How to sample in image space with checkpoints trained using prediction="score" or prediction="noise"?

[Graeme-Lee]

Problem Description

I have trained a SiT model on medical images in image space (not latent space) using the --prediction noise setting. When attempting to sample from the trained checkpoint, the generated images do not make any sense.

Sampling command:

python sample.py ODE --prediction noise

Current Sampling Code

import torch
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
from torchvision.utils import save_image
from diffusers.models import AutoencoderKL
from download import find_model
from models import SiT_models
from train_utils import parse_ode_args, parse_sde_args, parse_transport_args
from transport import create_transport, Sampler
import argparse
import sys
from time import time


def main(mode, args):
    # Setup PyTorch:
    torch.manual_seed(args.seed)
    torch.set_grad_enabled(False)
    device = "cuda" if torch.cuda.is_available() else "cpu"

    if args.ckpt is None:
        assert args.model == "SiT-XL/2", "Only SiT-XL/2 models are available for auto-download."
        assert args.image_size in [256, 512]
        assert args.num_classes == 1000
        assert args.image_size == 256, "512x512 models are not yet available for auto-download."
        learn_sigma = args.image_size == 256
    else:
        learn_sigma = False

    # Load model:
    latent_size = args.image_size // 1  # <- Is this correct for image-space training?
    model = SiT_models[args.model](
        input_size=latent_size,
        num_classes=args.num_classes,
        learn_sigma=learn_sigma,
    ).to(device)
    
    ckpt_path = args.ckpt or f"SiT-XL-2-{args.image_size}x{args.image_size}.pt"
    state_dict = find_model(ckpt_path)
    model.load_state_dict(state_dict)
    model.eval()
    
    transport = create_transport(
        args.path_type,
        args.prediction,
        args.loss_weight,
        args.train_eps,
        args.sample_eps
    )
    sampler = Sampler(transport)
    
    if mode == "ODE":
        if args.likelihood:
            assert args.cfg_scale == 1, "Likelihood is incompatible with guidance"
            sample_fn = sampler.sample_ode_likelihood(
                sampling_method=args.sampling_method,
                num_steps=args.num_sampling_steps,
                atol=args.atol,
                rtol=args.rtol,
            )
        else:
            sample_fn = sampler.sample_ode(
                sampling_method=args.sampling_method,
                num_steps=args.num_sampling_steps,
                atol=args.atol,
                rtol=args.rtol,
                reverse=args.reverse
            )
            
    elif mode == "SDE":
        sample_fn = sampler.sample_sde(
            sampling_method=args.sampling_method,
            diffusion_form=args.diffusion_form,
            diffusion_norm=args.diffusion_norm,
            last_step=args.last_step,
            last_step_size=args.last_step_size,
            num_steps=args.num_sampling_steps,
        )
    
    # Labels to condition the model with:
    class_labels = [0, 1]
    
    # Create sampling noise:
    n = len(class_labels)
    z = torch.randn(n, 1, latent_size, latent_size, device=device)  # <- Is 1 channel correct?
    y = torch.tensor(class_labels, device=device)

    # Setup classifier-free guidance:
    use_cfg = args.cfg_scale > 1.0
    if use_cfg:
        zs = torch.cat([z, z], 0)
        y_null = torch.tensor([1000] * n, device=device)
        ys = torch.cat([y, y_null], 0)
        model_kwargs = dict(y=ys, cfg_scale=args.cfg_scale)
        model_fn = model.forward_with_cfg
    else:
        model_kwargs = dict(y=y)
        model_fn = model.forward

    # Sample images:
    start_time = time()
    samples = sample_fn(z, model_fn, **model_kwargs)[-1]
    print(f"Sampling took {time() - start_time:.2f} seconds.")

    # Save images (VAE is loaded but never used for decoding):
    save_image(samples, f"sample.png", nrow=4, normalize=True, value_range=(0, 1))
if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    if len(sys.argv) < 2:
        print("Usage: program.py <mode> [options]")
        sys.exit(1)
    
    mode = sys.argv[1]

    assert mode[:2] != "--", "Usage: program.py <mode> [options]"
    assert mode in ["ODE", "SDE"], "Invalid mode. Please choose 'ODE' or 'SDE'"
    
    parser.add_argument("--model", type=str, choices=list(SiT_models.keys()), default="SiT-B/2")
    parser.add_argument("--vae", type=str, choices=["ema", "mse"], default="mse")
    parser.add_argument("--image-size", type=int, choices=[256, 512], default=256)
    parser.add_argument("--num-classes", type=int, default=1000)
    parser.add_argument("--cfg-scale", type=float, default=1.0)
    parser.add_argument("--num-sampling-steps", type=int, default=1000)
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument("--ckpt", type=str, default="/home/exouser/SiT/results/002-SiT-B-2-Linear-noise-None/checkpoints/0072000.pt",
                        help="Optional path to a SiT checkpoint (default: auto-download a pre-trained SiT-XL/2 model).")


    parse_transport_args(parser)
    if mode == "ODE":
        parse_ode_args(parser)
        # Further processing for ODE
    elif mode == "SDE":
        parse_sde_args(parser)
        # Further processing for SDE
    
    args = parser.parse_known_args()[0]
    main(mode, args)

Specific Questions

1. VAE usage with image-space training:

   • The pre-trained ImageNet examples seem to use latent diffusion with VAE, but I trained directly on images.

2. Latent size configuration:

   • I'm using latent_size = args.image_size // 1, making them equal. Is this correct for image-space training?

3. Transport parameters alignment:

   • How do I ensure --path_type, --prediction, --train_eps, and --sample_eps match between training and sampling?
   • Are there any other critical hyperparameters that must match?

4. Output value range:

   • What range should the sampled values be in before saving?
   • I'm using normalize=True, value_range=(0, 1) but unsure if this is correct.

Training Configuration

• Checkpoint: Trained with --prediction noise
• Training space: Image space (no VAE)
• Dataset: Medical imaging (grayscale/range [0,1], float32)
• Image size: 256x256
• Model: SiT-B/2
• Number of classes: 2

Thank you for your help!