inference.py

from pathlib import Path
import base64, io, torch
from PIL import Image
from huggingface_hub import hf_hub_download
from enhancer import ESRGANUpscaler, ESRGANUpscalerCheckpoints

# ------------------------------------------------------------------
# 1️⃣  Download all required checkpoints once (cached by HF DL cache)
# ------------------------------------------------------------------
def dl(repo, fname, rev=None):
    return Path(
        hf_hub_download(repo_id=repo, filename=fname, revision=rev)
    )

checkpoints = ESRGANUpscalerCheckpoints(
    # SD-1.5 Multi-Upscaler core
    unet              = dl("refiners/juggernaut.reborn.sd1_5.unet",         "model.safetensors",   "347d14c3c782c4959cc4d1bb1e336d19f7dda4d2"),
    clip_text_encoder = dl("refiners/juggernaut.reborn.sd1_5.text_encoder", "model.safetensors",   "744ad6a5c0437ec02ad826df9f6ede102bb27481"),
    lda               = dl("refiners/juggernaut.reborn.sd1_5.autoencoder",  "model.safetensors",   "3c1aae3fc3e03e4a2b7e0fa42b62ebb64f1a4c19"),
    controlnet_tile   = dl("refiners/controlnet.sd1_5.tile",                "model.safetensors",   "48ced6ff8bfa873a8976fa467c3629a240643387"),

    # ESRGAN 4× super-res
    esrgan            = dl("philz1337x/upscaler",                           "4x-UltraSharp.pth",   "011deacac8270114eb7d2eeff4fe6fa9a837be70"),

    # Negative prompt embedding
    negative_embedding     = dl("philz1337x/embeddings", "JuggernautNegative-neg.pt", "203caa7e9cc2bc225031a4021f6ab1ded283454a"),
    negative_embedding_key = "string_to_param.*",

    # LoRAs
    loras = {
        "more_details": dl("philz1337x/loras", "more_details.safetensors",     "a3802c0280c0d00c2ab18d37454a8744c44e474e"),
        "sdxl_render" : dl("philz1337x/loras", "SDXLrender_v2.0.safetensors",  "a3802c0280c0d00c2ab18d37454a8744c44e474e"),
    },
)

# ------------------------------------------------------------------
# 2️⃣  Instantiate the enhancer once (global singleton)
# ------------------------------------------------------------------
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype  = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32

enhancer = ESRGANUpscaler(checkpoints=checkpoints, device=device, dtype=dtype)

# ------------------------------------------------------------------
# 3️⃣  Hosted-Inference entry-point
# ------------------------------------------------------------------
class EndpointHandler:
    """
    Hugging Face Hosted Inference API entrypoint.
    Expects: {"image": "<BASE64-ENCODED>"}
    Returns: {enhanced_image, original_size, enhanced_size}
    """

    def __init__(self, path="."):
        pass  # all heavy work done globally above

    def __call__(self, inputs: dict) -> dict:
        if "image" not in inputs:
            return {"error": "No image provided"}

        # strip optional data:image/...;base64, prefix
        payload = inputs["image"].split(",", 1)[-1]
        try:
            img = Image.open(io.BytesIO(base64.b64decode(payload))).convert("RGB")
        except Exception as e:
            return {"error": f"Bad image payload: {e}"}

        # run the two-stage upscaler
        out = enhancer.upscale(img)

        buf = io.BytesIO()
        out.save(buf, format="PNG")
        out_b64 = base64.b64encode(buf.getvalue()).decode()

        return {
            "enhanced_image": out_b64,
            "original_size": img.size,
            "enhanced_size": out.size,
        }