CountLoop: Training-Free High-Instance Image Generation

Abstract

CountLoop is a training-free framework capable of generating a high number of instances with precise layout control and high aesthetic quality. Our method leverages iterative agent guidance, layout conditioning, and cross-instance texture consistency to achieve state-of-the-art results in high-instance image generation tasks.

Pipeline Overview

Given a text prompt, the Layout Designer constructs a planning graph encoding object attributes and spatial relations, which is converted into a pixel-aligned layout. Guided by instance masks and cumulative latent composition with an IP-Adapter, the image is synthesized. A Design-Critic evaluates the result and updates the planning graph via an iterative feedback loop. This loop repeats until the count and quality goals are met.

Benchmarks & Evaluation

GDINO as counting metric for robust, open-vocabulary evaluation
New benchmarks: CountLoop-S (single-category, high-count) and CountLoop-M (multi-category)
Superior performance vs. SLD, Make-It-Count, and others

Counting and Aesthetic Quality Across Four Benchmarks

**Comparing counting and aesthetic quality across four benchmarks.** For every dataset we report *Counting*—split into F1 score and **Acc**uracy—and *Spatial* (aesthetic quality).
Type	Model	COCO-Count			T2I-Compbench			CountLoop-S			CountLoop-M
Type	Model	F1	Acc.	Spatial	F1	Acc.	Spatial	F1	Acc.	Spatial	F1	Acc.	Spatial
T2I	SDXL	71.87	42.13	0.38	84.36	44.00	0.75	55.40	24.49	0.63	77.84	67.25	0.55
	FLUX	84.73	54.19	0.53	90.75	57.00	0.78	49.08	29.59	0.65	79.99	78.00	0.58
	SD 3.5	83.97	50.56	0.46	88.56	50.00	0.76	54.96	33.67	0.64	79.91	77.19	0.56
	Counting Guidance	67.54	18.50	0.63	71.41	17.50	0.56	36.67	10.20	0.47	64.42	25.90	0.41
	GPT-4o	92.91	72.50	0.55	94.19	68.00	0.80	49.45	39.64	0.69	79.10	50.11	0.60
Agentic	GenArtist	75.40	45.50	0.45	85.33	55.82	0.70	51.00	30.56	0.60	77.87	70.34	0.57
	SLD	90.34	69.90	0.70	91.50	65.50	0.77	55.04	40.07	0.75	82.46	74.35	0.65
	RPG-DiffusionMaster	84.89	60.73	0.60	91.32	60.00	0.75	51.89	34.38	0.70	80.16	71.46	0.62
L2I	LMD	54.69	29.81	0.24	71.44	35.50	0.73	49.24	28.57	0.66	80.28	77.67	0.64
	MIGC	73.82	36.11	0.36	71.47	33.00	0.65	54.16	25.17	0.65	81.06	79.08	0.62
	CountGen	58.99	50.00	0.61	63.75	19.78	0.75	48.18	41.40	0.72	72.00	45.33	0.69
	CountLoop (Ours)	98.47	93.33	0.93	95.38	78.50	0.79	60.00	55.00	0.97	85.43	83.67	0.73

BibTeX Citation

    @article{Mondal et.al., 
    title={CountLoop: Iterative Agent Guided High Instance Image Generation}, 
    url={https://openreview.net/forum?id=NZ0H1XtcZG}, 
    author={Mondal, Anindya and Banerjee, Ayan and Nag, Sauradip and Llados, Josep and Zhu, Xiatian and Dutta, Anjan},
    language={en}}

CountLoop

Iterative Agent Guided High Instance Image Generation

Abstract

Key Features

Pipeline Overview

Visual Results

Benchmarks & Evaluation

Counting and Aesthetic Quality Across Four Benchmarks

BibTeX Citation