main.py

import os
from fastapi import FastAPI, File, UploadFile, Form
from fastapi.responses import StreamingResponse
import torch
from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler, DPMSolverSinglestepScheduler
from diffusers.pipelines import StableDiffusionInpaintPipeline, StableDiffusionXLInpaintPipeline
from huggingface_hub import hf_hub_download
import numpy as np
import random
from PIL import Image
import io
import os

app = FastAPI()

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

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Load HF token from environment variable
HF_TOKEN = os.getenv("HF_TOKEN")

# Load pipelines
pipe_xl_final = StableDiffusionXLPipeline.from_single_file(
    hf_hub_download(repo_id="fluently/Fluently-XL-Final", filename="FluentlyXL-Final.safetensors", token=HF_TOKEN),
    torch_dtype=torch.float16,
    use_safetensors=True,
)
pipe_xl_final.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe_xl_final.scheduler.config)
pipe_xl_final.to(device)

pipe_anime = StableDiffusionPipeline.from_pretrained(
    "fluently/Fluently-anime",
    torch_dtype=torch.float16,
    use_safetensors=True,
)
pipe_anime.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe_anime.scheduler.config)
pipe_anime.to(device)

pipe_epic = StableDiffusionPipeline.from_pretrained(
    "fluently/Fluently-epic",
    torch_dtype=torch.float16,
    use_safetensors=True,
)
pipe_epic.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe_epic.scheduler.config)
pipe_epic.to(device)



pipe_inpaint = StableDiffusionInpaintPipeline.from_pretrained(
    "fluently/Fluently-v4-inpainting",
    torch_dtype=torch.float16,
    use_safetensors=True,
)
pipe_inpaint.to(device)

pipe_xl = StableDiffusionXLPipeline.from_pretrained(
    "fluently/Fluently-XL-v4",
    torch_dtype=torch.float16,
    use_safetensors=True,
)
pipe_xl.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe_xl.scheduler.config)
pipe_xl.to(device)

pipe_xl_lightning = StableDiffusionXLPipeline.from_pretrained(
    "fluently/Fluently-XL-v3-lightning",
    torch_dtype=torch.float16,
    use_safetensors=True,
)
pipe_xl_lightning.scheduler = DPMSolverSinglestepScheduler.from_config(pipe_xl_lightning.scheduler.config, use_karras_sigmas=False, timestep_spacing="trailing", lower_order_final=True)
pipe_xl_lightning.to(device)


def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    return seed


@app.post("/generate")
async def generate(
    model: str = Form(...),
    prompt: str = Form(...),
    negative_prompt: str = Form(""),
    use_negative_prompt: bool = Form(False),
    seed: int = Form(0),
    width: int = Form(1024),
    height: int = Form(1024),
    guidance_scale: float = Form(3),
    randomize_seed: bool = Form(False),
    inpaint_image: UploadFile = File(None),
    mask_image: UploadFile = File(None),
    blur_factor: float = Form(1.0),
    strength: float = Form(0.75)
):
    seed = int(randomize_seed_fn(seed, randomize_seed))

    if not use_negative_prompt:
        negative_prompt = ""

    inpaint_image_pil = Image.open(io.BytesIO(await inpaint_image.read())) if inpaint_image else None
    mask_image_pil = Image.open(io.BytesIO(await mask_image.read())) if mask_image else None

    if model == "Fluently XL Final":
        images = pipe_xl_final(
            prompt=prompt,
            negative_prompt=negative_prompt,
            width=width,
            height=height,
            guidance_scale=guidance_scale,
            num_inference_steps=25,
            num_images_per_prompt=1,
            output_type="pil",
        ).images
    elif model == "Fluently Anime":
        images = pipe_anime(
            prompt=prompt,
            negative_prompt=negative_prompt,
            width=width,
            height=height,
            guidance_scale=guidance_scale,
            num_inference_steps=30,
            num_images_per_prompt=1,
            output_type="pil",
        ).images
    elif model == "Fluently Epic":
        images = pipe_epic(
            prompt=prompt,
            negative_prompt=negative_prompt,
            width=width,
            height=height,
            guidance_scale=guidance_scale,
            num_inference_steps=30,
            num_images_per_prompt=1,
            output_type="pil",
        ).images
    elif model == "Fluently XL v4":
        images = pipe_xl(
            prompt=prompt,
            negative_prompt=negative_prompt,
            width=width,
            height=height,
            guidance_scale=guidance_scale,
            num_inference_steps=25,
            num_images_per_prompt=1,
            output_type="pil",
        ).images
    elif model == "Fluently XL v3 Lightning":
        images = pipe_xl_lightning(
            prompt=prompt,
            negative_prompt=negative_prompt,
            width=width,
            height=height,
            guidance_scale=2,
            num_inference_steps=5,
            num_images_per_prompt=1,
            output_type="pil",
        ).images
    elif model == "Fluently v4 inpaint" or model == "Fluently XL v3 inpaint":
        blurred_mask = pipe_inpaint.mask_processor.blur(mask_image_pil, blur_factor=blur_factor)
        images = pipe_inpaint(
            prompt=prompt,
            image=inpaint_image_pil,
            mask_image=blurred_mask,
            negative_prompt=negative_prompt,
            width=width,
            height=height,
            guidance_scale=guidance_scale,
            num_inference_steps=30,
            strength=strength,
            num_images_per_prompt=1,
            output_type="pil",
        ).images

    img = images[0]
    img_byte_arr = io.BytesIO()
    img.save(img_byte_arr, format='PNG')
    img_byte_arr.seek(0)

    return StreamingResponse(img_byte_arr, media_type="image/png")


if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)