BBQ-to-Image: Numeric Bounding Box and Qolor Control in Large-Scale Text-to-Image Models

Eliran Kachlon; Alexander Visheratin; Nimrod Sarid; Tal Hacham; Eyal Gutflaish; Saar Huberman; Hezi Zisman; David Ruppin

cs.CVFeb 24, 2026

BBQ-to-Image: Numeric Bounding Box and Qolor Control in Large-Scale Text-to-Image Models

Authors: Eliran Kachlon, Alexander Visheratin, Nimrod Sarid, Tal Hacham, Eyal Gutflaish, Saar Huberman, Hezi Zisman, David Ruppin,

Organizations: BRIA AI

Abstract

Text-to-image models have rapidly advanced in realism and controllability, with recent approaches leveraging long, detailed captions to support fine-grained generation. However, a fundamental parametric gap remains: existing models rely on descriptive language, whereas professional workflows require precise numeric control over object location, size, and color. In this work, we introduce BBQ, a large-scale text-to-image model that directly conditions on numeric bounding boxes and RGB triplets within a unified structured-text framework. We obtain precise spatial and chromatic control by training on captions enriched with parametric annotations, without architectural modifications or inference-time optimization. This also enables intuitive user interfaces such as object dragging and color pickers, replacing ambiguous iterative prompting with precise, familiar controls. Across comprehensive evaluations, BBQ achieves strong box alignment and improves RGB color fidelity over state-of-the-art baselines. More broadly, our results support a new paradigm in which user intent is translated into an intermediate structured language, consumed by a flow-based transformer acting as a renderer and naturally accommodating numeric parameters.

Figures & tables

Figure 2: End-to-end parametric workflow. A short prompt is expanded by a VLM into a structured JSON that includes numeric bounding boxes and RGB values (for clarity, we show only the parametric fields for the woman). The JSON is then provided to BBQ to generate an image. Users can edit specific fields (e.g., box coordinates or color values), and BBQ updates the output accordingly while preserving unrelated content, demonstrating native disentanglement. Notably, BBQ receives no image input, and consistency is maintained solely through the disentangle structured conditioning.

Panel of Figure 3 — Figure 3: Disentangled parametric refinement via structured re-generation. Each example starts from an image generated from a structured JSON prompt. We then edit only the relevant JSON fields and re-generate using the same random seed. Although the model does not observe the original image, it produces localized changes that follow the modified parameters while preserving the rest of the scene, demonstrating strong parametric disentanglement. Ground-truth bounding boxes are overlaid for visualization.

Panel of Figure 4 — Figure 4: Text-as-a-Bottleneck Reconstruction (TaBR). Starting from the original image (left), a detailed caption is generated and used as input to each model. The resulting reconstructions are compared against the original. BBQ more faithfully preserves scene layout, object relations, and fine-grained attributes than competing state-of-the-art models, demonstrating improved expressiveness.

Panel of Figure 5 — Figure 5: Color-conditioning accuracy. Each example shows the target color (left) and images generated by different models when conditioned on the same object and exact RGB value. BBQ achieves high chromatic fidelity to the target color and produces competitive results compared to state-of-the-art text-to-image models under identical color-conditioning prompts.

Panel of Figure 6 — Figure 6: Bounding-box accuracy. We compare BBQ with Nano Banana Pro and Flux.2 Pro on prompts that include explicit numeric bounding-box specifications (overlaid on the images). While the baseline models often struggle to consistently follow these spatial constraints, BBQ reliably places objects within the specified boxes.

Model (vs. BBQ)	BBQ win rate $\uparrow$	95% CI
Nano Banana Pro	65.2%	[50.8, 77.3]
FIBO	76.1%	[62.1, 86.1]
FLUX.2 Pro	93.3%	[82.1, 97.7]

Table 1: Text-as-a-Bottleneck Reconstruction (TaBR). Win rate is computed as the fraction of images where BBQ is preferred over the competing model among decisive comparisons (ties ignored). Confidence intervals correspond to 95% Wilson score intervals. BBQ outperforms all evaluated baselines across all comparisons.

	COCO			LVIS
Method	$\mathbf{AP}$	$\mathbf{AP_{50}}$	$\mathbf{AR}$	AP	$\mathbf{AP_{50}}$	$\mathbf{AP_{s}}$	$\mathbf{AP_{m}}$	$\mathbf{AP_{l}}$	$\mathbf{AP_{r}}$	$\mathbf{AP_{c}}$	$\mathbf{AP_{f}}$
Upper bound (real images)	50.2	66.7	61.0	44.6	57.7	33.2	55.0	66.1	31.4	44.5	50.5
Flux.2 Pro	3.5	8.7	4.6	2.1	4.8	0.1	0.8	7.3	0.1	0.2	2.4
Nano Banana Pro	5	11.3	5.5	4.1	10.6	0.1	1	15.9	3.3	4	3.9
GLIGEN [ 25 ]	19.6	35.0	30.7	9.9	9.5	1.6	10.5	31.1	7.4	10.0	10.9
BBQ (ours)	28.6	40.9	38.2	13.1	20.1	1.6	13.9	37.4	13.1	14.0	12.7

Table 2: Bounding-box alignment under box-conditioned generation on COCO and LVIS. We follow the InstanceDiffusion evaluation protocol using YOLO-based detection; the upper bound corresponds to detector performance on real images. Across both datasets, BBQ consistently outperforms strong text-to-image baselines (Nano Banana Pro and Flux.2 Pro) and GLIGEN, while trailing the specialized InstanceDiffusion approach. Importantly, BBQ achieves this without architectural modifications or grounding-specific components and is trained for high-fidelity image synthesis, providing strong spatial control within a general large-scale and disentangle model, also allowing intuitive refinement. Best results are in bold ; second best are underlined .

