Update custom_pipeline.py

custom_pipeline.py

@@ -1,151 +1,165 @@
-import torch
+import gradio as gr
 import numpy as np
-from diffusers import FluxPipeline, FlowMatchEulerDiscreteScheduler
-from typing import Any, Dict, List, Optional, Union
-from PIL import Image
-
-# Constants for shift calculation
-BASE_SEQ_LEN = 256
-MAX_SEQ_LEN = 4096
-BASE_SHIFT = 0.5
-MAX_SHIFT = 1.2
-
-# Helper functions
-def calculate_timestep_shift(image_seq_len: int) -> float:
-    """Calculates the timestep shift (mu) based on the image sequence length."""
-    m = (MAX_SHIFT - BASE_SHIFT) / (MAX_SEQ_LEN - BASE_SEQ_LEN)
-    b = BASE_SHIFT - m * BASE_SEQ_LEN
-    mu = image_seq_len * m + b
-    return mu
-
-def prepare_timesteps(
-    scheduler: FlowMatchEulerDiscreteScheduler,
-    num_inference_steps: Optional[int] = None,
-    device: Optional[Union[str, torch.device]] = None,
-    timesteps: Optional[List[int]] = None,
-    sigmas: Optional[List[float]] = None,
-    mu: Optional[float] = None,
-) -> (torch.Tensor, int):
-    """Prepares the timesteps for the diffusion process."""
-    if timesteps is not None and sigmas is not None:
-        raise ValueError("Only one of `timesteps` or `sigmas` can be passed.")
-
-    if timesteps is not None:
-        scheduler.set_timesteps(timesteps=timesteps, device=device)
-    elif sigmas is not None:
-        scheduler.set_timesteps(sigmas=sigmas, device=device)
-    else:
-        scheduler.set_timesteps(num_inference_steps, device=device, mu=mu)
-
-    timesteps = scheduler.timesteps
-    num_inference_steps = len(timesteps)
-    return timesteps, num_inference_steps
-
-# FLUX pipeline function
-class HighSpeedFluxPipeline(FluxPipeline):
-    """
-    Extends the FluxPipeline to yield intermediate images during the denoising process 
-    with progressively increasing resolution for faster generation.
-    """
-    @torch.inference_mode()
-    def generate_images(
-        self,
-        prompt: Union[str, List[str]] = None,
-        prompt_2: Optional[Union[str, List[str]]] = None,
-        height: Optional[int] = None,
-        width: Optional[int] = None,
-        num_inference_steps: int = 4,
-        timesteps: List[int] = None,
-        num_images_per_prompt: Optional[int] = 1,
-        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
-        latents: Optional[torch.FloatTensor] = None,
-        prompt_embeds: Optional[torch.FloatTensor] = None,
-        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
-        output_type: Optional[str] = "pil",
-        return_dict: bool = True,
-        max_sequence_length: int = 128,
-    ):
-        """Generates images and yields intermediate results during the denoising process."""
