seamless texture from flux space

app.py

@@ -1,148 +1,66 @@
 import gradio as gr
 import numpy as np
 import random
-import io
-from PIL import Image
-import tempfile
 import spaces
-from diffusers import DiffusionPipeline
 import torch
-from scipy.ndimage import sobel # For normal map generation
+from diffusers import  DiffusionPipeline, FlowMatchEulerDiscreteScheduler, AutoencoderTiny, AutoencoderKL
+from transformers import CLIPTextModel, CLIPTokenizer,T5EncoderModel, T5TokenizerFast
+from live_preview_helpers import calculate_shift, retrieve_timesteps, flux_pipe_call_that_returns_an_iterable_of_images
 
+dtype = torch.bfloat16
 device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "black-forest-labs/FLUX.1-dev"
 
-if torch.cuda.is_available():
-    torch_dtype = torch.bfloat16
-else:
-    torch_dtype = torch.float32
-
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
+taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device)
+good_vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae", torch_dtype=dtype).to(device)
+pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=dtype, vae=taef1).to(device)
+torch.cuda.empty_cache()
 
 MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
-
-def generate_normal_map_from_image(image: Image.Image, strength: float = 1.0) -> Image.Image:
-    """
-    Generates a normal map from a PIL Image using Sobel filters.
-    This is a simplified approach, not using an AI model for normal map generation.
-    It approximates normals from the intensity changes in the image.
-
-    Args:
-        image (PIL.Image.Image): The input image (e.g., the generated texture).
-        strength (float): Controls the intensity/depth of the normal map effect.
-                          Higher values make the bumps more pronounced.
-
-    Returns:
-        PIL.Image.Image: The generated normal map.
-    """
-    # Convert to grayscale
-    gray_image = image.convert("L")
-    pixels = np.array(gray_image, dtype=np.float32)
-
-    # Calculate gradients using Sobel filter
-    # dx (horizontal gradient)
-    dx = sobel(pixels, axis=1)
-    # dy (vertical gradient)
-    dy = sobel(pixels, axis=0)
-
-    # Create the normal vector components
-    # The Z component determines how "flat" or "bumpy" the surface appears.
-    # A higher strength means the Z component is relatively smaller, making the surface look bumpier.
-    # A lower strength means the Z component is relatively larger, making the surface look flatter.
-    # We scale dx and dy by strength, and keep Z as 1.0 (or a constant that will be normalized).
-    # Then normalize the vector (dx_scaled, dy_scaled, 1.0)
-
-    # Scale gradients by strength
-    normal_x = -dx * strength
-    normal_y = -dy * strength
-    normal_z = np.ones_like(pixels) * 255.0 # A large constant for Z before normalization
-
-    # Normalize the vectors (x, y, z)
-    magnitude = np.sqrt(normal_x**2 + normal_y**2 + normal_z**2)
-    # Avoid division by zero
-    magnitude[magnitude == 0] = 1e-6 # Small epsilon to prevent division by zero
-
-    normal_x_norm = normal_x / magnitude
-    normal_y_norm = normal_y / magnitude
-    normal_z_norm = normal_z / magnitude
-
-    # Map normalized components (-1 to 1) to 0-255 for RGB channels
-    # R: (normal_x_norm + 1) / 2 * 255
-    # G: (normal_y_norm + 1) / 2 * 255
-    # B: (normal_z_norm + 1) / 2 * 255
+MAX_IMAGE_SIZE = 2048
 
-    # Common convention for normal maps: R=X, G=Y, B=Z.
-    # Some engines (e.g., OpenGL) might require Y to be inverted.
-    # For general purpose, let's keep it standard.
-    r_channel = ((normal_x_norm + 1) / 2 * 255).astype(np.uint8)
-    g_channel = ((normal_y_norm + 1) / 2 * 255).astype(np.uint8)
-    b_channel = ((normal_z_norm + 1) / 2 * 255).astype(np.uint8)
+pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe)
 
