"""Generate a 512x512 image with EchoVision 1.5 Pro.""" import argparse from pathlib import Path import torch from diffusers import AutoencoderKL, DDIMScheduler, UNet2DConditionModel from safetensors.torch import load_file from transformers import CLIPTextModel, CLIPTokenizer from PIL import Image TEXT_ID = "openai/clip-vit-base-patch32" VAE_ID = "stabilityai/sd-vae-ft-mse" LATENT_SCALE = 0.18215 @torch.inference_mode() def main(): parser = argparse.ArgumentParser(description="EchoVision 1.5 Pro inference") parser.add_argument("--model-dir", default=".", help="Directory with config.json and model.safetensors") parser.add_argument("--prompt", default="A wooden cabin beside a mountain lake at sunrise") parser.add_argument("--negative-prompt", default="", help="Optional classifier-free guidance negative prompt") parser.add_argument("--steps", type=int, default=25) parser.add_argument("--guidance", type=float, default=7.5) parser.add_argument("--seed", type=int, default=42) parser.add_argument("--output", default="output.png") args = parser.parse_args() if not 1 <= args.steps <= 100: parser.error("--steps must be between 1 and 100") if not 0 <= args.guidance <= 20: parser.error("--guidance must be between 0 and 20") device = torch.device("cuda" if torch.cuda.is_available() else "cpu") dtype = torch.float16 if device.type == "cuda" else torch.float32 root = Path(args.model_dir) print(f"Loading EchoVision from {root.resolve()} on {device}...") unet = UNet2DConditionModel.from_config(root / "config.json") state = load_file(str(root / "model.safetensors"), device="cpu") unet.load_state_dict(state, strict=True) del state unet.to(device=device, dtype=dtype).eval() tokenizer = CLIPTokenizer.from_pretrained(TEXT_ID) text_encoder = CLIPTextModel.from_pretrained(TEXT_ID).eval().requires_grad_(False).to("cpu") vae = AutoencoderKL.from_pretrained(VAE_ID).eval().requires_grad_(False).to(device=device, dtype=dtype) def embed(text): tokens = tokenizer([text], padding="max_length", truncation=True, max_length=77, return_tensors="pt") # Text encoder is kept on CPU, matching the original EchoVision sampling code. return text_encoder(input_ids=tokens.input_ids, attention_mask=tokens.attention_mask).last_hidden_state.to(device=device, dtype=dtype) uncond = embed(args.negative_prompt) cond = embed(args.prompt) embeddings = torch.cat([uncond, cond], dim=0) generator = torch.Generator(device=device).manual_seed(args.seed) latents = torch.randn((1, 4, 64, 64), device=device, dtype=dtype, generator=generator) scheduler = DDIMScheduler( num_train_timesteps=1000, beta_start=0.0001, beta_end=0.02, beta_schedule="scaled_linear", prediction_type="epsilon", clip_sample=False, set_alpha_to_one=True, steps_offset=0, ) scheduler.set_timesteps(args.steps, device=device) for timestep in scheduler.timesteps: model_input = torch.cat([latents, latents], dim=0) if device.type == "cuda": with torch.autocast(device_type="cuda", dtype=torch.float16): prediction = unet(model_input, timestep, encoder_hidden_states=embeddings).sample else: prediction = unet(model_input, timestep, encoder_hidden_states=embeddings).sample prediction_uncond, prediction_cond = prediction.chunk(2) guided = prediction_uncond + args.guidance * (prediction_cond - prediction_uncond) latents = scheduler.step(guided, timestep, latents).prev_sample if device.type == "cuda": with torch.autocast(device_type="cuda", dtype=torch.float16): pixels = vae.decode(latents / LATENT_SCALE).sample else: pixels = vae.decode(latents / LATENT_SCALE).sample pixels = ((pixels / 2 + 0.5).clamp(0, 1)[0].permute(1, 2, 0) * 255).round().byte().cpu().numpy() Image.fromarray(pixels).save(args.output) print(f"Saved {args.output}") if __name__ == "__main__": main()