Migrate to HF Inference API

app.py

@@ -1,55 +1,20 @@
-import torch
-import spaces
 import gradio as gr
-from diffusers import FluxPipeline, FluxImg2ImgPipeline
-import os
+from huggingface_hub import InferenceClient
 
-print("Loading FLUX.1 Models (CPU)...")
-# Load the base FLUX model
-# We load the pipelines on CPU so ZeroGPU doesn't crash during startup.
-pipe_t2i = FluxPipeline.from_pretrained(
-    "black-forest-labs/FLUX.1-schnell",
-    torch_dtype=torch.bfloat16,
-    token=os.environ.get("HF_TOKEN")
-)
-
-pipe_edit = FluxImg2ImgPipeline.from_pipe(pipe_t2i)
-print("Models loaded successfully!")
-
-@spaces.GPU
-def generate_or_edit(prompt, input_image, denoising_strength, num_inference_steps, seed, randomize_seed, progress=gr.Progress(track_tqdm=True)):
-    # Move pipelines to GPU inside the ZeroGPU decorated function
-    pipe_t2i.to("cuda")
-    pipe_edit.to("cuda")
-    
-    if randomize_seed:
-        seed = torch.randint(0, 2**32 - 1, (1,)).item()
-    generator = torch.Generator("cuda").manual_seed(int(seed))
-    
+def generate_image(prompt, hf_token, progress=gr.Progress(track_tqdm=True)):
+    if not hf_token:
+        raise gr.Error("Please enter your Hugging Face API Token!")
     if not prompt:
         raise gr.Error("Please enter a prompt!")
-    
-    if input_image is not None:
-        # Edit mode (Img2Img)
-        input_image = input_image.convert("RGB")
-        image = pipe_edit(
-            prompt=prompt,
-            image=input_image,
-            strength=denoising_strength,
-            num_inference_steps=int(num_inference_steps),
-            guidance_scale=0.0, # FLUX.1-schnell uses 0 guidance scale
-            generator=generator,
-        ).images[0]
-    else:
-        # Generate mode (T2I)
-        image = pipe_t2i(
-            prompt=prompt,
-            num_inference_steps=int(num_inference_steps),
-            guidance_scale=0.0, # FLUX.1-schnell uses 0 guidance scale
-            generator=generator,
-        ).images[0]
         
-    return image, seed
+    client = InferenceClient(token=hf_token.strip())
+    
+    image = client.text_to_image(
+        prompt, 
+        model="black-forest-labs/FLUX.1-schnell"
+    )
+    
+    return image
 
 # UI
 custom_theme = gr.themes.Soft(
@@ -62,66 +27,40 @@ with gr.Blocks() as demo:
     gr.Markdown(
         """
         # ⚡ FLUX.1 Image Studio (Grok Quality)
-        Generate state-of-the-art images from scratch, or edit existing ones using the FLUX.1-schnell model.
+        Generate ultra-fast images from text using the real-time FLUX.1-schnell model via Hugging Face Serverless API.
+        *No local GPU Required! Generates in the cloud.*
         """
     )
     
     with gr.Row():
         with gr.Column(scale=1):
+            hf_token = gr.Textbox(
+                label="🔑 Hugging Face Access Token",
+                placeholder="hf_...",
+                type="password",
+                info="Paste your Hugging Face Token here"
+            )
             prompt = gr.Textbox(
                 label="✨ Prompt", 
                 lines=3, 
-                placeholder="e.g. A cyberpunk cat...", 
+                placeholder="e.g. A futuristic cyberpunk city at night...", 
                 autofocus=True
             )
-            input_image = gr.Image(
-                label="🖼️ Input Image (Optional - For editing)", 
-                type="pil"
-            )
             
-            with gr.Accordion("⚙️ Advanced Settings", open=False):
-                denoising_strength = gr.Slider(
-                    minimum=0.0, 
-                    maximum=1.0, 
-                    value=0.5, 
-                    step=0.05, 
-                    label="Denoising Strength (Editing Only)", 
-                    info="Lower = keeps more of original image. Higher = completely changes image to match prompt."
-                )
-                num_inference_steps = gr.Slider(
-                    minimum=1, 
-                    maximum=12, 
-                    value=4, 
-                    step=1, 
-                    label="Inference Steps", 
-                    info="FLUX.1-schnell is optimized for 4 steps."
-                )
-                
-                with gr.Row():
-                    randomize_seed = gr.Checkbox(label="🎲 Random Seed", value=True)
-                    seed = gr.Number(label="Seed", value=42, precision=0, visible=False)
-                
-                randomize_seed.change(
-                    lambda r: gr.Number(visible=not r), 
-                    inputs=[randomize_seed], 
-                    outputs=[seed]
-                )
-                
-            generate_btn = gr.Button("🚀 Generate / Edit Image", variant="primary", size="lg")
+            generate_btn = gr.Button("🎨 Generate Image", variant="primary", size="lg")
             
         with gr.Column(scale=1):
             output_image = gr.Image(label="Result", type="pil", interactive=False)
-            used_seed = gr.Number(label="Seed Used", interactive=False)
 
     generate_btn.click(
-        fn=generate_or_edit,
-        inputs=[prompt, input_image, denoising_strength, num_inference_steps, seed, randomize_seed],
-        outputs=[output_image, used_seed]
+        fn=generate_image,
+        inputs=[prompt, hf_token],
+        outputs=[output_image]
     )
     prompt.submit(
-        fn=generate_or_edit,
-        inputs=[prompt, input_image, denoising_strength, num_inference_steps, seed, randomize_seed],
-        outputs=[output_image, used_seed]
+        fn=generate_image,
+        inputs=[prompt, hf_token],
+        outputs=[output_image]
     )
 
 if __name__ == "__main__":

requirements.txt

@@ -1,9 +1,2 @@
-gradio
-git+https://github.com/huggingface/diffusers
-transformers
-kernels
-gradio[mcp]
-sentencepiece
-protobuf
-spaces
-accelerate
+gradio==4.36.1
+huggingface_hub