SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2310.02557 · ICLR · 2024

Generalization in diffusion models arises from geometry-adaptive harmonic representation

Stéphane Mallat, Eero Simoncelli, Florentin Guth, Zahra Kadkhodaie

arXiv · PDF · Open in the Atlas

Code that ran

11 of 16 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 7 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 4 failures of our instrument; 5 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2310.02557.json · badge: /api/ran/2310.02557.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
LabForComputationalVision/memorization_generalization_in_diffusion_models canonical (classifier label)
classifier basis: census
11 ran · 5 no recorded run
FunctionStatusWhere it lives
MSE Ran
ran, honoured the contract we drafted
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/quality_metrics_func.py
code served (permissive licence) · get_code("eb15c404ffac080f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# MSE function takes two image tensors and computes mean squared error
# Both parameters should be tensors of the same shape
# Using small batch size of 2, with 3 channels and spatial dimensions 4x4

ref_im = torch.randn(2, 3, 4, 4)
im = torch.randn(2, 3, 4, 4)

ARGS = {
    "ref_im": ref_im,
    "im": im,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
batch_psnr_numpy Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/quality_metrics_func.py
code served (permissive licence) · get_code("1aa782be31b7fc00")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 1, dtype family float, shape (2,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create small batch of reference images (N=2, W=4, H=3)
ref_ims_3d = np.random.rand(2, 4, 3).astype(np.float32)

# Create corresponding images to compare
images_3d = np.random.rand(2, 4, 3).astype(np.float32)

# Max intensity value
max_I = 1.0

ARGS = {
    "ref_ims": ref_ims_3d,
    "images": images_3d,
    "max_I": max_I,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (2,) — shape (2,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
calc_jacobian Ran
ran; our contract draft was wrong, not the code
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/linear_approx.py
code served (permissive licence) · get_code("d0324c885182c509")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (12, 12) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(1, 1, kernel_size=3, padding=1)
    
    def forward(self, x, layer_num=None, channel_num=None):
        # Model accepts optional layer_num and channel_num parameters
        # but we'll just apply convolution
        out = self.conv1(x)
        return out

# Create a small model
model = SimpleModel()

# Create input tensor: batch_size=2, channels=1, height=4, width=3
# The function accesses inp.size()[2] and inp.size()[3], so we need at least 4D tensor
inp = torch.randn(2, 1, 4, 3, requires_grad=True)

ARGS = {
    "inp": inp,
    "model": model,
    "layer_num": None,
    "channel_num": None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (12, 12) — shape (12, 12) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
calc_jacobian_rows Ran
ran; our contract draft was wrong, not the code
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/linear_approx.py
code served (permissive licence) · get_code("98dacb3ad56237ba")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (4, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create a minimal model that outputs a 4D tensor
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = nn.Conv2d(1, 1, kernel_size=3, padding=1)
    
    def forward(self, x):
        return self.conv(x)

# Create input tensor: batch_size=1, channels=1, height=4, width=3
inp = torch.randn(1, 1, 4, 3, requires_grad=False)

# Create model
model = SimpleModel()

# Indices for accessing output tensor dimensions [0, 0, i, j]
i = 2
j = 1

ARGS = {
    "inp": inp,
    "model": model,
    "i": i,
    "j": j
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (4, 3) — shape (4, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_mnist Ran
ran; our contract draft was wrong, not the code
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/dataloader_func.py
code served (permissive licence) · get_code("6ad8ccbe18be47c9")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import torch
import gzip
import os
import tempfile

torch.manual_seed(0)

# Create a temporary directory with mock MNIST data files
temp_dir = tempfile.mkdtemp()

# Create mock train-images file
train_images_path = os.path.join(temp_dir, 'train-images-idx3-ubyte.gz')
with gzip.open(train_images_path, 'wb') as f:
    # Write 16-byte header
    f.write(b'\x00' * 16)
    # Write mock image data (28*28*60000 bytes)
    mock_train_data = np.random.randint(0, 256, size=28*28*60000, dtype=np.uint8)
    f.write(mock_train_data.tobytes())

# Create mock test-images file
test_images_path = os.path.join(temp_dir, 't10k-images-idx3-ubyte.gz')
with gzip.open(test_images_path, 'wb') as f:
    # Write 16-byte header
    f.write(b'\x00' * 16)
    # Write mock image data (28*28*10000 bytes)
    mock_test_data = np.random.randint(0, 256, size=28*28*10000, dtype=np.uint8)
    f.write(mock_test_data.tobytes())

ARGS = {
    'data_dir_path': temp_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_set14 Ran
ran, honoured the contract we drafted
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/dataloader_func.py
code served (permissive licence) · get_code("31ac25abb6e95520")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 4, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import matplotlib.pyplot as plt
import os
import tempfile

# Create a temporary directory with test image files
temp_dir = tempfile.mkdtemp()

# Create a simple test image and save it
test_image = np.random.rand(10, 10, 3)
plt.imsave(os.path.join(temp_dir, 'test_image_1.png'), test_image)

test_image_2 = np.random.rand(10, 10, 3)
plt.imsave(os.path.join(temp_dir, 'test_image_2.png'), test_image_2)

# Ensure path ends with separator for proper concatenation in the function
path = temp_dir + os.sep

ARGS = {
    'path': path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 4, dtype family float — no output shape is stated: observe_signature.py could not observe it (read:AttributeError)
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_fft_filter_1d Ran
ran, honoured the contract we drafted
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/synthetic_data_generators.py
code served (permissive licence) · get_code("6eb2528b9343169a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 1, dtype family float, shape (16,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# The function make_fft_filter_1d takes a single parameter N (an integer)
# It computes FFT frequency filtering for a 1D signal of length N
# N should be a positive integer representing the size of the FFT

