Initial commit

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,4 @@
+.gradio/
+.ipynb_checkpoints/
+.DS_Store
+__pycache__/

README.md

@@ -1,5 +1,5 @@
 ---
-title: Text-to-Image Gradio Template
+title: FlexTok
 emoji: 🖼
 colorFrom: purple
 colorTo: red
@@ -7,6 +7,8 @@ sdk: gradio
 sdk_version: 5.0.1
 app_file: app.py
 pinned: false
+license: apache-2.0
+short_description: FlexTok flexible sequence length autoencoding demo
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -1,154 +1,181 @@
+from typing import List
+import os
+import spaces
 import gradio as gr
-import numpy as np
 import random
+from PIL import Image
+import matplotlib.pyplot as plt
 
-# import spaces #[uncomment to use ZeroGPU]
-from diffusers import DiffusionPipeline
+import einops
+import numpy as np
 import torch
+from torchvision import transforms
+import torchvision.transforms.functional as TF
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use
+from flextok.flextok_wrapper import FlexTokFromHub
+from flextok.utils.demo import imgs_from_urls, denormalize, batch_to_pil
+from flextok.utils.misc import detect_bf16_support, get_bf16_context, get_generator
 
+
+# We recommend running this demo on an A100 GPU
 if torch.cuda.is_available():
-    torch_dtype = torch.float16
+    device = "cuda"
+    gpu_type = torch.cuda.get_device_name(torch.cuda.current_device())
+    power_device = f"{gpu_type}"
+    torch.cuda.max_memory_allocated(device=device)
+    # Detect if bf16 is enabled or not
+    enable_bf16 = detect_bf16_support()
 else:
-    torch_dtype = torch.float32
+    device, power_device, enable_bf16 = "cpu", "CPU", False
+print(f'Device: {device}, GPU type: {gpu_type}')
+print('BF16 enabled:', enable_bf16)
+
+
+# The flag below controls whether to allow TF32 on matmul. This flag defaults to False in PyTorch 1.12 and later.
+torch.backends.cuda.matmul.allow_tf32 = True
+# The flag below controls whether to allow TF32 on cuDNN. This flag defaults to True.
+torch.backends.cudnn.allow_tf32 = True
+
+# Global no_grad
+torch.set_grad_enabled(False)
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
 
 MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
-
-
-# @spaces.GPU #[uncomment to use ZeroGPU]
-def infer(
-    prompt,
-    negative_prompt,
-    seed,
-    randomize_seed,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
-    progress=gr.Progress(track_tqdm=True),
-):
+
+MODEL_ID = 'EPFL-VILAB/flextok_d18_d28_dfn'
+MODEL_NAME = 'FlexTok d18-d28 (DFN)'
+
+# Load FlexTok model from HF Hub
+flextok_model = FlexTokFromHub.from_pretrained(MODEL_ID).to(device).eval()
+
+
+def img_from_path(
+    path: str,
+    img_size: int = 256,
+    mean: List[float] = [0.5, 0.5, 0.5],
+    std: List[float] = [0.5, 0.5, 0.5],
+) -> torch.Tensor:
+    # Image loading helper function
+    img_pil = Image.open(path).convert("RGB")
+
+    transform = transforms.Compose(
+        [
+            transforms.Resize(img_size),
+            transforms.CenterCrop(img_size),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=mean, std=std),
+        ]
+    )
+
+    return transform(img_pil).unsqueeze(0)
+
+
+@spaces.GPU(duration=20)
+def infer(img_path, seed=0, randomize_seed=False, timesteps=20, cfg_scale=7.5, perform_norm_guidance=True):
     if randomize_seed:
-        seed = random.randint(0, MAX_SEED)
+        seed = None
+
+    imgs = img_from_path(img_path).to(device)
+
+    # Tokenize images once
+    with get_bf16_context(enable_bf16):
+        tokens_list = flextok_model.tokenize(imgs)
 
-    generator = torch.Generator().manual_seed(seed)
+    # Create all token subsequences
+    k_keep_list = [1, 2, 4, 8, 16, 32, 64, 128, 256]
+    tokens_list = tokens_list*len(k_keep_list)
+    subseq_list = [seq[:,:k_keep].clone() for seq, k_keep in zip(tokens_list, k_keep_list)]
 
-    image = pipe(
-        prompt=prompt,
-        negative_prompt=negative_prompt,
-        guidance_scale=guidance_scale,
-        num_inference_steps=num_inference_steps,
-        width=width,
-        height=height,
-        generator=generator,
-    ).images[0]
+    # Detokenize various subsequences in parallel. Batch size is 9.
+    with get_bf16_context(enable_bf16):
+        generator = get_generator(seed=seed, device=device)
+        all_reconst = flextok_model.detokenize(
+            subseq_list, timesteps=timesteps, 
+            guidance_scale=cfg_scale, perform_norm_guidance=perform_norm_guidance,
+            generator=generator, verbose=False,
+        )
+    
+    # Transform to PIL images
+    all_images = [
+        (TF.to_pil_image(denormalize(reconst_k).clamp(0,1)), f'{k_keep} tokens') 
+        for reconst_k, k_keep in zip(all_reconst, k_keep_list)
+    ]
 
-    return image, seed
+    return all_images
 
 
 examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
+    'examples/0.png', 'examples/1.png', 'examples/2.png',
+    'examples/3.png', 'examples/4.png', 'examples/5.png',
 ]
 
