Initial commit.

README.md

@@ -8,7 +8,68 @@ sdk_version: 5.25.0
 app_file: app.py
 pinned: false
 license: openrail
-short_description: Unwritten Chinese Charecters Generation in Style
+short_description: Unwritten Chinese Charecters in Style
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# What is this?
+
+Generate New Characters by combining parts in creative ways. Write them in a controlled style.
+
+- Inspired by
+  - Lin Yutang's [Ming-Kwai typewriter](https://en.wikipedia.org/wiki/Chinese_typewriter#MingKwai_design) 
+  - Wu Yue's [Glyffuser](https://yue-here.com/posts/glyffuser/)
+
+# Why
+
+- Fun to generate valid but unseen characters. (Never in a dictionary, nor Unicode).
+- Implements Lin Yutang's ideas with AI/ML, without the mechanical marvel :-/ or limitations :-)
+- Extends a font to support new charsets, and beyond to non-existent chars.
+- Adds variation/diversity/personality to generated images. No boring duplicates from the same char.
+- Other [Creative Uses](#creative-uses)
+
+# How to use this app
+- Combine components or radicals in the following way
+- Specify the 'Structure' and 'Components', in a [Polish Notation](https://en.wikipedia.org/wiki/Polish_notation) fashion - Good for tree structures
+  - ⿰: 'LR' Left-Rigth
+  - ⿱: 'TB' Top-Bottom
+  - ⿸: 'TL' Top-Left
+  - ⿹: 'TR' Top-Right
+  - ⿺: 'BL' Bottom-Left
+  - ⿴: 'OI' Outer-Inner
+  - ⿻: 'OV' Overlap
+  - ⿲: 'LMR' Left-Middle-Right
+  - ⿳: 'TMB' Top-Middle-Bottom
+  - ⿵: 'BT' Bottom Open Enclosure
+  - ⿶: 'CT' Top Open Enclosure
+  - ⿷: 'RT' Right Open Enclosure
+- Select a 'Style' by clicking the sample images
+- Hit the 'Generate' button
+- Repeat
+
+# Usage Tips
+- Simple structures work best (⿰ ⿱ ⿴ etc.)
+- "Known radicals at seen positions" work best (釒on left better than right, but may also surprise you in a good way)
+- Noto font family (sans and serif) gives the best results, as there are many training examples
+- Cursive and handwritten styles usually give good results, as they are more tolerant
+- Fonts supporting less chars are challenging
+  - Current model was trained with 300k samples for only 20 epochs
+  - Training will continue if this app gets attention or likes
+
+- For dictionary chars, [decompose](https://github.com/cburgmer/cjklib/blob/master/cjklib/data/characterdecomposition.csv) first.
+- For a part is hard to describe, or you don't care, use '？' (full-width question mark, or does it matter?)
+
+- What to do when the results are not as expected
+  - Pick a different 'sytle' which may have trained the model better
+  - Try again with a different random seed. This will change the overall structure in an unpredictable way
+  - Try again with a different 'step' number. This will change the local details in a continuous way
+
+# Creative Uses
+## Turning a bug into a feature
+When you see a funny result you didn't expect (5 or 3 dots while it should be 4), don't throw it away immediately.
+- Save the results to confuse/train OCR
+- 3vade 3vil c3nsorship
+- Share in discussion. The input text/seed/step will reliably reproduce the result.
+
+# Future Features
+- Typewriter keyboard for hard-to-input radicals, filtered by pinyin prefix
+- Direct generation of a single char, auto decomposition%  

app.py

@@ -1,36 +1,39 @@
-import gradio as gr
-import numpy as np
+import os
+import re
 import random
+import numpy as np
+
+# !!! spaces must be imported before torch/CUDA
+import spaces
 
-# import spaces #[uncomment to use ZeroGPU]
+from huggingface_hub import login
 from diffusers import DiffusionPipeline
+import gradio as gr
 import torch
 
+from utils import QPipeline
+
+
 device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use
 
-if torch.cuda.is_available():
-    torch_dtype = torch.float16
-else:
-    torch_dtype = torch.float32
+login(token=os.environ["HF_TOKEN"])
+model_repo_id = os.environ["MODEL_ID"]
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
 
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
+pipe = QPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype).to(device)
 
