Spaces:
Paused
Paused
File size: 3,385 Bytes
118bda1 8feabfd 118bda1 |
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 |
from typing import Any
import gradio as gr
import PIL
import spaces
import torch
from hi_diffusers import HiDreamImagePipeline, HiDreamImageTransformer2DModel
from hi_diffusers.schedulers.flash_flow_match import (
FlashFlowMatchEulerDiscreteScheduler,
)
from transformers import AutoTokenizer, LlamaForCausalLM
# Constants
MODEL_PREFIX: str = "HiDream-ai"
LLAMA_MODEL_NAME: str = "meta-llama/Meta-Llama-3.1-8B-Instruct"
MODEL_PATH = "HiDream-ai/HiDream-I1-Dev"
MODEL_CONFIGS: dict[str, Any] = {
"guidance_scale": 0.0,
"num_inference_steps": 28,
"shift": 6.0,
"scheduler": FlashFlowMatchEulerDiscreteScheduler,
}
# Supported image sizes
RESOLUTION_OPTIONS: list[str] = [
"1024 x 1024 (Square)",
"768 x 1360 (Portrait)",
"1360 x 768 (Landscape)",
"880 x 1168 (Portrait)",
"1168 x 880 (Landscape)",
"1248 x 832 (Landscape)",
"832 x 1248 (Portrait)",
]
tokenizer = AutoTokenizer.from_pretrained(LLAMA_MODEL_NAME, use_fast=False)
text_encoder = LlamaForCausalLM.from_pretrained(
LLAMA_MODEL_NAME,
output_hidden_states=True,
output_attentions=True,
torch_dtype=torch.bfloat16,
).to("cuda")
transformer = HiDreamImageTransformer2DModel.from_pretrained(
MODEL_PATH,
subfolder="transformer",
torch_dtype=torch.bfloat16,
).to("cuda")
scheduler = MODEL_CONFIGS["scheduler"](
num_train_timesteps=1000,
shift=MODEL_CONFIGS["shift"],
use_dynamic_shifting=False,
)
pipe = HiDreamImagePipeline.from_pretrained(
MODEL_PATH,
scheduler=scheduler,
tokenizer_4=tokenizer,
text_encoder_4=text_encoder,
torch_dtype=torch.bfloat16,
).to("cuda", torch.bfloat16)
pipe.transformer = transformer
@spaces.GPU(duration=90)
def generate_image(
prompt: str,
resolution: str,
seed: int,
) -> tuple[PIL.Image.Image, int]:
if seed == -1:
seed = torch.randint(0, 1_000_000, (1,)).item()
height, width = tuple(map(int, resolution.replace(" ", "").split("x")))
generator = torch.Generator("cuda").manual_seed(seed)
image = pipe(
prompt=prompt,
height=height,
width=width,
guidance_scale=MODEL_CONFIGS["guidance_scale"],
num_inference_steps=MODEL_CONFIGS["num_inference_steps"],
generator=generator,
).images[0]
torch.cuda.empty_cache()
return image, seed
# Gradio UI
with gr.Blocks(title="HiDream Image Generator") as demo:
gr.Markdown("## 🌈 HiDream Image Generator")
with gr.Row():
with gr.Column():
prompt = gr.Textbox(
label="Prompt",
placeholder="e.g. A futuristic city with floating cars at sunset",
lines=3,
)
resolution = gr.Radio(
choices=RESOLUTION_OPTIONS,
value=RESOLUTION_OPTIONS[0],
label="Resolution",
)
seed = gr.Number(label="Seed (-1 for random)", value=-1, precision=0)
generate_btn = gr.Button("Generate Image", variant="primary")
seed_used = gr.Number(label="Seed Used", interactive=False)
with gr.Column():
output_image = gr.Image(label="Generated Image", type="pil")
generate_btn.click(
fn=generate_image,
inputs=[prompt, resolution, seed],
outputs=[output_image, seed_used],
)
if __name__ == "__main__":
demo.launch()
|