add rembg support

app.py

@@ -1,43 +1,142 @@
 import gradio as gr
 import numpy as np
 import random
+from PIL import Image
+from rembg import remove
 
 # import spaces #[uncomment to use ZeroGPU]
-from diffusers import DiffusionPipeline
+from peft import PeftModel
+from diffusers import DiffusionPipeline, StableDiffusionPipeline, ControlNetModel, StableDiffusionControlNetPipeline
+from diffusers.utils import load_image
 import torch
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use
+model_repo_id = "CompVis/stable-diffusion-v1-4"  # Replace to the model you would like to use
 
-if torch.cuda.is_available():
-    torch_dtype = torch.float16
-else:
-    torch_dtype = torch.float32
+torch_dtype = torch.float16
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
+pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
 pipe = pipe.to(device)
+# pipe.unet = PeftModel.from_pretrained(pipe.unet, "alexanz/SD14_lora_pusheen")
+pipe.safety_checker = None
+pipe.requires_safety_checker = False
 
 MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
+MAX_IMAGE_SIZE = 512
 
 
 # @spaces.GPU #[uncomment to use ZeroGPU]
+def load_model(model_id, lora_strength, use_controlnet=False, control_mode="edge_detection", use_ip_adapter=False, control_strength_ip=0.0):
+    global pipe
+    if pipe is not None:
+        del pipe
+        torch.cuda.empty_cache()
+    try:
+        if control_mode == "edge_detection" and (model_id == "CompVis/stable-diffusion-v1-4" or model_id == "alexanz/SD14_lora_pusheen"):
+            controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch_dtype)
+        elif control_mode == "pose_estimation"and (model_id == "CompVis/stable-diffusion-v1-4" or model_id == "alexanz/SD14_lora_pusheen"):
+            controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-openpose", torch_dtype=torch_dtype)
+        if control_mode == "edge_detection" and (model_id == "alexanz/SD15_lora_pusheen"):
+            controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_canny", torch_dtype=torch_dtype)
+        elif control_mode == "pose_estimation"and (model_id == "alexanz/SD15_lora_pusheen"):
+            controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_openpose", torch_dtype=torch_dtype)
+    
+        if model_id == "CompVis/stable-diffusion-v1-4":
+            if use_controlnet:
+                pipe = StableDiffusionControlNetPipeline.from_pretrained(
+                    model_id,
+                    safety_checker=None,
+                    controlnet=controlnet,
+                    torch_dtype=torch_dtype
+                )
+            else:
+                pipe = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch_dtype)
+
+        elif model_id == "alexanz/SD14_lora_pusheen":
+            if use_controlnet:
+                pipe = StableDiffusionControlNetPipeline.from_pretrained(
+                    "CompVis/stable-diffusion-v1-4",
+                    safety_checker=None,
+                    controlnet=controlnet,
+                    torch_dtype=torch_dtype
+                )
+                pipe.unet = PeftModel.from_pretrained(pipe.unet, model_id, scaling=lora_strength, torch_dtype=torch_dtype)
+            else:
+                pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch_dtype)
+                pipe.unet = PeftModel.from_pretrained(pipe.unet, model_id, scaling=lora_strength)
+
+        elif model_id == "alexanz/SD15_lora_pusheen":
+            if use_controlnet:
+                pipe = StableDiffusionControlNetPipeline.from_pretrained(
+                    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+                    safety_checker=None,
+                    controlnet=controlnet,
+                    torch_dtype=torch_dtype
+                )
+                pipe.unet = PeftModel.from_pretrained(pipe.unet, model_id, scaling=lora_strength, torch_dtype=torch_dtype)
+            else:
+                pipe = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch_dtype)
+                pipe.unet = PeftModel.from_pretrained(pipe.unet, model_id, scaling=lora_strength)
+
+        if use_ip_adapter:
+            pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
+            pipe.set_ip_adapter_scale(control_strength_ip)
+
+        pipe = pipe.to(device)
+        pipe.safety_checker = None
+        pipe.requires_safety_checker = False
+        pipe.enable_model_cpu_offload()
+        return f"Model {model_id} loaded with ControlNet: {use_controlnet}, mode: {control_mode}"
+    except Exception as e:
+        return f"Error: {str(e)}"
+
+
 def infer(
-    prompt,
-    negative_prompt,
-    seed,
-    randomize_seed,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
-    progress=gr.Progress(track_tqdm=True),
+        prompt,
+        negative_prompt,
+        seed,
+        randomize_seed,
+        width,
+        height,
+        lora_strength,
+        guidance_scale,
+        num_inference_steps,
+        use_controlnet,
+        control_image_cont,
+        control_strength_cont,
+        model_dropdown,
+        control_mode,
+        use_ip_adapter,
+        control_strength_ip,
+        control_image_ip,
+        use_rmbg,
+        progress=gr.Progress(track_tqdm=True),
 ):
+    load_status = load_model(
+        model_dropdown,
+        lora_strength,
+        use_controlnet,
+        control_mode,
+        use_ip_adapter,
+        control_strength_ip
+    )
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
 
     generator = torch.Generator().manual_seed(seed)
 