-        height = height or self.default_sample_size * self.vae_scale_factor
-        width = width or self.default_sample_size * self.vae_scale_factor
-
-        # 1. Check inputs
-        self.check_inputs(
-            prompt,
-            prompt_2,
-            height,
-            width,
-            prompt_embeds=prompt_embeds,
-            pooled_prompt_embeds=pooled_prompt_embeds,
-            max_sequence_length=max_sequence_length,
-        )
+import random
+import spaces
+import torch
+import time
+from diffusers import DiffusionPipeline
+from custom_pipeline import FLUXPipelineWithIntermediateOutputs
+
+# Constants
+MAX_SEED = np.iinfo(np.int32).max
+MAX_IMAGE_SIZE = 2048
+DEFAULT_WIDTH = 1024
+DEFAULT_HEIGHT = 1024
+DEFAULT_INFERENCE_STEPS = 1
+
+# Device and model setup
+dtype = torch.float16
+pipe = FLUXPipelineWithIntermediateOutputs.from_pretrained(
+    "black-forest-labs/FLUX.1-schnell", torch_dtype=dtype
+).to("cuda")
+torch.cuda.empty_cache()
+
+# Inference function
+@spaces.GPU(duration=25)
+def generate_image(prompt, seed=42, width=DEFAULT_WIDTH, height=DEFAULT_HEIGHT, randomize_seed=False, num_inference_steps=2, progress=gr.Progress(track_tqdm=True)):
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    generator = torch.Generator().manual_seed(int(float(seed)))
 
-        # 2. Define call parameters
-        batch_size = 1 if isinstance(prompt, str) else len(prompt)
-        device = self._execution_device
+    start_time = time.time()
 
-        prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
+    # Only generate the last image in the sequence
+    for img in pipe.generate_images(  
             prompt=prompt,
-            prompt_2=prompt_2,
-            prompt_embeds=prompt_embeds,
-            pooled_prompt_embeds=pooled_prompt_embeds,
-            device=device,
-            num_images_per_prompt=num_images_per_prompt,
-            max_sequence_length=max_sequence_length,
-        )
-        # 4. Prepare latent variables
-        num_channels_latents = self.transformer.config.in_channels // 4
-        latents, latent_image_ids = self.prepare_latents(
-            batch_size * num_images_per_prompt,
-            num_channels_latents,
-            height,
-            width,
-            prompt_embeds.dtype,
-            device,
-            generator,
-            latents,
-        )
-        # 5. Prepare timesteps
-        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
-        image_seq_len = latents.shape[1]
-        mu = calculate_timestep_shift(image_seq_len)
-        timesteps, num_inference_steps = prepare_timesteps(
-            self.scheduler,
-            num_inference_steps,
-            device,
-            timesteps,
-            sigmas,
-            mu=mu,
+            guidance_scale=0, # as Flux schnell is guidance free
+            num_inference_steps=num_inference_steps,
+            width=width,
+            height=height,
+            generator=generator
+        ): 
+        latency = f"Latency: {(time.time()-start_time):.2f} seconds"    
+        yield img, seed, latency
+
+# Example prompts
+examples = [
+    "a tiny astronaut hatching from an egg on the moon",
+    "a cute white cat holding a sign that says hello world",
+    "an anime illustration of a wiener schnitzel",
+    "Create mage of Modern house in minecraft style",
+    "Imagine steve jobs as Star Wars movie character",
+    "Lion",
+    "Photo of a young woman with long, wavy brown hair tied in a bun and glasses. She has a fair complexion and is wearing subtle makeup, emphasizing her eyes and lips. She is dressed in a black top. The background appears to be an urban setting with a building facade, and the sunlight casts a warm glow on her face.",
+]
+
+# --- Gradio UI ---
+with gr.Blocks() as demo:
+    with gr.Column(elem_id="app-container"):
+        gr.Markdown("# 🎨 Realtime FLUX Image Generator")
+        gr.Markdown("Generate stunning images in real-time with Modified Flux.Schnell pipeline.")
