app.py

import re
import nltk
import torch
import spaces

import gradio as gr

from threading import Thread

from nltk.tag import pos_tag
from nltk.chunk import ne_chunk
from nltk.tokenize import word_tokenize

from peft import PeftModel
from transformers import CsmForConditionalGeneration
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer, AutoProcessor

try:
    nltk.data.find('tokenizers/punkt')
    nltk.data.find('taggers/averaged_perceptron_tagger')
    nltk.data.find('chunkers/maxent_ne_chunker')
    nltk.data.find('corpora/words')
except LookupError:
    nltk.download('punkt', quiet=True)
    nltk.download('averaged_perceptron_tagger', quiet=True)
    nltk.download('maxent_ne_chunker', quiet=True)
    nltk.download('words', quiet=True)


SYSTEM_PROMPT = """You are an expert creative director specializing in visual descriptions for image generation.

Your task: Transform the user's concept into a rich, detailed image description while PRESERVING their core idea.

IMPORTANT RULES:
1. Keep ALL key elements (intents, entities) from the original concept
2. Enhance with artistic details, NOT change the fundamental idea
3. Maintain the user's intended subject, action, and setting

You should elaborate on:
• Visual composition and perspective (bird's eye, close-up, wide angle, etc.)
• Artistic style (photorealistic, impressionist, specific artist like Van Gogh, etc.)
• Color palette and color temperature
• Lighting (golden hour, dramatic shadows, soft diffused, etc.)
• Atmosphere and mood
• Textures and materials (rough, smooth, metallic, organic, etc.)
• Technical details (medium, brushwork, rendering style)
• Environmental context (time of day, weather, season, era)
• Level of detail and focus points

Output format: A single, flowing paragraph that reads naturally as an image prompt."""

prompt_model = AutoModelForCausalLM.from_pretrained("shb777/PromptTuner-v0.1")
tokenizer = AutoTokenizer.from_pretrained("shb777/PromptTuner-v0.1")
prompt_model.eval()

CSM_BASE_MODEL_ID = "sesame/csm-1b"
CSM_ADAPTER_ID = "shb777/csm-maya-exp2"
SPEAKER_ID = 4  # Was trained on this speaker ID
device = "cuda" if torch.cuda.is_available() else "cpu"

csm_processor = AutoProcessor.from_pretrained(CSM_BASE_MODEL_ID)
csm_model = CsmForConditionalGeneration.from_pretrained(CSM_BASE_MODEL_ID, device_map=device)
csm_model = PeftModel.from_pretrained(csm_model, CSM_ADAPTER_ID)
csm_model.eval()


def extract_key_phrases(text: str) -> list: # We will highlight key phrases in the enhanced prompt
    phrases = []
    try:
        tokens = word_tokenize(text)
        tagged = pos_tag(tokens)
        chunks = ne_chunk(tagged)

        current_phrase = []
        for chunk in chunks:
            if hasattr(chunk, 'label'):
                phrase = ' '.join([token for token, _ in chunk.leaves()])
                phrases.append(phrase.lower())
            elif chunk[1].startswith('NN'):
                current_phrase.append(chunk[0])
            elif chunk[1].startswith('JJ') and current_phrase:
                current_phrase.append(chunk[0])
            else:
                if current_phrase:
                    phrases.append(' '.join(current_phrase).lower())
                    current_phrase = []

        if current_phrase:
            phrases.append(' '.join(current_phrase).lower())

        for word, tag in tagged:
            if tag.startswith('JJ') or tag in ('RB', 'RBR', 'RBS'):
                phrases.append(word.lower())

    except Exception:
        words = re.findall(r'\b[a-zA-Z]{3,}\b', text.lower())
        phrases = list(set(words))

    multi_word = re.findall(r'\b[a-zA-Z]{3,}(?:\s+[a-zA-Z]{3,}){1,3}\b', text)
    phrases.extend([mw.lower() for mw in multi_word])

    phrases = list(set(phrases))
    phrases.sort(key=len, reverse=True)

    return phrases[:20]


def highlight_matches(original_input: str, enhanced_output: str) -> str:
    if not original_input.strip():
        return f'<p class="output-text">{enhanced_output}</p>'

    key_phrases = extract_key_phrases(original_input)
    if not key_phrases:
        return f'<p class="output-text">{enhanced_output}</p>'