+MAX_SEED = 65535
+MAX_IMAGE_SIZE = 128
 
-# @spaces.GPU #[uncomment to use ZeroGPU]
+
+@spaces.GPU  # Enable ZeroGPU if needed
 def infer(
     prompt,
     negative_prompt,
     seed,
     randomize_seed,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
+    num_inference_steps=10,
     progress=gr.Progress(track_tqdm=True),
 ):
     if randomize_seed:
@@ -39,34 +42,140 @@ def infer(
     generator = torch.Generator().manual_seed(seed)
 
     image = pipe(
-        prompt=prompt,
-        negative_prompt=negative_prompt,
-        guidance_scale=guidance_scale,
-        num_inference_steps=num_inference_steps,
-        width=width,
-        height=height,
+        [prompt],
+        batch_size=1,
         generator=generator,
+        num_inference_steps=num_inference_steps
     ).images[0]
 
     return image, seed
 
 
 examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
+    "Structure: (LR 文 英). Style: style001",
+    "Structure: (TL 广 東). Style: style028",
+    "Structure: (TB 艹 (LR 禾 魚)). Style: style015",
+    "Structure: (TB 敬 音). Style: style013",
+    "Structure: (LR 釒 馬). Style: style018",
+    "Structure: (BL 走 羽). Style: style022",
+    "Structure: (LR 羊 大). Style: style005",
+    "Structure: (LR 鹿 孚). Style: style017",
+    "Structure: (OI 口 也). Style: style002",
 ]
 
+# Map style images to style names (use real image files later)
+style_options = {
+    "images/style001.png": "style001",
+    "images/style002.png": "style002",
+    "images/style003.png": "style003",
+    "images/style004.png": "style004",
+    "images/style005.png": "style005",
+    "images/style006.png": "style006",
+    "images/style007.png": "style007",
+    "images/style008.png": "style008",
+    "images/style009.png": "style009",
+    "images/style010.png": "style010",
+    "images/style011.png": "style011",
+    "images/style012.png": "style012",
+    "images/style013.png": "style013",
+    "images/style014.png": "style014",
+    "images/style015.png": "style015",
+    # "images/style016.png": "style016", very similar to 002
+    "images/style017.png": "style017",
+    "images/style018.png": "style018",
+    "images/style019.png": "style019",
+    "images/style020.png": "style020",
+    "images/style021.png": "style021",
+    "images/style022.png": "style022",
+    "images/style023.png": "style023",
+    "images/style024.png": "style024",
+    "images/style025.png": "style025",
+    "images/style026.png": "style026",
+    "images/style027.png": "style027",
+    "images/style028.png": "style028",
+    "images/style029.png": "style029",
+}
+
+
+def apply_style_on_click(evt: gr.SelectData, prompt_text):
+    index = evt.index
+    style_label = list(style_options.values())[index]
+
+    if re.search(r"Style: [^\n]+", prompt_text):
+        return re.sub(r"Style: [^\n]+", f"Style: {style_label}", prompt_text)
+    else:
+        return prompt_text.strip() + f" Style: {style_label}"
+
+
+# CSS for fixing Gallery layout
 css = """
 #col-container {
     margin: 0 auto;
-    max-width: 640px;
+    max-width: 800px;
 }
 """
 
 with gr.Blocks(css=css) as demo:
     with gr.Column(elem_id="col-container"):
-        gr.Markdown(" # Text-to-Image Gradio Template")
+        gr.Markdown(" # NeoChar ")
+        gr.Markdown(" ## What: Create New Characters. Write them in Style. ")
+        gr.Markdown(" * NO more missing glyphs - Make them when fonts don't support! ")
+        gr.Markdown(" * Create valid and new Hanzi/Kanji that never existed before ")
+        gr.Markdown(" * Calligraphy with diversity - powered by controlled chaos")
+        gr.Markdown(" ## How ")
+        gr.Markdown(" * Specify 'Structure' and 'Components'. LR: Left-Right, TB: Top-Bottom, etc. ")
+        gr.Markdown(" * Select a 'Style' ")
+        gr.Markdown(" * (examples below)")
+        gr.Markdown(" ## README for more ")
+
+
+        gr.HTML("""
+        <style>
+        .gallery-container .gallery-item {
+            width: 60px !important;
+            height: 60px !important;
+            padding: 0 !important;
+            margin: 4px !important;
+            border-radius: 4px;
+            overflow: hidden;
+            background: none !important;
+            box-shadow: none !important;
+        }
+
+        .gallery-container .gallery-item img {
+            width: 64px !important;
+            height: 64px !important;
+            object-fit: cover;
+            display: block;
+            margin: auto;
+        }
+
+        .gallery-container button {
+            all: unset !important;
+            padding: 0 !important;
+            margin: 0 !important;
+            border: none !important;
+            background: none !important;
+            box-shadow: none !important;
+        }
+
+        .gallery__modal,
+        .gallery-container .preview,
+        .gallery-container .gallery-item:focus-visible {
+            display: none !important;
+            pointer-events: none !important;
+        }
+        </style>
+        """)
+
+        gallery = gr.Gallery(
+            value=list(style_options.keys()),
+            label="Click any image",
+            columns=7,
+            allow_preview=False,
+            height=None,
+            elem_classes=["gallery-container"]
+        )
 
         with gr.Row():
             prompt = gr.Text(
@@ -76,9 +185,14 @@ with gr.Blocks(css=css) as demo:
                 placeholder="Enter your prompt",
                 container=False,
             )
-
             run_button = gr.Button("Run", scale=0, variant="primary")
 