+        gr.Markdown("<span style='color: red;'>Note: Sometimes it stucks or stops generating images (I don't know why). In that situation just refresh the site.</span>")
+
+        with gr.Row():
+            with gr.Column(scale=3):
+                result = gr.Image(label="Generated Image", show_label=False, interactive=False)
+            with gr.Column(scale=1):
+                prompt = gr.Text(
+                    label="Prompt",
+                    placeholder="Describe the image you want to generate...",
+                    lines=3,
+                    show_label=False,
+                    container=False,
+                )
+                generateBtn = gr.Button("🖼️ Generate Image")
+                enhanceBtn = gr.Button("🚀 Enhance Image")
+
+                with gr.Column("Advanced Options"):
+                    with gr.Row():
+                        realtime = gr.Checkbox(label="Realtime Toggler", info="If TRUE then uses more GPU but create image in realtime.", value=False)
+                        latency = gr.Text(label="Latency")
+                    with gr.Row():
+                        seed = gr.Number(label="Seed", value=42)
+                        randomize_seed = gr.Checkbox(label="Randomize Seed", value=False)
+                    with gr.Row():
+                        width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=DEFAULT_WIDTH)
+                        height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=DEFAULT_HEIGHT)
+                        num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=4, step=1, value=DEFAULT_INFERENCE_STEPS)
+
+        with gr.Row():
+            gr.Markdown("### 🌟 Inspiration Gallery")
+        with gr.Row():
+            gr.Examples(
+                examples=examples,
+                fn=generate_image,
+                inputs=[prompt],
+                outputs=[result, seed, latency],
+                cache_examples="lazy" 
+            )
+
+    def enhance_image(*args):
+        gr.Info("Enhancing Image") # currently just runs optimized pipeline for 2 steps. Further implementations later.
+        return next(generate_image(*args))
+
+    enhanceBtn.click(
+        fn=enhance_image,
+        inputs=[prompt, seed, width, height],
+        outputs=[result, seed, latency],
+        show_progress="hidden",
+        api_name=False,
+        queue=False,
+        concurrency_limit=None
+    )
+
+    generateBtn.click(
+        fn=generate_image,
+        inputs=[prompt, seed, width, height, randomize_seed, num_inference_steps],
+        outputs=[result, seed, latency],
+        show_progress="full",
+        api_name="RealtimeFlux",
+        queue=False,
+        concurrency_limit=None
+    )
+
+    def update_ui(realtime_enabled):
+        return {
+            prompt: gr.update(interactive=True),
+            generateBtn: gr.update(visible=not realtime_enabled)
+        }
+
+    realtime.change(
+        fn=update_ui,
+        inputs=[realtime],
+        outputs=[prompt, generateBtn],
+        queue=False,
+        concurrency_limit=None
+    )
+
+    def realtime_generation(*args):
+        if args[0]:  # If realtime is enabled
+            return next(generate_image(*args[1:]))
+
+    prompt.submit(
+        fn=generate_image,
+        inputs=[prompt, seed, width, height, randomize_seed, num_inference_steps],
+        outputs=[result, seed, latency],
+        show_progress="full",
+        api_name=False,
+        queue=False,
+        concurrency_limit=None
+    )
+
+    for component in [prompt, width, height, num_inference_steps]:
+        component.input(
+            fn=realtime_generation,
+            inputs=[realtime, prompt, seed, width, height, randomize_seed, num_inference_steps],
+            outputs=[result, seed, latency],
+            show_progress="hidden",
+            api_name=False,
+            trigger_mode="always_last",
+            queue=False,
+            concurrency_limit=None
         )
-        self._num_timesteps = len(timesteps)
-
-        # 6. Denoising loop
-        for i, t in enumerate(timesteps):
-            
-            timestep = t.expand(latents.shape[0]).to(latents.dtype)
-
-            noise_pred = self.transformer(
-                hidden_states=latents,
-                timestep=timestep / 1000,
-                pooled_projections=pooled_prompt_embeds,
-                encoder_hidden_states=prompt_embeds,
-                txt_ids=text_ids,
-                img_ids=latent_image_ids,
-                return_dict=False,
-            )[0]
-
-            latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
-            torch.cuda.empty_cache()
-
-        # Final image
-        return self._decode_latents_to_image(latents, height, width, output_type)
-        self.maybe_free_model_hooks()
-        torch.cuda.empty_cache()
-
-    def _decode_latents_to_image(self, latents, height, width, output_type, vae=None):
-        """Decodes the given latents into an image."""
-        vae = vae or self.vae
-        latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
-        latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor
-        image = vae.decode(latents, return_dict=False)[0]
-        return self.image_processor.postprocess(image, output_type=output_type)[0]
+
+# Launch the app
+demo.launch()