-    normal_map_array = np.stack([r_channel, g_channel, b_channel], axis=-1)
-
-    return Image.fromarray(normal_map_array, 'RGB')
-
-@spaces.GPU(duration=65)
-def infer(
-    prompt,
-    negative_prompt="",
-    seed=42,
-    randomize_seed=False,
-    width=1024,
-    height=1024,
-    guidance_scale=4.5,
-    num_inference_steps=40,
-    normal_map_strength=1.0,
-    progress=gr.Progress(track_tqdm=True),
-):
+@spaces.GPU(duration=75)
+def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, guidance_scale=3.5, num_inference_steps=28, progress=gr.Progress(track_tqdm=True)):
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-
     generator = torch.Generator().manual_seed(seed)
-
-    image = pipe(
-        prompt=f"smlstxtr, {prompt}, seamless texture",
-        negative_prompt=negative_prompt,
-        guidance_scale=guidance_scale,
-        num_inference_steps=num_inference_steps,
-        width=width,
-        height=height,
-        generator=generator,
-    ).images[0]
-
-    # Generate the normal map from the created image
-    normal_map_image = generate_normal_map_from_image(image, strength=normal_map_strength)
     
-    # Create temporary files with proper cleanup
-    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as normal_tmp_file:
-        normal_map_image.save(normal_tmp_file.name, format="JPEG")
-        normal_map_path = normal_tmp_file.name
-
-    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as tmp_file:
-        image.save(tmp_file.name, format="JPEG")
-        image_path = tmp_file.name
-
-    return image_path, normal_map_path
-
-
+    for img in pipe.flux_pipe_call_that_returns_an_iterable_of_images(
+            prompt= f"smlstxtr, {prompt} , seamless texture",
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            width=width,
+            height=height,
+            generator=generator,
+            output_type="pil",
+            good_vae=good_vae,
+        ):
+            yield img
+    
 examples = [
-        "A seamless grassy ground texture, 4k, realistic, for game environment",
-    "Rough concrete wall texture, weathered, high detail",
-    "Smooth metallic surface with subtle scratches, sci-fi style",
+    "A rusty metal surface with peeling paint and corrosion spots",
+    " A dense cluster of white snowflakes on dark blue background",
+    "A collection of colorful beach pebbles packed tightly together",
 ]
 
-css = """
+css="""
 #col-container {
     margin: 0 auto;
-    max-width: 640px;
+    max-width: 520px;
 }
 """
 
 with gr.Blocks(css=css) as demo:
+    
     with gr.Column(elem_id="col-container"):
+        gr.Markdown(f"""# FLUX.1 [dev]
+12B param rectified flow transformer guidance-distilled from [FLUX.1 [pro]](https://blackforestlabs.ai/)  
+[[non-commercial license](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md)] [[blog](https://blackforestlabs.ai/announcing-black-forest-labs/)] [[model](https://huggingface.co/black-forest-labs/FLUX.1-dev)]
+        """)
+        
         with gr.Row():
+            
             prompt = gr.Text(
                 label="Prompt",
                 show_label=False,
@@ -150,21 +68,13 @@ with gr.Blocks(css=css) as demo:
                 placeholder="Enter your prompt",
                 container=False,
             )
-
-            run_button = gr.Button("Run", scale=0, variant="primary")
-
-        with gr.Row():
-            result_image = gr.Image(label="Generated Texture", type="filepath")
-            normal_map_result = gr.Image(label="Normal Map", type="filepath")
-
+            
+            run_button = gr.Button("Run", scale=0)
+        
+        result = gr.Image(label="Result", show_label=False)
+        
         with gr.Accordion("Advanced Settings", open=False):
-            negative_prompt = gr.Text(
-                label="Negative prompt",
-                max_lines=1,
-                placeholder="Enter a negative prompt",
-                visible=False,
-            )
-
+            
             seed = gr.Slider(
                 label="Seed",
                 minimum=0,
@@ -172,69 +82,58 @@ with gr.Blocks(css=css) as demo:
                 step=1,
                 value=0,
             )
-
+            
             randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
-
+            
             with gr.Row():
+                
                 width = gr.Slider(
                     label="Width",
-                    minimum=512,
+                    minimum=256,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024, 
+                    value=1024,
                 )
-
+                
                 height = gr.Slider(
                     label="Height",
-                    minimum=512,
+                    minimum=256,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
                     value=1024,
                 )
-
+            
             with gr.Row():
+
                 guidance_scale = gr.Slider(
-                    label="Guidance scale",
-                    minimum=0.0,
-                    maximum=7.5,
+                    label="Guidance Scale",
+                    minimum=1,
+                    maximum=15,
                     step=0.1,
-                    value=4.5,
+                    value=3.5,
                 )
-
+  
                 num_inference_steps = gr.Slider(
                     label="Number of inference steps",
                     minimum=1,
                     maximum=50,
                     step=1,
-                    value=40, 
+                    value=28,
                 )
-                # New slider for normal map strength
-            normal_map_strength = gr.Slider(
-                label="Normal Map Strength",
-                minimum=0.1,
-                maximum=5.0,
-                step=0.1,
-                value=1.0,
-                info="Adjusts the perceived depth/bumpiness of the normal map."
-            )
+        
+        gr.Examples(
+            examples = examples,
+            fn = infer,
+            inputs = [prompt],
+            outputs = [result],
+            cache_examples="lazy"
+        )
 
