Spaces:
Paused
Paused
File size: 5,250 Bytes
755bbb7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 |
import gradio as gr
import spaces
import torch
from hi_diffusers import HiDreamImagePipeline, HiDreamImageTransformer2DModel
from hi_diffusers.schedulers.flash_flow_match import (
FlashFlowMatchEulerDiscreteScheduler,
)
from hi_diffusers.schedulers.fm_solvers_unipc import FlowUniPCMultistepScheduler
from transformers import LlamaForCausalLM, PreTrainedTokenizerFast
# Constants
MODEL_PREFIX: str = "HiDream-ai"
LLAMA_MODEL_NAME: str = "meta-llama/Meta-Llama-3.1-8B-Instruct"
# Model configurations
MODEL_CONFIGS: dict[str, dict] = {
"dev": {
"path": f"{MODEL_PREFIX}/HiDream-I1-Dev",
"guidance_scale": 0.0,
"num_inference_steps": 28,
"shift": 6.0,
"scheduler": FlashFlowMatchEulerDiscreteScheduler,
},
"full": {
"path": f"{MODEL_PREFIX}/HiDream-I1-Full",
"guidance_scale": 5.0,
"num_inference_steps": 50,
"shift": 3.0,
"scheduler": FlowUniPCMultistepScheduler,
},
"fast": {
"path": f"{MODEL_PREFIX}/HiDream-I1-Fast",
"guidance_scale": 0.0,
"num_inference_steps": 16,
"shift": 3.0,
"scheduler": FlashFlowMatchEulerDiscreteScheduler,
},
}
# Supported image sizes
RESOLUTION_OPTIONS: list[str] = [
"1024 × 1024 (Square)",
"768 × 1360 (Portrait)",
"1360 × 768 (Landscape)",
"880 × 1168 (Portrait)",
"1168 × 880 (Landscape)",
"1248 × 832 (Landscape)",
"832 × 1248 (Portrait)",
]
# Model cache
loaded_models: dict[str, HiDreamImagePipeline] = {}
def parse_resolution(res_str: str) -> tuple[int, int]:
"""Parse resolution string like '1024 × 1024' into (1024, 1024)"""
return tuple(map(int, res_str.replace("×", "x").replace(" ", "").split("x")))
def load_models(model_type: str) -> HiDreamImagePipeline:
"""Load and initialize the HiDream model pipeline for a given model type."""
config = MODEL_CONFIGS[model_type]
pretrained_model = config["path"]
tokenizer = PreTrainedTokenizerFast.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(
pretrained_model,
subfolder="transformer",
torch_dtype=torch.bfloat16,
).to("cuda")
scheduler = config["scheduler"](
num_train_timesteps=1000,
shift=config["shift"],
use_dynamic_shifting=False,
)
pipe = HiDreamImagePipeline.from_pretrained(
pretrained_model,
scheduler=scheduler,
tokenizer_4=tokenizer,
text_encoder_4=text_encoder,
torch_dtype=torch.bfloat16,
).to("cuda", torch.bfloat16)
pipe.transformer = transformer
return pipe
# Preload default model
print("🔧 Preloading default model (full)...")
loaded_models["full"] = load_models("full")
print("✅ Model loaded.")
@spaces.GPU(duration=90)
def generate_image(
model_type: str,
prompt: str,
resolution: str,
seed: int,
) -> tuple[object, int]:
"""Generate image using HiDream pipeline."""
if model_type not in loaded_models:
print(f"📦 Lazy-loading model {model_type}...")
loaded_models[model_type] = load_models(model_type)
pipe: HiDreamImagePipeline = loaded_models[model_type]
config = MODEL_CONFIGS[model_type]
if seed == -1:
seed = torch.randint(0, 1_000_000, (1,)).item()
height, width = parse_resolution(resolution)
generator = torch.Generator("cuda").manual_seed(seed)
image = pipe(
prompt=prompt,
height=height,
width=width,
guidance_scale=config["guidance_scale"],
num_inference_steps=config["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():
model_type = gr.Radio(
choices=list(MODEL_CONFIGS.keys()),
value="full",
label="Model Type",
info="Choose between full, fast or dev variants",
)
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=[model_type, prompt, resolution, seed],
outputs=[output_image, seed_used],
)
if __name__ == "__main__":
demo.launch()
|