SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2503.23367 · ICCV · 2025

FastVAR: Linear Visual Autoregressive Modeling via Cached Token Pruning

Shu-Tao Xia, Yawei Li, Luca Benini, Hang Guo, Tao Dai, Jiangshan Wang, Taolin Zhang

arXiv · PDF · Open in the Atlas

Code that ran

11 of 14 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; 0 honoured the contract we drafted, 0 violated it and 8 have no contract verdict; 3 failures of our instrument; 3 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/2503.23367.json · badge: /api/ran/2503.23367.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
csguoh/fastvar canonical (classifier label)
classifier basis not recorded on the link
11 ran · 3 no recorded run
FunctionStatusWhere it lives
apply_rotary_emb Ran
ran; our contract draft was wrong, not the code
csguoh/fastvar/Infinity/infinity/models/basic.py
code served (permissive licence) · get_code("42125baaa5c021b6")
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, ndim 5, 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 torch
import numpy as np

torch.manual_seed(0)

# Construct tiny tensors for q and k
# Shape: (batch, heads, seq_len, dim)
batch_size = 2
num_heads = 3
seq_len = 4
dim = 8

q = torch.randn(batch_size, num_heads, seq_len, dim)
k = torch.randn(batch_size, num_heads, seq_len, dim)

# scale_schedule: list of tuples, each with 3 elements (used in the function)
# The function computes: start = sum([item[0] * item[1] * item[2] for item in scale_schedule[:scale_ind]])
scale_schedule = ((2, 3, 4), (5, 6, 7))

# rope2d_freqs_grid: dict with string keys mapping to tensors
# Key is str(tuple(scale_schedule))
# Tensor shape should support: rope_cache = rope2d_freqs_grid[key][:, :, :, :, start:start + seq_len]
# So it needs shape (?, ?, ?, ?, large_seq_dim)
# The function reshapes qk from (..., seq_len, dim) to (..., seq_len, -1, 2)
# After rope application, it reshapes back to (..., seq_len, dim)
# rope_cache is indexed as [0] and [1] for cos and sin components
# Shape: (2, batch, heads, ?, total_seq_len) where total_seq_len >= seq_len

rope2d_freqs_grid = {
    str(tuple(scale_schedule)): torch.randn(2, batch_size, num_heads, 1, 20)
}

# pad_to_multiplier: likely an integer or not used in the function body
# Looking at the function, pad_to_multiplier is never referenced
pad_to_multiplier = 8

# rope2d_normalized_by_hw: likely a boolean or not used in the function body
# Looking at the function, rope2d_normalized_by_hw is never referenced
rope2d_normalized_by_hw = True

# scale_ind: integer index into scale_schedule
scale_ind = 0

ARGS = {
    "q": q,
    "k": k,
    "scale_schedule": scale_schedule,
    "rope2d_freqs_grid": rope2d_freqs_grid,
    "pad_to_multiplier": pad_to_multiplier,
    "rope2d_normalized_by_hw": rope2d_normalized_by_hw,
    "scale_ind": scale_ind,
}
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, ndim 5, dtype family float — a tuple output has no array shape of its own; none is stated
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
compute_merge Ran
ran on a synthesized input
csguoh/fastvar/Infinity/infinity/models/fastvar_utils.py
code served (permissive licence) · get_code("bb07b7c6bc141e2d")
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, ndim 1, 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 torch

torch.manual_seed(0)

# The function `compute_merge` takes:
# - x: torch.Tensor
# - prune_scale_list: list (default [32, 40])
# - is_later_layer: bool (default False)
# - x_shape: tuple or list with shape info

# Looking at the function body:
# - x is used with x.shape[1] to get sequence length
# - x_shape is unpacked as (_, original_h, original_w)
# - x_shape is used to compute original_tokens and check original_w
# - Inside masked_previous_scale_cache, cur_x is reshaped using cur_shape

# Create a small tensor: batch_size=2, seq_len=5, channels=4
x = torch.randn(2, 5, 4)

# x_shape should be (channels, height, width) based on how it's used in masked_previous_scale_cache
# We use (4, 3, 2) for distinct dimensions
x_shape = (4, 3, 2)

