app.py

# app.py
from datasets import load_dataset
import gradio as gr
from gradio_client import Client
import json, os, random, torch
from diffusers import FluxPipeline, AutoencoderKL
from live_preview_helpers import flux_pipe_call_that_returns_an_iterable_of_images
import spaces

# ─────────────────────────────── 1. Device ────────────────────────────────
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# ─────────────────────── 2. Image / FLUX pipeline ─────────────────────────
pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.bfloat16
).to(device)
good_vae = AutoencoderKL.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="vae",
    torch_dtype=torch.bfloat16
).to(device)
pipe.flux_pipe_call_that_returns_an_iterable_of_images = (
    flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe)
)

# ───────────────────────── 3. LLM (Zephyr-chat) ───────────────────────────
llm_client = Client("HuggingFaceH4/zephyr-chat")     # public Space
CHAT_API   = llm_client.view_api()[0]["api_name"]    # e.g. "/chat"

def call_llm(
    user_prompt: str,
    system_prompt: str = "You are Zephyr, a helpful and creative assistant.",
    history: list | None = None,
    temperature: float = 0.7,
    top_p: float = 0.9,
    max_tokens: int = 1024,
) -> str:
    """
    Robust wrapper around the Zephyr chat Space.
    Falls back to '...' on any error so the Gradio UI never crashes.
    """
    history = history or []
    try:
        # Zephyr-chat expects: prompt, system_prompt, history, temperature, top_p, max_new_tokens
        result = llm_client.predict(
            user_prompt,
            system_prompt,
            history,
            temperature,
            top_p,
            max_tokens,
            api_name=CHAT_API,
        )
        # Some Spaces return a plain string, others return the old tuple format.
        return result.strip() if isinstance(result, str) else result[1][0][-1].strip()
    except Exception as e:
        print(f"[LLM error] {e}")
        return "..."

# ───────────────────────── 4. Persona dataset ────────────────────────────
ds = load_dataset("MohamedRashad/FinePersonas-Lite", split="train")

def get_random_persona_description() -> str:
    idx = random.randint(0, len(ds) - 1)
    return ds[idx]["persona"]

# ─────────────────────────── 5. Prompts  ─────────────────────────────────
prompt_template = """Generate a character with this persona description:

{persona_description}

In a world with this description:

{world_description}

Write the character in JSON format with these keys:
name, background, appearance, personality, skills_and_abilities, goals,
conflicts, backstory, current_situation, spoken_lines (list of strings).

Respond with **only** the JSON (no markdown, no fencing)."""

world_description_prompt = (
    "Invent a short, unique and vivid world description. "
    "Respond with the description only."
)

# ─────────────────────── 6. Gradio helper funcs ─────────────────────────
def get_random_world_description() -> str:
    return call_llm(world_description_prompt)

@spaces.GPU(duration=75)
def infer_flux(character_json):
    """Stream intermediate images while FLUX denoises."""
    for image in pipe.flux_pipe_call_that_returns_an_iterable_of_images(
        prompt=character_json["appearance"],
        guidance_scale=3.5,
        num_inference_steps=28,
        width=1024,
        height=1024,
        generator=torch.Generator("cpu").manual_seed(0),
        output_type="pil",
        good_vae=good_vae,
    ):
        yield image

def generate_character(world_description: str,
                       persona_description: str,
                       progress=gr.Progress(track_tqdm=True)):
    raw = call_llm(
        prompt_template.format(
            persona_description=persona_description,
            world_description=world_description,
        ),
        max_tokens=1024,
    )
    try:
        return json.loads(raw)
    except json.JSONDecodeError:
        # One retry if the LLM hallucinated
        raw = call_llm(
            prompt_template.format(
                persona_description=persona_description,
                world_description=world_description,
            ),
            max_tokens=1024,
        )
        return json.loads(raw)

# ───────────────────────────── 7. UI ─────────────────────────────────────
app_description = """
- Generates a character profile (JSON) from a world + persona description.
- **Appearance** images come from [FLUX-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev).
- **Back-stories** come from [Zephyr-7B-β](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta).
- Personas are sampled from [FinePersonas-Lite](https://huggingface.co/datasets/MohamedRashad/FinePersonas-Lite).

Tip → Write or randomise a world, then spin the persona box to see how the same
world shapes different heroes.
"""

with gr.Blocks(title="Character Generator", theme="Nymbo/Nymbo_Theme") as demo:
    gr.Markdown("<h1 style='text-align:center'>🧙‍♀️ Character Generator</h1>")
    gr.Markdown(app_description.strip())

    with gr.Row():
        world_description = gr.Textbox(label="World Description", lines=10, scale=4)
        persona_description = gr.Textbox(
            label="Persona Description",
            value=get_random_persona_description(),
            lines=10,
            scale=1,
        )

    with gr.Row():
        random_world_btn   = gr.Button("🔄 Random World",   variant="secondary")
        submit_btn         = gr.Button("✨ Generate Character", variant="primary", scale=5)
        random_persona_btn = gr.Button("🔄 Random Persona", variant="secondary")

    with gr.Row():
        character_image = gr.Image(label="Character Image")
        character_json  = gr.JSON(label="Character Description")

    # Hooks
    submit_btn.click(
        generate_character,
        inputs=[world_description, persona_description],
        outputs=[character_json],
    ).then(
        infer_flux,
        inputs=[character_json],
        outputs=[character_image],
    )

    random_world_btn.click(
        get_random_world_description,
        outputs=[world_description],
    )
    random_persona_btn.click(
        get_random_persona_description,
        outputs=[persona_description],
    )

demo.queue().launch(share=False)