SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2410.07171 · ICLR · 2025

IterComp: Iterative Composition-Aware Feedback Learning from Model Gallery for Text-to-Image Generation

Mengdi Wang, Guohao Li, Ling Yang, Yujiu Yang, Bin Cui, Xinchen Zhang, Yaqi Cai, Jiake Xie, Yong Tang

arXiv · PDF · Open in the Atlas

Code that ran

11 of 18 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; 3 honoured the contract we drafted, 2 violated it and 0 have no contract verdict; 6 failures of our instrument; 7 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/2410.07171.json · badge: /api/ran/2410.07171.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
yangling0818/rpg-diffusionmaster canonical (classifier label)
classifier basis: census
8 ran · 2 no recorded run
yangling0818/itercomp canonical (classifier label)
classifier basis: census
2 ran · 5 no recorded run
YangLing0818/IterComp canonical (classifier label)
classifier basis: classifier_2025
1 ran · 0 no recorded run
FunctionStatusWhere it lives
floatdef Ran
ran, honoured the contract we drafted
yangling0818/rpg-diffusionmaster/matrix.py
code served (permissive licence) · get_code("9b32f5c5fdc1fed6")
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 float, 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 numpy as np

torch.manual_seed(0)

ARGS = {
    "x": "123.45",
    "vdef": 0.0
}
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 float, 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
get_params_dict Ran
ran; our contract draft was wrong, not the code
yangling0818/rpg-diffusionmaster/mllm.py
code served (permissive licence) · get_code("8a3533b6e18fc64e")
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 re

def get_params_dict(output_text):
    response = output_text
    split_ratio_match = re.search('Final split ratio: ([\\d.,;]+)', response)
    if split_ratio_match:
        final_split_ratio = split_ratio_match.group(1)
        print('Final split ratio:', final_split_ratio)
    else:
        print('Final split ratio not found.')
    prompt_match = re.search('Regional Prompt: (.*?)(?=\\n\\n|\\Z)', response, re.DOTALL)
    if prompt_match:
        regional_prompt = prompt_match.group(1).strip()
        print('Regional Prompt:', regional_prompt)
    else:
        print('Regional Prompt not found.')
    image_region_dict = {'Final split ratio': final_split_ratio, 'Regional Prompt': regional_prompt}
    return image_region_dict

output_text_example = """Final split ratio: 0.7,0.3;0.5
Regional Prompt: A beautiful landscape with mountains and trees
Some other content here"""

ARGS = {
    'output_text': output_text_example
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — 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
is_l2 Ran
ran, violated the contract we drafted
yangling0818/rpg-diffusionmaster/matrix.py
code served (permissive licence) · get_code("98feb4cf3a68bbf2")
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 bool, 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)

ARGS = {
    'l': [[1, 2, 3], [4, 5, 6]]
}
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 bool, 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
is_url Ran
ran, violated the contract we drafted
yangling0818/itercomp/train/models/blip.py
pointer only (licence: MIT); the stored copy get_code serves, from jylins/videoxum, is pointer only (licence: NONE) · get_code("4d1d9c47619920ba")
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 jylins/videoxum, 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.

l2_count Ran
ran, honoured the contract we drafted
yangling0818/rpg-diffusionmaster/matrix.py
code served (permissive licence) · get_code("44c2fbeedf9741bb")
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 int, ndim 0, dtype family int, 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 numpy as np

torch.manual_seed(0)

# The function l2_count iterates over l (a list-like object)
# and for each row in l, it calls len(row)
# So l should be a list of objects that have a len()

# Create a list of lists with distinct sizes
l = [
    [1, 2, 3],           # len = 3
    [4, 5, 6, 7],        # len = 4
    [8, 9],              # len = 2
]

ARGS = {
    "l": l
}
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 int, ndim 0, dtype family int, 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
load Ran
ran; our contract draft was wrong, not the code
yangling0818/itercomp/Iterative_feedback/utils.py
code served (permissive licence) · get_code("96b66ffedde07222")
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 ImageReward. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 os
import torch
import tempfile
from unittest.mock import Mock, patch
from pathlib import Path

torch.manual_seed(0)

# Create a minimal ImageReward model class that satisfies the function's needs
class ImageReward(torch.nn.Module):
    def __init__(self, device='cpu', med_config=None):
        super().__init__()
        self.device_type = device
        self.med_config = med_config
        self.linear = torch.nn.Linear(4, 2)
    
    def forward(self, x):
        return self.linear(x)
    
    def eval(self):
        return super().eval()
    
    def to(self, device):
        return super().to(device)
    
    def load_state_dict(self, state_dict, strict=False):
        # Mock implementation - just accept the state dict
        return Mock()

# Create a temporary directory for model files
temp_dir = tempfile.mkdtemp()

# Create mock model checkpoint file
mock_checkpoint_path = os.path.join(temp_dir, 'ImageReward.pt')
torch.save({'linear.weight': torch.randn(2, 4), 'linear.bias': torch.randn(2)}, mock_checkpoint_path)