-css = """
+css="""
 #col-container {
     margin: 0 auto;
-    max-width: 640px;
+    max-width: 1500px;
+}
+#col-input-container {
+    margin: 0 auto;
+    max-width: 400px;
+}
+#run-button {
+    margin: 0 auto;
+}
+#gallery {
+    aspect-ratio: 1/1 !important;
+    height: auto !important;
 }
 """
 
-with gr.Blocks(css=css) as demo:
+with gr.Blocks(css=css, theme=gr.themes.Base()) as demo:
+    
     with gr.Column(elem_id="col-container"):
-        gr.Markdown(" # Text-to-Image Gradio Template")
-
+        gr.Markdown(f"""
+        # FlexTok: Resampling Images into 1D Token Sequences of Flexible Length
+        """)
+        
         with gr.Row():
-            prompt = gr.Text(
-                label="Prompt",
-                show_label=False,
-                max_lines=1,
-                placeholder="Enter your prompt",
-                container=False,
-            )
-
-            run_button = gr.Button("Run", scale=0, variant="primary")
-
-        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,
-                maximum=MAX_SEED,
-                step=1,
-                value=0,
+            with gr.Column(elem_id="col-input-container"):
+                gr.Markdown(f"""
+                [`Website`](https://flextok.epfl.ch) | [`arXiv`](https://arxiv.org/abs/2502.13967) | [`GitHub`](https://github.com/apple/ml-flextok)
+
+                Official demo for: <br>
+                [**FlexTok: Resampling Images into 1D Token Sequences of Flexible Length**](https://arxiv.org/abs/2502.13967), arXiv 2025 <br>
+                *[Roman Bachmann](https://roman-bachmann.github.io/)\*, [Jesse Allardice](https://github.com/JesseAllardice)\*, [David Mizrahi](https://dmizrahi.com/)\*, [Enrico Fini](https://scholar.google.com/citations?user=OQMtSKIAAAAJ), [Oğuzhan Fatih Kar](https://ofkar.github.io/), [Elmira Amirloo](https://elamirloo.github.io/), [Alaaeldin El-Nouby](https://aelnouby.github.io/), [Amir Zamir](https://vilab.epfl.ch/zamir/), [Afshin Dehghan](https://scholar.google.com/citations?user=wcX-UW4AAAAJ)*
+
+                This demo uses the FlexTok tokenizer to autoencode the given RGB input, using [{MODEL_ID}](https://huggingface.co/{MODEL_ID}), running on *{power_device}*. The FlexTok encoder produces a 1D sequence of discrete tokens that are ordered in a coarse-to-fine manner. We show reconstructions from truncated subsequences, using the first 1, 2, 4, 8, ..., 256 tokens. As you will see, the first tokens capture the high-level semantic content, while subsequent ones add more fine-grained detail.
+                """)
+                
+                img_path = gr.Image(label='RGB input image', type='filepath')
+                run_button = gr.Button(f"Autoencode with {MODEL_NAME}", scale=0, elem_id="run-button")
+
+                with gr.Accordion("Advanced Settings", open=False):
+                    gr.Markdown(f"""
+                    The FlexTok decoder is a rectified flow model. The following settings control the seed of the initial noise, the number of denoising timesteps, the guidance scale, and whether to perform [Adaptive Projected Guidance](https://arxiv.org/abs/2410.02416) (we recommend enabling it).
+                    """)
+                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
+                    randomize_seed = gr.Checkbox(label="Randomize seed", value=False)
+                    timesteps = gr.Slider(label="Denoising timesteps", minimum=1, maximum=1000, step=1, value=20)
+                    cfg_scale = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=15.0, step=0.1, value=7.5)
+                    perform_norm_guidance = gr.Checkbox(label="Perform Adaptive Projected Guidance", value=True)
+        
+            result = gr.Gallery(
+                label="Reconstructions", show_label=True, elem_id="gallery", type='pil',
+                columns=[3], rows=None, object_fit="contain", height=800
             )
 
-            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,  # Replace with defaults that work for your model
-                )
-
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,  # Replace with defaults that work for your model
-                )
-
-            with gr.Row():
-                guidance_scale = gr.Slider(
-                    label="Guidance scale",
-                    minimum=0.0,
-                    maximum=10.0,
-                    step=0.1,
-                    value=0.0,  # Replace with defaults that work for your model
-                )
-
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=50,
-                    step=1,
-                    value=2,  # Replace with defaults that work for your model
-                )
-
-        gr.Examples(examples=examples, inputs=[prompt])
-    gr.on(
-        triggers=[run_button.click, prompt.submit],
-        fn=infer,
-        inputs=[
-            prompt,
-            negative_prompt,
-            seed,
-            randomize_seed,
-            width,
-            height,
-            guidance_scale,
-            num_inference_steps,
-        ],
-        outputs=[result, seed],
+        gr.Examples(
+            examples = examples,
+            fn = infer,
+            inputs = [img_path],
+            outputs = [result],
+            cache_examples='lazy',
+        )
+
+    run_button.click(
+        fn = infer,
+        inputs = [img_path, seed, randomize_seed, timesteps, cfg_scale, perform_norm_guidance],
+        outputs = [result]
     )
 
-if __name__ == "__main__":
-    demo.launch()
+demo.queue(max_size=10).launch(share=True)

requirements.txt

@@ -1,6 +1,2 @@
-accelerate
-diffusers
-invisible_watermark
-torch
-transformers
-xformers
+flextok @ git+https://github.com/apple/ml-flextok #@e115399
+spaces