# Standard parameters
prune_scale_list = [32, 40]
is_later_layer = False

ARGS = {
    "x": x,
    "prune_scale_list": prune_scale_list,
    "is_later_layer": is_later_layer,
    "x_shape": x_shape,
}
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 tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
compute_merge Ran
ran on a synthesized input
csguoh/fastvar/HART/hart/modules/networks/fastvar_utils.py
code served (permissive licence) · get_code("974751f30ae5dccd")
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, ndim 1, 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 torch
import math

torch.manual_seed(0)

# Create a small tensor with shape (B, L, c) where L is a perfect square
# B=2 (batch), L=16 (4x4 spatial), c=4 (channels)
x = torch.randn(2, 16, 4)

# ratio parameter - controls token pruning ratio
ratio = 0.5

# prune_scale_list - list of scales where hard-coded ratios apply
prune_scale_list = [4, 8]

# is_later_layer - boolean flag for whether this is a later layer
is_later_layer = False

ARGS = {
    "x": x,
    "ratio": ratio,
    "prune_scale_list": prune_scale_list,
    "is_later_layer": is_later_layer,
}
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 tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
do_nothing Ran
ran on a synthesized input
csguoh/fastvar/Infinity/infinity/models/fastvar_utils.py
code served (permissive licence) · get_code("942c492abe601d24")
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 (2, 4) 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 do_nothing takes a torch.Tensor x and accepts *args and **kwargs
# It simply returns x without using any of the additional arguments

# Create a small tensor (batch=2, dim=4)
x = torch.randn(2, 4)

# *args can be empty or contain additional positional arguments
args_tuple = ()

# **kwargs can be empty or contain additional keyword arguments
kwargs_dict = {}

# Map all parameters to their values
ARGS = {
    "x": x,
    "args": args_tuple,
    "kwargs": kwargs_dict,
}
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 (2, 4) — shape (2, 4) 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
get_dropout_layer Ran
ran on a synthesized input
csguoh/fastvar/Infinity/infinity/models/fastvar_basic.py
code served (permissive licence) · get_code("b4a8295838aac5c1")
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 Dropout. 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 torch
import torch.nn as nn

torch.manual_seed(0)

# The function takes a single parameter 'p' which is a dropout probability
# It should be a float between 0 and 1
# We'll test with a value > 0 to trigger the Dropout case

ARGS = {
    "p": 0.5
}
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 Dropout — a Dropout output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Dropout output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not 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
get_position_ids Ran
ran on a synthesized input
csguoh/fastvar/HART/hart/modules/networks/fastvar_basic.py
code served (permissive licence) · get_code("9bd25e1e6d3188a3")
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 (2, 336, 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 torch
import functools

torch.manual_seed(0)

# Analyze the function:
# - batch_size: int, used in repeat()
# - patch_nums: tuple/list of ints, indexed with si and iterated over
# - device: torch.device, passed to torch.arange()
# - si: int, default -1, used as index/slice into patch_nums
# - m_maskgit: None or tensor used for indexing pos_ids

# Construct arguments:
batch_size = 2
patch_nums = (4, 8, 16)  # tuple of patch numbers at different levels
device = torch.device('cpu')
si = -1  # use all patch_nums
m_maskgit = None  # optional masking

ARGS = {
    'batch_size': batch_size,
    'patch_nums': patch_nums,
    'device': device,
    'si': si,
    'm_maskgit': m_maskgit
}
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 3, dtype family float, shape (2, 336, 2) — shape (2, 336, 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
get_position_ids_1d Ran
ran on a synthesized input
csguoh/fastvar/HART/hart/modules/networks/fastvar_basic.py
code served (permissive licence) · get_code("82c722b23631ab21")
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 int, shape (2, 4) 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 functools

torch.manual_seed(0)

def get_position_ids_1d(batch_size, L, device):
    return torch.arange(L, device=device).unsqueeze(0).repeat(batch_size, 1)