N = 16

ARGS = {
    "N": N
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 1, dtype family float, shape (16,) — shape (16,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_fft_filter_2d Ran
ran, honoured the contract we drafted
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/synthetic_data_generators.py
code served (permissive licence) · get_code("faa53db8ba9db059")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (8, 8) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    "N": 8
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (8, 8) — shape (8, 8) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_fft_filter_2d_seprable Ran
ran, honoured the contract we drafted
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/synthetic_data_generators.py
code served (permissive licence) · get_code("56aaa6b239a0ae60")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (8, 8) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    "N": 8
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (8, 8) — shape (8, 8) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sig_to_psnr Ran
ran, honoured the contract we drafted
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/quality_metrics_func.py
code served (permissive licence) · get_code("7ff802a286a451e5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float64, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# The function sig_to_psnr takes:
# - std: a standard deviation value (numeric)
# - I_max: maximum intensity value (default 255)

# Looking at the function body:
# - std is used in arithmetic operations: (std / 255) ** 2
# - I_max has a default value of 255 and is used as a divisor
# - The function computes -10 * np.log10((std / 255) ** 2)

# Create valid arguments
std_value = 10.0  # A reasonable standard deviation value
I_max_value = 255  # The default maximum intensity for 8-bit images

ARGS = {
    "std": std_value,
    "I_max": I_max_value,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
single_image_loader Ran
ran, honoured the contract we drafted
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/dataloader_func.py
code served (permissive licence) · get_code("bf7b787bf71b1055")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 3, dtype family float, shape (1, 28, 28) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
import tempfile
import os
import gzip

torch.manual_seed(0)

# Create a temporary directory with mock MNIST data
temp_dir = tempfile.mkdtemp()
mnist_dir = os.path.join(temp_dir, 'mnist')
os.makedirs(mnist_dir, exist_ok=True)

# Create a mock MNIST gzip file with the expected format
mnist_file_path = os.path.join(mnist_dir, 't10k-images-idx3-ubyte.gz')
with gzip.open(mnist_file_path, 'wb') as f:
    # Write 16-byte header
    f.write(b'\x00' * 16)
    # Write 10000 images of 28x28 pixels (10000 * 28 * 28 = 7,840,000 bytes)
    image_data = np.random.randint(0, 256, size=(10000, 28, 28), dtype=np.uint8).tobytes()
    f.write(image_data)

# Construct the arguments
ARGS = {
    'data_set_dire_path': mnist_dir,
    'image_number': 5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float, shape (1, 28, 28) — shape (1, 28, 28) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
calc_jacobian_row_MS Not yet run
unverified (harvested, no recorded run)
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/linear_approx.py
code served (permissive licence) · get_code("53ef977c62a678b6")
How it ran

no invocation recorded.

init_BF_CNN_RF Not yet run
unverified (harvested, no recorded run)
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/model_loader_func.py
code served (permissive licence) · get_code("561901e6303b8afc")
How it ran

no invocation recorded.

init_UNet Not yet run
unverified (harvested, no recorded run)
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/model_loader_func.py
code served (permissive licence) · get_code("0a2d1378af0b3c3c")
How it ran

no invocation recorded.

initialize_network Not yet run
unverified (harvested, no recorded run)
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/model_loader_func.py
code served (permissive licence) · get_code("706dea6429c294c3")
How it ran

no invocation recorded.

univ_inv_sol Not yet run
unverified (harvested, no recorded run)
LabForComputationalVision/memorization_generalization_in_diffusion_models/code/algorithm_inv_prob.py
code served (permissive licence) · get_code("4835e614b7b50339")
How it ran

no invocation recorded.

Also present in other papers' code

Of 16 samples lifted from this paper's linked repositories, 16 carry a function identity and 0 do not. Of the 16 distinct functions with an identity, 4 are also present in other papers whose code we hold (4 with at least one paper dated earlier, 0 the same month, 0 later) and 12 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
MSE 1 other paper
1 dated earlier
2303.02984 (2023-03)
Learning multi-scale local conditional probability models of images
calc_jacobian_row_MS 1 other paper
1 dated earlier
2303.02984 (2023-03)
Learning multi-scale local conditional probability models of images
calc_jacobian_rows 1 other paper
1 dated earlier
2303.02984 (2023-03)
Learning multi-scale local conditional probability models of images
load_set14 1 other paper
1 dated earlier
2303.02984 (2023-03)
Learning multi-scale local conditional probability models of images

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Deep neural networks (DNNs) trained for image denoising are able to generate highquality samples with score-based reverse diffusion algorithms. These impressive capabilities seem to imply an escape from the curse of dimensionality, but recent reports of memorization of the training set raise the question of whether these networks are learning the "true" continuous density of the data. Here, we show that two DNNs trained on non-overlapping subsets of a dataset learn nearly the same score function, and thus the same density, when the number of training images is large enough. In this regime of strong generalization, diffusion-generated images are distinct from the training set, and are of high visual quality, suggesting that the inductive biases of the DNNs are well-aligned with the data density. We analyze the learned denoising functions and show that the inductive biases give rise to a shrinkage operation in a basis adapted to the underlying image. Examination of these bases reveals oscillating harmonic structures along contours and in homogeneous regions. We demonstrate that trained denoisers are inductively biased towards these geometry-adaptive harmonic bases since they arise not only when the network is trained on photographic images, but also when it is trained on image classes supported on low-dimensional manifolds for which the harmonic basis is suboptimal. Finally, we show that when trained on regular image classes for which the optimal basis is known to be geometry-adaptive and harmonic, the denoising performance of the networks is near-optimal.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2310.02557")
get_code_for_paper("2310.02557")
have("2310.02557")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2310.02557.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2310.02557.svg)](https://syntology.ai/paper/2310.02557)

Connect an agent — have() is free.