app.py

import spaces
import gradio as gr
import numpy as np
import PIL.Image
from PIL import Image
import random
from diffusers import StableDiffusionXLPipeline
from diffusers import EulerAncestralDiscreteScheduler
import torch
from compel import Compel, ReturnedEmbeddingsType
from huggingface_hub import hf_hub_download

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Make sure to use torch.float16 consistently throughout the pipeline
pipe = StableDiffusionXLPipeline.from_pretrained(
    "lehehroi/ill-14",
    torch_dtype=torch.float16,
    # Explicitly use fp16 variant
    use_safetensors=True  # Use safetensors if available
)

pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.to(device)

# Force all components to use the same dtype
pipe.text_encoder.to(torch.float16)
pipe.text_encoder_2.to(torch.float16)
pipe.vae.to(torch.float16)
pipe.unet.to(torch.float16)

# Initialize Compel for long prompt processing
compel = Compel(
    tokenizer=[pipe.tokenizer, pipe.tokenizer_2],
    text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
    returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
    requires_pooled=[False, True],
    truncate_long_prompts=False
)

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1216

# === LoRA loader ===
# Public HF model repo holding all character LoRAs as <name>.safetensors
LORA_REPO = "lehehroi/imagegen_loras"
_loaded_lora = {"name": None}

def _ensure_lora(lora_name, lora_weight):
    """Reconcile pipeline LoRA state to the requested (name, weight).
    - lora_name "" / "none" / None  -> unload any active LoRA
    - same name as currently loaded -> just update weight
    - different name                 -> unload current, load new
    """
    if not lora_name or lora_name == "none" or str(lora_name).strip() == "":
        if _loaded_lora["name"] is not None:
            pipe.unload_lora_weights()
            _loaded_lora["name"] = None
        return
    if _loaded_lora["name"] == lora_name:
        pipe.set_adapters([lora_name], adapter_weights=[float(lora_weight)])
        return
    if _loaded_lora["name"] is not None:
        pipe.unload_lora_weights()
    path = hf_hub_download(repo_id=LORA_REPO, filename=f"{lora_name}.safetensors")
    pipe.load_lora_weights(path, adapter_name=lora_name)
    pipe.set_adapters([lora_name], adapter_weights=[float(lora_weight)])
    _loaded_lora["name"] = lora_name
# === end LoRA loader ===

# Simple long prompt processing function
def process_long_prompt(prompt, negative_prompt=""):
    """Simple long prompt processing using Compel"""
    try:
        conditioning, pooled = compel([prompt, negative_prompt])
        return conditioning, pooled
    except Exception as e:
        print(f"Long prompt processing failed: {e}, falling back to standard processing")
        return None, None

@spaces.GPU
def infer(prompt, negative_prompt, seed, randomize_seed, width, height,
          guidance_scale, num_inference_steps, lora_name="none", lora_weight=0.85):
    # Reconcile LoRA state to requested name/weight before generation
    _ensure_lora(lora_name, lora_weight)

    use_long_prompt = len(prompt.split()) > 60 or len(prompt) > 300

    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    generator = torch.Generator(device=device).manual_seed(seed)

    try:
        # Try long prompt processing first if prompt is long
        if use_long_prompt:
            print("Using long prompt processing...")
            conditioning, pooled = process_long_prompt(prompt, negative_prompt)

            if conditioning is not None:
                output_image = pipe(
                    prompt_embeds=conditioning[0:1],
                    pooled_prompt_embeds=pooled[0:1],
                    negative_prompt_embeds=conditioning[1:2],
                    negative_pooled_prompt_embeds=pooled[1:2],
                    guidance_scale=guidance_scale,
                    num_inference_steps=num_inference_steps,
                    width=width,
                    height=height,
                    generator=generator
                ).images[0]
                return output_image

        # Fall back to standard processing
        output_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]

        return output_image
    except RuntimeError as e:
        print(f"Error during generation: {e}")
        # Return a blank image with error message
        error_img = Image.new('RGB', (width, height), color=(0, 0, 0))
        return error_img


css = """
#col-container {
    margin: 0 auto;
    max-width: 1024px;
}
"""

with gr.Blocks(css=css) as demo:

    with gr.Column(elem_id="col-container"):

        with gr.Row():
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt (long prompts are automatically supported)",
                container=False,
            )

            run_button = gr.Button("Run", scale=0)

        result = gr.Image(format="png", 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",
                value="monochrome, (low quality, worst quality:1.2), very displeasing, 3d, watermark, signature, ugly, poorly drawn,"
            )

            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )

            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,
                )

                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=MAX_IMAGE_SIZE,
                )

            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.0,
                    maximum=20.0,
                    step=0.1,
                    value=7,
                )

                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=28,
                    step=1,
                    value=28,
                )

            lora_name = gr.Textbox(
                label="LoRA name (filename without .safetensors, or 'none')",
                value="none",
            )

            lora_weight = gr.Slider(
                label="LoRA weight",
                minimum=0.0,
                maximum=1.5,
                step=0.05,
                value=0.85,
            )

    run_button.click(
        fn=infer,
        inputs=[prompt, negative_prompt, seed, randomize_seed, width, height,
                guidance_scale, num_inference_steps, lora_name, lora_weight],
        outputs=[result]
    )

demo.queue().launch()