# Construct arguments
batch_size = 2
L = 4
device = 'cpu'

ARGS = {
    'batch_size': batch_size,
    'L': L,
    'device': device
}
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 int, shape (2, 4) — shape (2, 4) 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
masked_previous_scale_cache Ran
ran on a synthesized input
csguoh/fastvar/Infinity/infinity/models/fastvar_utils.py
code served (permissive licence) · get_code("a2661ae675e65815")
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, ndim 1, 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 torch

torch.manual_seed(0)

# Analyzing the function:
# - cur_x: tensor with shape (B, L, c), used in view/permute/gather/scatter operations
# - num_remain: int, used for slicing select_indices
# - cur_shape: tuple/list with at least [_, H, W] indices, used for view and interpolate size

# The function expects:
# - cur_x.shape = (B, L, c)
# - cur_shape[1] and cur_shape[2] are spatial dimensions (H, W)
# - cur_x.view(B, cur_shape[1], cur_shape[2], -1) must work
#   So L must equal cur_shape[1] * cur_shape[2] * (c / last_dim)
#   Actually: B * L * c = B * cur_shape[1] * cur_shape[2] * (-1)
#   So: L = cur_shape[1] * cur_shape[2] when the last dimension is 1
#   Or more generally: L * c = cur_shape[1] * cur_shape[2] * something

# Let's use:
B = 2
H = 3
W = 4
c = 5

# For the view to work: B, L, c -> B, H, W, -1
# We need: L * c = H * W * d for some d
# Let's set L = H * W = 12, so d = c / 1 = 5... no that won't work with -1
# Actually if L=12, c=5: 12*5 = 60, and H*W = 12, so -1 = 5. That works.
# Let's simplify: make L = H*W

L = H * W  # L = 12

cur_x = torch.randn(B, L, c)  # shape (2, 12, 5)
num_remain = 3
cur_shape = [0, H, W]  # cur_shape[1]=H, cur_shape[2]=W

ARGS = {
    "cur_x": cur_x,
    "num_remain": num_remain,
    "cur_shape": cur_shape,
}
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 tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
masked_previous_scale_cache Ran
ran on a synthesized input
csguoh/fastvar/HART/hart/modules/networks/fastvar_utils.py
code served (permissive licence) · get_code("c24fea416022c9ec")
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, ndim 1, 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 torch
import math

torch.manual_seed(0)

# cur_x: shape (B, L, c) where L must be a perfect square
# The function does: cur_x.view(B, cur_x_hw, cur_x_hw, -1)
# So L = cur_x_hw * cur_x_hw
# Let's use B=2, cur_x_hw=4, so L=16, c=3
B = 2
cur_x_hw = 4
L = cur_x_hw * cur_x_hw
c = 3

cur_x = torch.randn(B, L, c)

# cur_scale: used in interpolate(size=(cur_scale, cur_scale))
# This should be a reasonable spatial dimension
cur_scale = 8

# num_remain: number of tokens to keep after filtering
# Should be less than L
num_remain = 5

ARGS = {
    "cur_x": cur_x,
    "cur_scale": cur_scale,
    "num_remain": num_remain,
}
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 tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
precompute_rope2d_freqs_grid Ran
ran; our contract draft was wrong, not the code
csguoh/fastvar/Infinity/infinity/models/basic.py
code served (permissive licence) · get_code("94b02b98b65ce1b4")
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 dict. 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 torch
import torch.nn as nn

torch.manual_seed(0)

# Build dynamic_resolution_h_w structure
# This mimics the nested dict structure expected by the function
dynamic_resolution_h_w = {
    "1:1": {
        "1M": {
            "scales": [
                (0, 16, 16),
                (1, 32, 32),
            ]
        },
        "2M": {
            "scales": [
                (0, 24, 24),
                (1, 48, 48),
            ]
        }
    },
    "2:1": {
        "1M": {
            "scales": [
                (0, 16, 32),
                (1, 32, 64),
            ]
        },
        "2M": {
            "scales": [
                (0, 24, 48),
                (1, 48, 96),
            ]
        }
    }
}