$K$	Model	$a$ – $b$ Mean $\downarrow$	$a$ – $b$ Median $\downarrow$	$a$ – $b$ p90 $\downarrow$	$\Delta E_{00}$ Mean $\downarrow$	$\Delta E_{00}$ Median $\downarrow$	$\Delta E_{00}$ p90 $\downarrow$
5	BBQ (ours)	7.16	6.67	13.30	5.93	5.76	9.62
	Nano Banana Pro	10.91	7.52	20.80	6.50	5.72	11.00
	FLUX.2 Pro	10.07	8.16	19.30	6.64	6.12	10.17
	FIBO	10.32	9.44	20.36	6.74	7.32	10.52
8	BBQ (ours)	7.48	6.23	14.33	5.74	5.07	9.89
	Nano Banana Pro	10.64	6.87	21.05	5.91	4.82	10.80

Table 3: Color fidelity comparison.

\Delta E_{00}

(CIEDE2000) measures perceptual color difference, while

a

–

b

distance captures chromaticity (hue and saturation) independently of lightness. We report mean, median, and 90th percentile (p90), where lower values indicate better color accuracy.. Across both

K=5

and

K=8

, BBQ achieves the lowest

a

–

b

errors in all statistics, indicating the most accurate chromaticity control and fewer severe failures, while remaining competitive under

\Delta E_{00}

that penalizes lightness differences. Best results are in bold and second-best are underlined (computed per

K

Geometry-Conditioned Image Generation

See all 11 papers

NumColor: Precise Numeric Color Control in Text-to-Image Generation
Mar 13, 2026Muhammad Atif Butt, Diego Hernandez, Alexandra Gomez-Villa +3Controllable Image Generation T2I Diffusion Models
Computer Vision Center, Spain · Computer Sciences Department, Universitat Autònoma de Barcelona, Spain · Program of Computer Science, City University of Hong Kong (Dongguan) +1
Camera Control for Text-to-Image Generation via Learning Viewpoint Tokens
Apr 21, 2026Xinxuan Lu, Charless Fowlkes, Alexander C. BergT2I Generation Controllable Image Generation
University of California, Irvine
NumBench: Diagnosing Counting Failures in Text-to-Image Models
Aug 28, 2026Sandeep Wadhwa, Mayank Vatsa, Richa Singh +2T2I Generation T2I Generation Evaluation
Indian Institute of Technology Jodhpur, India · Shiv Nadar University, India

Mar 13, 2026J/K move · Enter open · S save

NumColor: Precise Numeric Color Control in Text-to-Image Generation

Muhammad Atif Butt, Diego Hernandez, Alexandra Gomez-Villa, Kai Wang+2

Computer Vision Center, Spain · Computer Sciences Department, Universitat Autònoma de Barcelona, Spain · Program of Computer Science, City University of Hong Kong (Dongguan)+1

Controllable Image Generation T2I Diffusion Models

Open paper AskPDF arXiv

Text-to-image diffusion models excel at generating images from natural language descriptions, yet fail to interpret numerical colors such as hex codes (#FF5733) and RGB values (rgb(255,87,51)). This limitation stems from subword tokenization, which fragments color codes into semantically meaningless tokens that text encoders cannot map to coherent color representations. We present NumColor, that enables precise numerical color control across multiple diffusion architectures. NumColor comprises two components: a Color Token Aggregator that detects color specifications regardless of tokenization, and a ColorBook containing 6,707 learnable embeddings that map colors to embedding space of text encoder in perceptually uniform CIE Lab space. We introduce two auxiliary losses, directional alignment and interpolation consistency, to enforce geometric correspondence between Lab and embedding spaces, enabling smooth color interpolation. To train the ColorBook, we construct NumColor-Data, a synthetic dataset of 500K rendered images with unambiguous color-to-pixel correspondence, eliminating the annotation ambiguity inherent in photographic datasets. Although trained solely on FLUX, NumColor transfers zero-shot to SD3, SD3.5, PixArt-α, and PixArt-Σ without model-specific adaptation. NumColor improves numerical color accuracy by 4-9x across five models, while simultaneously improving color harmony scores by 10-30x on GenColorBench benchmark.

Abstract

Geometry-Conditioned Image Generation

Explore similar work

NumColor: Precise Numeric Color Control in Text-to-Image Generation

Camera Control for Text-to-Image Generation via Learning Viewpoint Tokens

NumBench: Diagnosing Counting Failures in Text-to-Image Models