    key_phrases.sort(key=len, reverse=True)

    output = enhanced_output
    highlighted_spans = []

    for phrase in key_phrases:
        pattern = re.compile(r'\b' + re.escape(phrase) + r'\b', re.IGNORECASE)

        def replace_with_highlight(match):
            matched_text = match.group(0)
            start = match.start()
            for h_start, h_end in highlighted_spans:
                if start >= h_start and start <= h_end:
                    return matched_text
            highlighted_spans.append((start, match.end()))
            return f'<mark class="highlight-keyword">{matched_text}</mark>'

        output = pattern.sub(replace_with_highlight, output)

    return f'<p class="output-text">{output}</p>'


def enhance_prompt(user_prompt: str):
    if not user_prompt or not user_prompt.strip():
        yield (
            '<span class="placeholder-text">Please enter a prompt to enhance.</span>',
            "",
            gr.update(interactive=True),
            gr.update(interactive=True)
        )
        return

    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": user_prompt}
    ]

    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(prompt, return_tensors="pt")

    streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True, skip_prompt=True)
    generation_kwargs = {
        'max_new_tokens': 512,
        'streamer': streamer,
        'do_sample': True,
        'temperature': 1,
        'top_p': 0.95,
        'top_k': 64
    }

    placeholder = '<span class="placeholder-text">Your enhanced prompt will appear here</span>'
    yield placeholder, "", gr.update(interactive=False), gr.update(interactive=False)

    thread = Thread(target=prompt_model.generate, kwargs={**inputs, **generation_kwargs})
    thread.start()

    output = ""
    for text in streamer:
        output += text
        highlighted = highlight_matches(user_prompt, output)
        yield highlighted, output, gr.update(), gr.update()

    final_highlighted = highlight_matches(user_prompt, output)
    yield final_highlighted, output, gr.update(interactive=True), gr.update(interactive=True)


@spaces.GPU
def generate_tts_gpu(text):
    conversation = [
        {"role": str(SPEAKER_ID), "content": [{"type": "text", "text": text}]},
    ]

    enc = csm_processor.apply_chat_template(
        conversation,
        tokenize=True,
        return_dict=True,
        return_tensors="pt",
    ).to(device)

    gen_kwargs = {
        "max_new_tokens": 375,
        # "do_sample": True,
        # "temperature": 0.7,
        # "depth_decoder_do_sample": True,
        # "depth_decoder_temperature": 0.7,
        # "depth_decoder_top_k": 20,
        # "depth_decoder_top_p": 0.95,
    }

    audio = csm_model.generate(
        **enc,
        **gen_kwargs,
        output_audio=True
    )
    audio_array = audio[0].to(torch.float32).cpu().numpy()
    return (24000, audio_array)


def text_to_speech(text: str):
    if not text or not text.strip():
        raise gr.Error("Please enter text to convert to speech.")

    if len(text) > 200:
        raise gr.Error("Text too long. Please limit to 200 characters or split into sentences.")

    # Preprocess text - remove characters that CSM struggles with
    text = text.replace('(', '').replace(')', '')
    text = text.replace('"', '').replace('"', '').replace('"', '')
    text = text.replace(';', ',')
    text = text.replace('!', ' ')
    text = text.replace('[', '').replace(']', '')
    text = text.replace('/', ' ')

    try:
        audio_result = generate_tts_gpu(text)
        return audio_result
    except Exception as e:
        raise gr.Error(f"Failed to generate speech: {str(e)}")

with open("style.css", "r") as f: # Load CSS for styling
    custom_css = f.read()

with gr.Blocks(css=custom_css, title="TinkerSpace") as demo:

    with gr.Tabs():
        with gr.Tab("CSM Maya TTS"):
            with gr.Row(elem_classes=["main-grid"]):
                with gr.Column(elem_classes=["card"]):
                    gr.HTML('<label class="form-label">Text Input</label>')

                    tts_input = gr.Textbox(
                        placeholder="Enter text to convert to speech...",
                        lines=5,
                        show_label=False,
                        container=False,
                        elem_classes=["input-textarea"]
                    )