ARGS = {
    "dim": 4,
    "dynamic_resolution_h_w": dynamic_resolution_h_w,
    "rope2d_normalized_by_hw": 0,
    "pad_to_multiplier": 1,
    "max_height": 128,
    "max_width": 128,
    "base": 10000.0,
    "device": None,
    "scaling_factor": 1.0,
}
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 dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not 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
slow_attn Ran
ran; our contract draft was wrong, not the code
csguoh/fastvar/HART/hart/modules/networks/fastvar_basic.py
pointer only (licence: MIT); the stored copy get_code serves, from oliverrensu/mvar, is pointer only (licence: NONE) · get_code("f85408c04c9c626f")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from oliverrensu/mvar, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

apply_rotary_emb_fastvar Not yet run
unverified (harvested, no recorded run)
csguoh/fastvar/Infinity/infinity/models/fastvar_basic.py
code served (permissive licence) · get_code("9e16906b7f949f04")
How it ran

no invocation recorded.

features2image Not yet run
unverified (harvested, no recorded run)
csguoh/fastvar/Infinity/infinity/models/bitwise_self_correction.py
code served (permissive licence) · get_code("0a9d5c3882889998")
How it ran

no invocation recorded.

labels2image Not yet run
unverified (harvested, no recorded run)
csguoh/fastvar/Infinity/infinity/models/bitwise_self_correction.py
code served (permissive licence) · get_code("9ba6c0e0e8e4e50e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 distinct functions with an identity, 6 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 5 later) and 8 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
features2image 3 other papers
2 later, 1 same year, order not determinable
Chung_Fine-Tuning_Visual_Autogressive_Models_for_Subject-Driven_Generation_ICCV_2025_paper (2025 (year only))
Fine-Tuning Visual Autoregressive Models for Subject-Driven Generation
2508.15772 (2025-08)
Visual Autoregressive Modeling for Instruction-Guided Image Editing
2604.21450 (2026-04)
VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution
get_dropout_layer 3 other papers
2 later, 1 same year, order not determinable
Chung_Fine-Tuning_Visual_Autogressive_Models_for_Subject-Driven_Generation_ICCV_2025_paper (2025 (year only))
Fine-Tuning Visual Autoregressive Models for Subject-Driven Generation
2508.15772 (2025-08)
Visual Autoregressive Modeling for Instruction-Guided Image Editing
2604.21450 (2026-04)
VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution
labels2image 3 other papers
2 later, 1 same year, order not determinable
Chung_Fine-Tuning_Visual_Autogressive_Models_for_Subject-Driven_Generation_ICCV_2025_paper (2025 (year only))
Fine-Tuning Visual Autoregressive Models for Subject-Driven Generation
2508.15772 (2025-08)
Visual Autoregressive Modeling for Instruction-Guided Image Editing
2604.21450 (2026-04)
VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution
apply_rotary_emb 2 other papers
1 later, 1 same year, order not determinable
Chung_Fine-Tuning_Visual_Autogressive_Models_for_Subject-Driven_Generation_ICCV_2025_paper (2025 (year only))
Fine-Tuning Visual Autoregressive Models for Subject-Driven Generation
2604.21450 (2026-04)
VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution
precompute_rope2d_freqs_grid 2 other papers
1 later, 1 same year, order not determinable
Chung_Fine-Tuning_Visual_Autogressive_Models_for_Subject-Driven_Generation_ICCV_2025_paper (2025 (year only))
Fine-Tuning Visual Autoregressive Models for Subject-Driven Generation
2604.21450 (2026-04)
VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution
slow_attn 1 other paper
1 dated earlier
2411.10433 (2024-11)
M-VAR: Decoupled Scale-wise Autoregressive Modeling for High-Quality Image Generation

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

Figure 1. Our FastVAR can generate one 2K image using one NVIDIA 3090 GPU, while existing baseline fails due to out of memory.

For agents

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

get_harvested_code_for_paper("2503.23367")
get_code_for_paper("2503.23367")
have("2503.23367")

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

curl https://syntology.ai/api/ran/2503.23367.json

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

[![Syntology run record](https://syntology.ai/api/ran/2503.23367.svg)](https://syntology.ai/paper/2503.23367)

Connect an agent — have() is free.