+    if use_controlnet and control_image_cont is None:
+        return None, seed, "⚠️ ControlNet need control_image!"
+    
+    if use_ip_adapter and control_image_ip is None:
+        return None, seed, "⚠️ IP-adapter need control_image!"
+
+    if use_controlnet:
+        control_image_cont= Image.fromarray(control_image_cont)
+        control_strength_cont = float(control_strength_cont)
+    if use_ip_adapter:
+        control_image_ip = Image.fromarray(control_image_ip)
+
     image = pipe(
         prompt=prompt,
         negative_prompt=negative_prompt,
@@ -46,15 +145,21 @@ def infer(
         width=width,
         height=height,
         generator=generator,
+        image=control_image_cont if use_controlnet else None,
+        controlnet_conditioning_scale=control_strength_cont if use_controlnet else None,
+        ip_adapter_image=control_image_ip if use_ip_adapter else None
     ).images[0]
 
-    return image, seed
+    if use_rmbg:
+        image = remove(image)
+
+    return image, seed, "Model ready"
 
 
 examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
+    "Sticker of Pusheen. Cartoon image of a gray cat with cap of tea.",
+    "Sticker of Pusheen. Gray cat holding a guitar, sitting under a disco ball, with colorful lights and a happy face.",
+    "Sticker of Pusheen. A cute cartoon fluffy cat.",
 ]
 
 css = """
@@ -67,6 +172,10 @@ css = """
 with gr.Blocks(css=css) as demo:
     with gr.Column(elem_id="col-container"):
         gr.Markdown(" # Text-to-Image Gradio Template")
+        model_dropdown = gr.Dropdown(label="Model ID",
+                                     choices=["alexanz/SD14_lora_pusheen", "CompVis/stable-diffusion-v1-4", "alexanz/SD15_lora_pusheen"],
+                                     value="CompVis/stable-diffusion-v1-4")
+        model_status = gr.Textbox(label="Model Status", interactive=False)
 
         with gr.Row():
             prompt = gr.Text(
@@ -86,7 +195,14 @@ with gr.Blocks(css=css) as demo:
                 label="Negative prompt",
                 max_lines=1,
                 placeholder="Enter a negative prompt",
-                visible=False,
+            )
+
+            lora_strength = gr.Slider(
+                label="Lora strength",
+                minimum=0.0,
+                maximum=1.0,
+                step=0.1,
+                value=1.0,
             )
 
             seed = gr.Slider(
@@ -105,7 +221,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=256,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024,  # Replace with defaults that work for your model
+                    value=512,  # Replace with defaults that work for your model
                 )
 
                 height = gr.Slider(
@@ -113,7 +229,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=256,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024,  # Replace with defaults that work for your model
+                    value=512,  # Replace with defaults that work for your model
                 )
 
             with gr.Row():
@@ -122,7 +238,7 @@ with gr.Blocks(css=css) as demo:
                     minimum=0.0,
                     maximum=10.0,
                     step=0.1,
-                    value=0.0,  # Replace with defaults that work for your model
+                    value=7.5,  # Replace with defaults that work for your model
                 )
 
                 num_inference_steps = gr.Slider(
@@ -130,10 +246,40 @@ with gr.Blocks(css=css) as demo:
                     minimum=1,
                     maximum=50,
                     step=1,
-                    value=2,  # Replace with defaults that work for your model
+                    value=20,  # Replace with defaults that work for your model
                 )
 
+        use_controlnet = gr.Checkbox(label="Use ControlNet", value=False)
+        with gr.Accordion("ControlNet Settings", open=True, visible=False) as controlnet_settings:
+            control_mode = gr.Dropdown(
+                label="ControlNet Mode",
+                choices=["edge_detection", "pose_estimation"],
+                value="edge_detection"
+            )
+            control_strength_cont = gr.Slider(
+                label="Control Strength",
+                minimum=0.0,
+                maximum=2.0,
+                step=0.1,
+                value=1.0
+            )
+            control_image_cont = gr.Image(label="Control Image", type="numpy")
+
+        use_ip_adapter = gr.Checkbox(label="Use IP-adapter", value=False)
+        with gr.Accordion("IP-adapter Settings", open=True, visible=False) as ip_adapter_settings:
+            control_strength_ip = gr.Slider(
+                label="Control Strength",
+                minimum=0.0,
+                maximum=2.0,
+                step=0.1,
+                value=1.0
+            )
+            control_image_ip = gr.Image(label="Control Image (IP-adapter)", type="numpy")
+
+        use_rmbg = gr.Checkbox(label="Delete background?", value=False)
+
         gr.Examples(examples=examples, inputs=[prompt])
+
     gr.on(
         triggers=[run_button.click, prompt.submit],
         fn=infer,
@@ -144,11 +290,38 @@ with gr.Blocks(css=css) as demo:
             randomize_seed,
             width,
             height,
+            lora_strength,
             guidance_scale,
             num_inference_steps,
+            use_controlnet,
+            control_image_cont,
+            control_strength_cont,
+            model_dropdown,
+            control_mode,
+            use_ip_adapter,
+            control_strength_ip,
+            control_image_ip,
+            use_rmbg
         ],
-        outputs=[result, seed],
+        outputs=[result, seed, model_status],
+    )
+
+    use_controlnet.change(
+        fn=lambda x: gr.update(visible=x, value=None),
+        inputs=[use_controlnet],
+        outputs=[controlnet_settings]
+    )
+
+    use_ip_adapter.change(
+        fn=lambda x: gr.update(visible=x, value=None),
+        inputs=[use_ip_adapter],
+        outputs=[ip_adapter_settings]
+    )
+
+    use_rmbg.change(
+        fn=lambda x: gr.update(visible=x, value=None),
+        inputs=[use_rmbg]
     )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()