File size: 5,283 Bytes
b09c64f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 | import base64
import io
from typing import Any, Dict
import torch
from PIL import Image
from diffusers import (
StableDiffusionXLPipeline,
StableDiffusionXLImg2ImgPipeline,
DPMSolverMultistepScheduler,
)
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
def _decode_image(b64: str) -> Image.Image:
"""Decode a base64 string (optionally a data: URL) into a PIL RGB image."""
if b64.strip().startswith("data:") and "," in b64:
b64 = b64.split(",", 1)[1]
raw = base64.b64decode(b64)
return Image.open(io.BytesIO(raw)).convert("RGB")
def _encode_image(img: Image.Image) -> str:
"""Encode a PIL image as a base64 PNG string."""
buf = io.BytesIO()
img.save(buf, format="PNG")
return base64.b64encode(buf.getvalue()).decode("utf-8")
class EndpointHandler:
"""
Dual-mode SDXL handler for LUSTIFY-v2.0.
Request shape (HF Inference Endpoints):
{
"inputs": "<prompt>",
"parameters": {
"negative_prompt": "...", # optional
"num_inference_steps": 30, # optional
"guidance_scale": 5.0, # optional (author recommends 4-7)
"width": 1024, "height": 1024, # txt2img only
"seed": 12345, # optional, for reproducibility
"image": "<base64>", # PRESENCE switches to img2img
"strength": 0.6 # img2img only (0-1)
}
}
Response: {"image": "<base64 png>", "mode": "txt2img"|"img2img", "parameters": {...}}
"""
def __init__(self, path: str = ""):
# Base text-to-image pipeline. add_watermarker=False avoids the optional
# invisible-watermark dependency.
self.txt2img = StableDiffusionXLPipeline.from_pretrained(
path,
torch_dtype=DTYPE,
use_safetensors=True,
add_watermarker=False,
)
# DPM++ 2M SDE Karras — the checkpoint author's recommended sampler.
self.txt2img.scheduler = DPMSolverMultistepScheduler.from_config(
self.txt2img.scheduler.config,
algorithm_type="sde-dpmsolver++",
use_karras_sigmas=True,
)
self.txt2img.to(DEVICE)
# img2img reuses the exact same weights/components — no extra VRAM cost.
self.img2img = StableDiffusionXLImg2ImgPipeline(**self.txt2img.components)
self.img2img.to(DEVICE)
if DEVICE == "cuda":
self.txt2img.enable_vae_slicing()
try:
self.txt2img.enable_xformers_memory_efficient_attention()
self.img2img.enable_xformers_memory_efficient_attention()
except Exception:
# xformers is optional; the pipelines run fine without it.
pass
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
prompt = data.get("inputs") or data.get("prompt")
params = data.get("parameters") or {}
if not prompt:
return {"error": "No prompt provided. Send {'inputs': '<prompt>'}."}
negative_prompt = params.get("negative_prompt")
num_inference_steps = int(params.get("num_inference_steps", 30))
guidance_scale = float(params.get("guidance_scale", 5.0))
width = int(params.get("width", 1024))
height = int(params.get("height", 1024))
seed = params.get("seed")
generator = None
if seed is not None:
generator = torch.Generator(device=DEVICE).manual_seed(int(seed))
init_b64 = params.get("image")
strength = float(params.get("strength", 0.6))
try:
if init_b64:
init_image = _decode_image(init_b64)
result = self.img2img(
prompt=prompt,
negative_prompt=negative_prompt,
image=init_image,
strength=strength,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
generator=generator,
)
mode = "img2img"
else:
result = self.txt2img(
prompt=prompt,
negative_prompt=negative_prompt,
width=width,
height=height,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
generator=generator,
)
mode = "txt2img"
except Exception as e:
return {
"error": f"{type(e).__name__}: {e}",
"mode": "img2img" if init_b64 else "txt2img",
}
image = result.images[0]
return {
"image": _encode_image(image),
"mode": mode,
"parameters": {
"num_inference_steps": num_inference_steps,
"guidance_scale": guidance_scale,
"strength": strength if init_b64 else None,
"width": width,
"height": height,
"seed": int(seed) if seed is not None else None,
},
}
|