                    with gr.Row(elem_classes=["flex gap-2 mt-6"]):
                        generate_tts_btn = gr.Button(
                            "Generate Speech",
                            variant="primary",
                            scale=2,
                            elem_classes=["btn", "btn-primary"]
                        )
                        clear_tts_btn = gr.Button(
                            "Clear",
                            scale=1,
                            elem_classes=["btn", "btn-secondary"]
                        )

                with gr.Column(elem_classes=["card"]):
                    gr.HTML('<label class="form-label">Generated Audio</label>')

                    tts_output = gr.Audio(
                        label=None,
                        show_label=False,
                        autoplay=False,
                        interactive=False,
                        elem_classes=["output-container"]
                    )

            with gr.Column(elem_classes=["examples-section"]):
                gr.Examples(
                    examples=[
                        ['You went to the party, even though I explicitly told you not to?'],
                        ["With a gentle touch and a loving smile, she reassured, 'Dont worry, my love. We'll get through this together, just like we always have. I love you.'"],
                        ["After years of work, Heisenberg finally published a ground-breaking cutting-edge research paper on quantum physics."],
                        ['"Uh, are you sure about this?" Tim asked nervously, looking at the steep slope before them. "Whoa, it\'s higher than I thought," he continued, his voice filled with trepidation.'],
                        ['"Shh, Lucy, shh, we mustn\'t wake your baby brother," Tom whispered, as they tiptoed past the nursery.']
                    ],
                    inputs=tts_input,
                    label="Examples"
                )

            with gr.Row():
                gr.Markdown(
                    "Powered by [CSM Maya](https://huggingface.co/shb777/csm-maya-exp2)"
                )

        with gr.Tab("Prompt Enhancer"):
            with gr.Row(elem_classes=["main-grid"]):
                with gr.Column(elem_classes=["card"]):
                    gr.HTML('<label class="form-label">Input Prompt</label>')

                    input_text = gr.Textbox(
                        placeholder="Describe your image concept... e.g., fox, red tail, blue moon, clouds",
                        lines=5,
                        show_label=False,
                        autofocus=True,
                        container=False,
                        elem_classes=["input-textarea"]
                    )

                    with gr.Row(elem_classes=["flex gap-2 mt-6"]):
                        enhance_btn = gr.Button(
                            "Enhance Prompt",
                            variant="primary",
                            scale=2,
                            elem_classes=["btn", "btn-primary"]
                        )
                        clear_btn = gr.Button(
                            "Clear",
                            scale=1,
                            elem_classes=["btn", "btn-secondary"]
                        )

                with gr.Column(elem_classes=["card"]):
                    gr.HTML('<label class="form-label">Enhanced Prompt</label>')

                    output_html = gr.HTML(
                        value='<span class="placeholder-text">Your enhanced prompt will appear here</span>',
                        elem_classes=["output-container"]
                    )

                    raw_output = gr.Textbox(visible=False)

            with gr.Column(elem_classes=["examples-section"]):
                gr.Examples(
                    examples=[
                        ["fox, red tail, blue moon, clouds"],
                        ["room with french window, cozy morning vibes, minimal"],
                        ["anime style, sunset, japan"]
                    ],
                    inputs=input_text,
                    label="Examples"
                )

            with gr.Row():
                gr.Markdown(
                    "Powered by [PromptTuner](https://huggingface.co/shb777/PromptTuner-v0.1), "
                    "a finetuned gemma3-270M model specifically designed to enhance text prompts "
                    "for text-to-image generation."
                )

    enhance_btn.click(
        fn=enhance_prompt,
        inputs=[input_text],
        outputs=[output_html, raw_output, enhance_btn, clear_btn]
    )

    clear_btn.click(
        fn=lambda: (
            "",
            '<span class="placeholder-text">Your enhanced prompt will appear here</span>',
            "",
            gr.update(interactive=True),
            gr.update(interactive=True)
        ),
        inputs=None,
        outputs=[input_text, output_html, raw_output, enhance_btn, clear_btn]
    )

    generate_tts_btn.click(
        fn=text_to_speech,
        inputs=[tts_input],
        outputs=[tts_output]
    )

    clear_tts_btn.click(
        fn=lambda: ("", None),
        inputs=None,
        outputs=[tts_input, tts_output]
    )

if __name__ == "__main__":
    demo.queue(max_size=20).launch(mcp_server=True)