+        gallery.select(
+            fn=apply_style_on_click,
+            inputs=[prompt],
+            outputs=prompt
+        )
+
         result = gr.Image(label="Result", show_label=False)
 
         with gr.Accordion("Advanced Settings", open=False):
@@ -100,40 +214,16 @@ with gr.Blocks(css=css) as demo:
             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,
+                    maximum=20,
                     step=1,
-                    value=2,  # Replace with defaults that work for your model
+                    value=10,
                 )
 
         gr.Examples(examples=examples, inputs=[prompt])
+
     gr.on(
         triggers=[run_button.click, prompt.submit],
         fn=infer,
@@ -142,9 +232,6 @@ with gr.Blocks(css=css) as demo:
             negative_prompt,
             seed,
             randomize_seed,
-            width,
-            height,
-            guidance_scale,
             num_inference_steps,
         ],
         outputs=[result, seed],
@@ -152,3 +239,4 @@ with gr.Blocks(css=css) as demo:
 
 if __name__ == "__main__":
     demo.launch()
+

requirements.txt

@@ -1,6 +1,14 @@
+numpy
 accelerate
 diffusers
 invisible_watermark
 torch
 transformers
-xformers
+xformers
+sentencepiece
+datasets
+einops
+Pillow
+torchvision
+tqdm
+imwatermark

utils.py

@@ -0,0 +1,108 @@
+import os
+import numpy as np
+import torch
+from torch import nn
+from torch.utils.data import Dataset, DataLoader
+from torchvision import transforms as T
+from PIL import Image as PILImage, ImageDraw, ImageFont
+from imwatermark import WatermarkEncoder
+
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from diffusers.utils.torch_utils import randn_tensor
+from transformers import MT5Tokenizer, MT5EncoderModel
+from typing import List, Optional, Tuple, Union
+
+# Determine device and torch dtype
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+
+# Load MT5 tokenizer and encoder (can be replaced with private model + token if needed)
+tokenizer = MT5Tokenizer.from_pretrained("google/mt5-small", use_safetensors=True)
+encoder_model = MT5EncoderModel.from_pretrained("google/mt5-small", use_safetensors=True).to(device=device, dtype=torch_dtype)
+encoder_model.eval()
+
+class QPipeline(DiffusionPipeline):
+    def __init__(self, unet, scheduler):
+        super().__init__()
+        self.register_modules(unet=unet, scheduler=scheduler)
+
+    def add_watermark(self, img: PILImage.Image) -> PILImage.Image:
+        # Resize image to 256, as 128 is too small for watermark
+        img = img.resize((256, 256), resample=PILImage.BICUBIC)
+
+        watermark_str = os.getenv("WATERMARK_URL", "hf.co/lqume/new-hanzi")
+        encoder = WatermarkEncoder()
+        encoder.set_watermark('bytes', watermark_str.encode('utf-8'))
+
+        # Convert PIL image to NumPy array
+        img_np = np.asarray(img.convert("RGB"))  # ensure 3-channel RGB
+        watermarked_np = encoder.encode(img_np, 'dwtDct')
+
+        # Convert back to PIL
+        return PILImage.fromarray(watermarked_np)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        texts: List[str],
+        batch_size: int = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        num_inference_steps: int = 20,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+    ) -> Union[ImagePipelineOutput, Tuple[List[PILImage.Image]]]:
+
+        batch_size = len(texts)
+
+        # Tokenize input text
+        tokenized = tokenizer(
+            texts,
+            return_tensors="pt",
+            padding="max_length",
+            truncation=True,
+            max_length=48
+        )
+        input_ids = tokenized["input_ids"].to(device=device, dtype=torch.long)
+        attention_mask = tokenized["attention_mask"].to(device=device, dtype=torch.long)
+
+        # Encode to latent space
+        encoded = encoder_model.encoder(input_ids=input_ids, attention_mask=attention_mask)
+
+        # Prepare noise tensor
+        if isinstance(self.unet.config.sample_size, int):
+            image_shape = (
+                batch_size,
+                self.unet.config.in_channels,
+                self.unet.config.sample_size,
+                self.unet.config.sample_size,
+            )
+        else:
+            image_shape = (batch_size, self.unet.config.in_channels, *self.unet.config.sample_size)
+
+        image = randn_tensor(image_shape, generator=generator, device=self.device, dtype=torch_dtype)
+
+        # Run denoising loop
+        self.scheduler.set_timesteps(num_inference_steps)
+
+        for timestep in self.progress_bar(self.scheduler.timesteps):
+            noise_pred = self.unet(
+                image,
+                timestep,
+                encoder_hidden_states=encoded.last_hidden_state,
+                encoder_attention_mask=attention_mask.bool(),
+                return_dict=False
+            )[0]
+
+            image = self.scheduler.step(noise_pred, timestep, image, generator=generator, return_dict=False)[0]
+
+        # Final image post-processing
+        image = image.clamp(0, 1).cpu().permute(0, 2, 3, 1).numpy()
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+            image = [self.add_watermark(img) for img in image]
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)
+