-        gr.Examples(examples=examples, inputs=[prompt], outputs=[result_image, normal_map_result], fn=infer, cache_examples=True, cache_mode="lazy")
     gr.on(
         triggers=[run_button.click, prompt.submit],
-        fn=infer,
-        inputs=[
-            prompt,
-            negative_prompt,
-            seed,
-            randomize_seed,
-            width,
-            height,
-            guidance_scale,
-            num_inference_steps,
-            normal_map_strength,
-        ],
-        outputs=[result_image, normal_map_result],
+        fn = infer,
+        inputs = [prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
+        outputs = [result]
     )
 
-if __name__ == "__main__":
-    demo.launch()
+demo.launch()

live_preview_helpers.py

@@ -0,0 +1,166 @@
+import torch
+import numpy as np
+from diffusers import FluxPipeline, AutoencoderTiny, FlowMatchEulerDiscreteScheduler
+from typing import Any, Dict, List, Optional, Union
+
+# Helper functions
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    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 retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+# FLUX pipeline function
+@torch.inference_mode()
+def flux_pipe_call_that_returns_an_iterable_of_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 = 28,
+    timesteps: List[int] = None,
+    guidance_scale: float = 3.5,
+    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,
+    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    max_sequence_length: int = 512,
+    good_vae: Optional[Any] = None,
+):
+    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,
+    )
+
+    self._guidance_scale = guidance_scale
+    self._joint_attention_kwargs = joint_attention_kwargs
+    self._interrupt = False
+
+    # 2. Define call parameters
+    batch_size = 1 if isinstance(prompt, str) else len(prompt)
+    device = self._execution_device
+
+    # 3. Encode prompt
+    lora_scale = joint_attention_kwargs.get("scale", None) if joint_attention_kwargs is not None else None
+    prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
+        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,
+        lora_scale=lora_scale,
+    )
+    # 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_shift(
+        image_seq_len,
+        self.scheduler.config.base_image_seq_len,
+        self.scheduler.config.max_image_seq_len,
+        self.scheduler.config.base_shift,
+        self.scheduler.config.max_shift,
+    )
+    timesteps, num_inference_steps = retrieve_timesteps(
+        self.scheduler,
+        num_inference_steps,
+        device,
+        timesteps,
+        sigmas,
+        mu=mu,
+    )
+    self._num_timesteps = len(timesteps)
+
+    # Handle guidance
+    guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
+
+    # 6. Denoising loop
+    for i, t in enumerate(timesteps):
+        if self.interrupt:
+            continue
+
+        timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+        noise_pred = self.transformer(
+            hidden_states=latents,
+            timestep=timestep / 1000,
+            guidance=guidance,
+            pooled_projections=pooled_prompt_embeds,
+            encoder_hidden_states=prompt_embeds,
+            txt_ids=text_ids,
+            img_ids=latent_image_ids,
+            joint_attention_kwargs=self.joint_attention_kwargs,
+            return_dict=False,
+        )[0]
+        # Yield intermediate result
+        latents_for_image = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+        latents_for_image = (latents_for_image / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+        image = self.vae.decode(latents_for_image, return_dict=False)[0]
+        yield self.image_processor.postprocess(image, output_type=output_type)[0]
+        
+        latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+        torch.cuda.empty_cache()
+
+    # Final image using good_vae
+    latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+    latents = (latents / good_vae.config.scaling_factor) + good_vae.config.shift_factor
+    image = good_vae.decode(latents, return_dict=False)[0]
+    self.maybe_free_model_hooks()
+    torch.cuda.empty_cache()
+    yield self.image_processor.postprocess(image, output_type=output_type)[0]

requirements.txt

@@ -4,6 +4,7 @@ diffusers
 torch
 numpy
 transformers
+xformers
 git+https://github.com/huggingface/diffusers.git
 sentencepiece
 scipy