# Create mock med_config file
mock_config_path = os.path.join(temp_dir, 'med_config.json')
with open(mock_config_path, 'w') as f:
    f.write('{}')

# Mock the ImageReward class to be importable from this module
import sys
current_module = sys.modules[__name__]
setattr(current_module, 'ImageReward', ImageReward)

# Define ARGS dictionary with all parameters for the load function
ARGS = {
    'name': mock_checkpoint_path,  # Use local file path instead of model name
    'device': 'cpu',
    'download_root': temp_dir,
    'med_config': mock_config_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
huggingface_hub, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ImageReward — a ImageReward output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ImageReward 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
loss_func Ran
ran; our contract draft was wrong, not the code
YangLing0818/IterComp/train/train_reward_models.py
code served (permissive licence) · get_code("1164a205d991f3d8")
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) 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 torch
import torch.nn.functional as F

torch.manual_seed(0)

# The function expects a tensor called 'reward' with shape (batch_size, num_classes)
# where num_classes >= 2 (since it accesses reward[:, 0] and reward[:, 1])
# It uses cross_entropy which expects logits, so we create a tensor with appropriate shape

reward = torch.randn(2, 4)  # batch_size=2, num_classes=4

ARGS = {
    "reward": reward
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
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
rescale_noise_cfg Ran
ran; our contract draft was wrong, not the code
yangling0818/rpg-diffusionmaster/RegionalDiffusion_base.py
code served (permissive licence) · get_code("bea2d776a332f2b0")
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) and returned Tensor, ndim 4, dtype family float, shape (2, 4, 3, 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

torch.manual_seed(0)

# Create small tensors with distinct dimensions
# noise_cfg: shape (batch=2, channels=4, height=3, width=3)
noise_cfg = torch.randn(2, 4, 3, 3)

# noise_pred_text: shape (batch=2, channels=4, height=3, width=3)
noise_pred_text = torch.randn(2, 4, 3, 3)

# guidance_rescale: float between 0 and 1
guidance_rescale = 0.7

ARGS = {
    "noise_cfg": noise_cfg,
    "noise_pred_text": noise_pred_text,
    "guidance_rescale": guidance_rescale,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 4, dtype family float, shape (2, 4, 3, 3) — shape (2, 4, 3, 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
retrieve_timesteps Ran
ran; our contract draft was wrong, not the code
yangling0818/rpg-diffusionmaster/RegionalDiffusion_xl.py
pointer only (licence: MIT); the stored copy get_code serves, from susunghong/seg-sdxl, is pointer only (licence: NONE) · get_code("63e7d99da5ea137f")
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 susunghong/seg-sdxl, 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.

retrieve_timesteps Ran
ran; our contract draft was wrong, not the code
yangling0818/rpg-diffusionmaster/RegionalDiffusion_base.py
code served (permissive licence) · get_code("6cccd001a059f194")
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 torch
import inspect
from typing import List, Optional, Union

torch.manual_seed(0)

# Create a minimal scheduler class that satisfies the function's needs
class MockScheduler:
    def __init__(self, supports_timesteps=True):
        self.supports_timesteps = supports_timesteps
        self.timesteps = torch.tensor([999, 500, 100, 10])
    
    def set_timesteps(self, num_inference_steps=None, device=None, timesteps=None, **kwargs):
        if timesteps is not None:
            self.timesteps = torch.tensor(timesteps, device=device)
        else:
            # Simulate creating timesteps based on num_inference_steps
            self.timesteps = torch.linspace(999, 0, num_inference_steps, dtype=torch.long, device=device)

# Create scheduler instance
scheduler = MockScheduler(supports_timesteps=True)

# Define arguments for the retrieve_timesteps function
ARGS = {
    "scheduler": scheduler,
    "num_inference_steps": 4,
    "device": "cpu",
    "timesteps": 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 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
split_dims Ran
ran, honoured the contract we drafted
yangling0818/rpg-diffusionmaster/cross_attention.py
pointer only (licence: MIT); the stored copy get_code serves, from krafton-ai/rare-to-frequent, is pointer only (licence: NONE) · get_code("56156523c6123e73")
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 krafton-ai/rare-to-frequent, 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.

GPT4 Not yet run
unverified (harvested, no recorded run)
yangling0818/rpg-diffusionmaster/mllm.py
code served (permissive licence) · get_code("24399db18bc9c34f")
How it ran

no invocation recorded.

