app.py

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

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

# ────────────────────── 2.  FLUX image 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_iter.__get__(pipe)

# ───────────────────────── 3.  LLM client (robust) ──────────────────────────
LLM_SPACES = [
    "https://huggingfaceh4-zephyr-chat.hf.space",
    "meta-llama/Llama-3.3-70B-Instruct",
    "huggingface-projects/gemma-2-9b-it",
]

def first_live_space(space_ids: list[str]) -> Client:
    """
    Return the first Space whose /chat endpoint answers a 1-token echo.
    """
    for sid in space_ids:
        try:
            print(f"[info] probing {sid}")
            c = Client(sid, hf_token=os.getenv("HF_TOKEN"))
            _ = c.predict("ping", 8, api_name="/chat")  # simple health check
            print(f"[info] using {sid}")
            return c
        except Exception as e:
            print(f"[warn] {sid} unusable → {e}")
    raise RuntimeError("No live chat Space found!")

llm_client = first_live_space(LLM_SPACES)
CHAT_API   = "/chat"  # universal endpoint for TGI-style Spaces

def call_llm(prompt: str,
             max_tokens: int = 256,
             temperature: float = 0.6,
             top_p: float = 0.9) -> str:
    """
    Send a single-message chat to the Space. Extra sliders in the remote UI must
    be supplied in positional order after the prompt, so we match Zephyr/Gemma:
        [prompt, max_tokens, temperature, top_p, repeat_penalty, presence_penalty]
    We pass only the first four; the Space will fill the rest with defaults.
    """
    try:
        return llm_client.predict(
            prompt, max_tokens, temperature, top_p, api_name=CHAT_API
        ).strip()
    except Exception as exc:
        print(f"[error] LLM failure → {exc}")
        return "…"

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

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

# ─────────────────────────── 5.  Text 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 with keys:
name, background, appearance, personality, skills_and_abilities,
goals, conflicts, backstory, current_situation,
spoken_lines (list of strings).

Respond with JSON only (no markdown)."""

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

# ───────────────────────── 6.  Helper functions ─────────────────────────────
def random_world() -> str:
    return call_llm(WORLD_PROMPT, max_tokens=120)

@spaces.GPU(duration=75)
def infer_flux(character_json):
    for img 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 img

def generate_character(world_desc: str, persona_desc: str,
                       progress=gr.Progress(track_tqdm=True)):
    raw = call_llm(
        PROMPT_TEMPLATE.format(
            persona_description=persona_desc,
            world_description=world_desc,
        ),
        max_tokens=1024,
    )
    try:
        return json.loads(raw)
    except json.JSONDecodeError:
        # retry once if the model didn’t return valid JSON
        raw = call_llm(
            PROMPT_TEMPLATE.format(
                persona_description=persona_desc,
                world_description=world_desc,
            ),
            max_tokens=1024,
        )
        return json.loads(raw)

# ─────────────────────────── 7.  Gradio UI ──────────────────────────────────
DESCRIPTION = """
* Generates a JSON character sheet from a world + persona.
* Appearance images via **FLUX-dev**; story text via Zephyr-chat or Gemma fallback.
* Personas sampled from **FinePersonas-Lite**.

Tip → Shuffle the world then persona for rapid inspiration.
"""

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

    with gr.Row():
        world_tb = gr.Textbox(label="World Description", lines=10, scale=4)
        persona_tb = gr.Textbox(
            label="Persona Description", value=random_persona(), lines=10, scale=1
        )

    with gr.Row():
        btn_world   = gr.Button("🔄 Random World",   variant="secondary")
        btn_generate = gr.Button("✨ Generate Character", variant="primary", scale=5)
        btn_persona = gr.Button("🔄 Random Persona", variant="secondary")

    with gr.Row():
        img_out  = gr.Image(label="Character Image")
        json_out = gr.JSON(label="Character Description")

    btn_generate.click(
        generate_character, [world_tb, persona_tb], [json_out]
    ).then(
        infer_flux, [json_out], [img_out]
    )

    btn_world.click(random_world, outputs=[world_tb])
    btn_persona.click(random_persona, outputs=[persona_tb])

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