed_answer["<DESCRIPTION>"])
 
-css="""
-#col-left {
-    margin: 0 auto;
-    max-width: 600px;
-}
-#col-right {
-    margin: 0 auto;
-    max-width: 750px;
-}
-"""
-
 with gr.Blocks(css=css) as Kolors:
-    with gr.Row():
-        with gr.Column(elem_id="col-left"):
-            with gr.Row():
-                korean_prompt = gr.Textbox(
-                    label="한국어 프롬프트 입력",
-                    placeholder="한국어로 원하는 프롬프트를 입력하세요",
-                    lines=2
-                )
-            with gr.Row():
-                generate_prompt_button = gr.Button("Generate Prompt")
-            with gr.Row():
-                optimized_prompt = gr.Textbox(
-                    label="최적화된 프롬프트 생성",
-                    placeholder=" ",
-                    lines=2,
-                    interactive=False
-                )
-            with gr.Row():
-                generated_prompt = gr.Textbox(
-                    label="프롬프트 입력",
-                    placeholder="이미지 생성에 사용할 프롬프트를 입력하세요",
-                    lines=2
-                )
-            with gr.Row():
-                ip_adapter_image = gr.Image(label="Image Prompt (optional)", type="pil")
-            with gr.Row(visible=False):  # Advanced Settings 숨김
-                negative_prompt = gr.Textbox(
-                    label="Negative prompt",
-                    placeholder="Enter a negative prompt",
-                    visible=True,
-                )
-                seed = gr.Slider(
-                    label="Seed",
-                    minimum=0,
-                    maximum=MAX_SEED,
-                    step=1,
-                    value=0,
-                )
-                randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+    with gr.Tab("Image Generation"):
+        with gr.Row():
+            with gr.Column(elem_id="col-left"):
                 with gr.Row():
-                    width = gr.Slider(
-                        label="Width",
-                        minimum=256,
-                        maximum=MAX_IMAGE_SIZE,
-                        step=32,
-                        value=1024,
-                    )
-                    height = gr.Slider(
-                        label="Height",
-                        minimum=256,
-                        maximum=MAX_IMAGE_SIZE,
-                        step=32,
-                        value=1024,
+                    generated_prompt = gr.Textbox(
+                        label="프롬프트 입력",
+                        placeholder="이미지 생성에 사용할 프롬프트를 입력하세요",
+                        lines=2
                     )
                 with gr.Row():
-                    guidance_scale = gr.Slider(
-                        label="Guidance scale",
-                        minimum=0.0,
-                        maximum=10.0,
-                        step=0.1,
-                        value=5.0,
+                    ip_adapter_image = gr.Image(label="Image Prompt (optional)", type="pil")
+                with gr.Row(visible=False):  # Advanced Settings 숨김
+                    negative_prompt = gr.Textbox(
+                        label="Negative prompt",
+                        placeholder="Enter a negative prompt",
+                        visible=True,
                     )
-                    num_inference_steps = gr.Slider(
-                        label="Number of inference steps",
-                        minimum=10,
-                        maximum=50,
+                    seed = gr.Slider(
+                        label="Seed",
+                        minimum=0,
+                        maximum=MAX_SEED,
                         step=1,
-                        value=25,
+                        value=0,
                     )
+                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+                    with gr.Row():
+                        width = gr.Slider(
+                            label="Width",
+                            minimum=256,
+                            maximum=MAX_IMAGE_SIZE,
+                            step=32,
+                            value=1024,
+                        )
+                        height = gr.Slider(
+                            label="Height",
+                            minimum=256,
+                            maximum=MAX_IMAGE_SIZE,
+                            step=32,
+                            value=1024,
+                        )
+                    with gr.Row():
+                        guidance_scale = gr.Slider(
+                            label="Guidance scale",
+                            minimum=0.0,
+                            maximum=10.0,
+                            step=0.1,
+                            value=5.0,
+                        )
+                        num_inference_steps = gr.Slider(
+                            label="Number of inference steps",
+                            minimum=10,
+                            maximum=50,
+                            step=1,
+                            value=25,
+                        )
+                    with gr.Row():
+                        ip_adapter_scale = gr.Slider(
+                            label="Image influence scale",
+                            info="Use 1 for creating variations",
+                            minimum=0.0,
+                            maximum=1.0,
+                            step=0.05,
+                            value=0.5,
+                        )
                 with gr.Row():
-                    ip_adapter_scale = gr.Slider(
-                        label="Image influence scale",
-                        info="Use 1 for creating variations",
-                        minimum=0.0,
-                        maximum=1.0,
-                        step=0.05,
-                        value=0.5,
-                    )
-            with gr.Row():
-                run_button = gr.Button("Generate Image")
-            
-        with gr.Column(elem_id="col-right"):
-            result = gr.Image(label="Result", show_label=False)
-            download_button = gr.File(label="Download Image")
-            image_description = gr.Textbox(label="Image Description", placeholder="이미지 분석 결과가 여기에 표시됩니다.", interactive=False)
-            analyze_button = gr.Button("Analyze Image")
-
-    # 최적화된 프롬프트 생성 및 결과 표시
-    generate_prompt_button.click(
-        fn=generate_prompt,
-        inputs=[korean_prompt],
-        outputs=[optimized_prompt]
-    )
-    
-    # 이미지 생성 및 다운로드 파일 경로 설정
-    run_button.click(
-        fn=infer,
-        inputs=[generated_prompt, ip_adapter_image, ip_adapter_scale, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
-        outputs=[result, download_button]
-    )
+                    run_button = gr.Button("Generate Image")
+                
+            with gr.Column(elem_id="col-right"):
+                result = gr.Image(label="Result", show_label=False)
+                download_button = gr.File(label="Download Image")
+
+        # 이미지 생성 및 다운로드 파일 경로 설정
+        run_button.click(
+            fn=infer,
+            inputs=[generated_prompt, ip_adapter_image, ip_adapter_scale, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
+            outputs=[result, download_button]
+        )
     
-    # 이미지 설명 생성
-    analyze_button.click(
-        fn=describe_image,
-        inputs=[ip_adapter_image],
-        outputs=[image_description]
-    )
+    with gr.Tab("Florence-2 SD3 Prompts"):
+        gr.Markdown("# [Florence-2 SD3 Long Captioner](https://huggingface.co/gokaygokay/Florence-2-SD3-Captioner/)")
+        gr.Markdown("[Florence-2 Base](https://huggingface.co/microsoft/Florence-2-base-ft) fine-tuned on Long SD3 Prompt and Image pairs. Check above link for datasets that are used for fine-tuning.")
+        with gr.Row():
+            with gr.Column():
+                input_img = gr.Image(label="Input Picture")
+                submit_btn = gr.Button(value="Submit")
+            with gr.Column():
+                output_text = gr.Textbox(label="Output Text")
+
+
+        submit_btn.click(run_example, [input_img], [output_text])
 
 Kolors.queue().launch(debug=True)