ImageReward_download Not yet run
unverified (harvested, no recorded run)
yangling0818/itercomp/Iterative_feedback/utils.py
code served (permissive licence) · get_code("b1ae6f0bdf766e6c")
How it ran

no invocation recorded.

encode_prompt_rm Not yet run
unverified (harvested, no recorded run)
yangling0818/itercomp/Iterative_feedback/feedback_train.py
code served (permissive licence) · get_code("2d955b94f1d7c1b6")
How it ran

no invocation recorded.

import_model_class_from_model_name_or_path Not yet run
unverified (harvested, no recorded run)
yangling0818/itercomp/Iterative_feedback/feedback_train.py
code served (permissive licence) · get_code("01e3b39e9e10abfb")
How it ran

no invocation recorded.

load_score Not yet run
unverified (harvested, no recorded run)
yangling0818/itercomp/Iterative_feedback/utils.py
code served (permissive licence) · get_code("ef6e5967acc37116")
How it ran

no invocation recorded.

local_llm Not yet run
unverified (harvested, no recorded run)
yangling0818/rpg-diffusionmaster/mllm.py
code served (permissive licence) · get_code("0861ce73c0e98185")
How it ran

no invocation recorded.

parse_args Not yet run
unverified (harvested, no recorded run)
yangling0818/itercomp/Iterative_feedback/feedback_train.py
code served (permissive licence) · get_code("46770865c030e572")
How it ran

no invocation recorded.

Also present in other papers' code

Of 18 samples lifted from this paper's linked repositories, 18 carry a function identity and 0 do not. Of the 17 distinct functions with an identity, 14 are also present in other papers whose code we hold (14 with at least one paper dated earlier, 3 the same month, 5 later) and 3 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
rescale_noise_cfg 62 other papers
38 dated earlier, 7 the same month, 16 later, 1 same year, order not determinable
2206.02779 (2022-06)
Blended Latent Diffusion
2304.06720 (2023-04)
Expressive Text-to-Image Generation with Rich Text
2309.06380 (2023-09)
InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation
and 59 more not named here
retrieve_timesteps 21 other papers
14 dated earlier, 3 the same month, 4 later
2310.06313 (2023-10)
Advancing Pose-Guided Image Synthesis with Progressive Conditional Diffusion Models
2401.11708 (2024-01)
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs
2403.06976 (2024-03)
BrushNet: A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion
and 18 more not named here
is_url 11 other papers
9 dated earlier, 1 later, 1 same year, order not determinable
2211.09790 (2022-11)
ConStruct-VL: Data-Free Continual Structured VL Concepts Learning
2303.12060 (2023-03)
VideoXum: Cross-modal Visual and Textural Summarization of Videos
2304.05977 (2023-04)
ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation
and 8 more not named here
floatdef 2 other papers
1 dated earlier, 1 later
2401.11708 (2024-01)
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs
2504.09697 (2025-04)
SPICE: A Synergistic, Precise, Iterative, and Customizable Image Editing Workflow
import_model_class_from_model_name_or_path 2 other papers
1 dated earlier, 1 later
2304.05977 (2023-04)
ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation
2510.23974 (2025-10)
Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models
split_dims 2 other papers
1 dated earlier, 1 the same month
2401.11708 (2024-01)
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs
2410.22376 (2024-10)
Rare-to-Frequent: Unlocking Compositional Generation Power of Diffusion Models on Rare Concepts with LLM Guidance
GPT4 1 other paper
1 dated earlier
2401.11708 (2024-01)
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs
ImageReward_download 1 other paper
1 dated earlier
2304.05977 (2023-04)
ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation
get_params_dict 1 other paper
1 dated earlier
2401.11708 (2024-01)
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs
is_l2 1 other paper
1 dated earlier
2401.11708 (2024-01)
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs
l2_count 1 other paper
1 dated earlier
2401.11708 (2024-01)
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs
load 1 other paper
1 dated earlier
2304.05977 (2023-04)
ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation
load_score 1 other paper
1 dated earlier
2304.05977 (2023-04)
ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation
local_llm 1 other paper
1 dated earlier
2401.11708 (2024-01)
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs

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

Advanced diffusion models like Stable Diffusion 3, Omost, and FLUX have made notable strides in compositional text-to-image generation. However, these methods typically exhibit distinct strengths for compositional generation, with some excelling in handling attribute binding and others in spatial relationships. This disparity highlights the need for an approach that can leverage the complementary strengths of various models to comprehensively improve the composition capability. To this end, we introduce IterComp, a novel framework that aggregates composition-aware model preferences from multiple models and employs an iterative feedback learning approach to enhance compositional generation. Specifically, we curate a gallery of six powerful open-source diffusion models and evaluate their three key compositional metrics: attribute binding, spatial relationships, and non-spatial relationships. Based on these metrics, we develop a compositionaware model preference dataset comprising numerous image-rank pairs to train composition-aware reward models. Then, we propose an iterative feedback learning method to enhance compositionality in a closed-loop manner, enabling the progressive self-refinement of both the base diffusion model and reward models over multiple iterations. Detailed theoretical proof demonstrates the effectiveness of this method. Extensive experiments demonstrate our significant superiority over previous methods, particularly in multi-category object composition and complex semantic alignment. IterComp opens new research avenues in reward feedback learning for diffusion models and compositional generation.

For agents

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

get_harvested_code_for_paper("2410.07171")
get_code_for_paper("2410.07171")
have("2410.07171")

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

curl https://syntology.ai/api/ran/2410.07171.json

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

[![Syntology run record](https://syntology.ai/api/ran/2410.07171.svg)](https://syntology.ai/paper/2410.07171)

Connect an agent — have() is free.