Text-to-image diffusion models can render a photorealistic analog clock but cannot control the time it displays. Across four pretrained models, fewer than 1% of generated clocks place both hands within 6 degrees of the requested time; the models instead collapse onto visually dominant configurations such as the canonical 10:10 pose. The same failure appears when a specific fraction of a circular object is requested, where models return complete or conventionally sliced objects. We observe that these two failures share a geometric structure: both target quantities can be parameterized by angular variables. We introduce a geometric token encoding framework that conditions generation directly on these angles. Each angular component is encoded as a periodic sine-cosine feature and projected into a role-specific token embedding that is injected into the prompt before a frozen text encoder. To spatially ground these tokens without requiring localization input at inference, we introduce Stochastic Localization Relaxation (SLR), which stochastically relaxes attention supervision from component-level to object-level to unconstrained localization during training. For clocks, we additionally introduce a hand-swap preference objective that discourages assigning the correct angles to the wrong hand roles. On analog clocks our method reaches 63-78% Acc@6 degrees across evaluation methods, compared with below 1% for the baselines. On circular-fraction generation, it achieves 89.5% [email protected], while maintaining 88.8% on unseen object categories and 84.9% on unseen target values, compared with roughly 20% for pretrained baselines.
Figures & tables
Figure 2: Left: Overview of our geometric token encoding framework, illustrated for clock-time control. The target clock time (h,m) or fraction-sector boundaries are encoded as sinusoidal pairs and mapped through the projection network P to produce role-specific geometric tokens eh,em (or es,ee ) injected into the prompt before the frozen text encoder. Both tasks share the same MLP architecture and condition a LoRA-adapted SD 2.1 denoising U-Net. Right: Stochastic Localization Relaxation (SLR). Training samples one of three localization regimes (55/30/15%): component-level CALL using component boxes, object-level CALL using a shared object box, and unconstrained standard cross-attention. No localization input is required at inference.
Figure 3: Dataset construction for the two domains. Top: Procedurally rendered clocks are restyled using ControlNet conditioned on Canny edges while preserving exact hand angles. Bottom: Complete circular objects generated by Flux.1 are segmented and arc-masked according to the target fraction.
GPT-5.6
Claude Opus 5
IAT
Human
User Study
Model
A@ 6∘↑
Avg. MAE ↓
A@ 6∘↑
Avg. MAE ↓
A@ 6∘↑
Avg. MAE ↓
A@ 6∘↑
Avg. MAE ↓
Ctrl. ↑
App. ↑
SD 2.1 [ 29 ]
0.4%
88.1∘
0.9%
91.3∘
0.5%
89.7∘
0.0%
104.5∘
1.5%
5.7%
SDXL [ 27 ]
0.2%
90.7∘
0.7%
91.2∘
0.4%
90.7∘
0.0%
112.8∘
1.3%
7.4%
SD 3.5 [ 14 ]
0.0%
88.4∘
0.0%
87.8∘
0.1%
90.7∘
0.0%
89.8∘
1.9%
9.7%
Flux.1 [ 6 ]
0.0%
91.9∘
0.0%
87.9∘
0.6%
90.4∘
0.0%
92.6∘
5.7%
26.9%
Ours (train)
71.9%
21.0∘
63.4%
23.6∘
76.6%
19.9∘
75.3%
20.7∘
89.6%
50.3%
Table 1: Clock generation results under three automatic readers, human annotation, and a user study with 33 participants. Acc@ 6∘ : both hand errors ≤6∘ . Avg. MAE: mean angular error averaged over both hands. The user study reports the percentage of selections for geometric controllability (Ctrl.) and appearance quality (App.). Additional times: 48 non-quarter-hour targets (960 images).
Table 2: Fraction generation results. Acc@ k : fraction error ≤k . Baselines are evaluated without fine-tuning. Unseen objects: 14 novel categories. Unseen fractions: held-out fraction values. Dense 360∘ : all 1∘ steps over 3 objects. User study reports controllability/appearance preference (%).
Figure 4: Clock generation. Under the same random seeds, baseline models often produce common clock configurations instead of the requested time, whereas our model accurately follows the target time.
Figure 5: Fraction generation. Base models frequently produce complete objects or conventional slices, whereas our model accurately follows the requested fraction for both pizza and pie across five target values.
Acc@ 6∘↑
Conditioning (SD 2.1)
GPT-5.6
Claude
IAT
Text prompt
0.4%
0.9%
0.5%
+ prompt engineering
0.3%
0.7%
0.8%
+ text-only LoRA
0.7%
0.2%
0.7%
+ geometric tokens (ours)
71.8%
63.1%
77.7%
Table 3: Controlled comparison on a fixed SD 2.1 backbone over the same 960 clock generations. Neither prompt engineering nor text-conditioned fine-tuning provides meaningful geometric control; only geometric tokens achieve high accuracy.
Figure 6: Generalization to unseen fraction objects. Our model follows the requested arc fraction across four object categories not seen during training (rows) and four target fractions (columns), without additional fine-tuning.
GPT-5.6
Claude Opus 5
IAT
Variant
A@ 6∘↑
MAE ↓
A@ 6∘↑
MAE ↓
A@ 6∘↑
MAE ↓
Ours
71.8%
21.8 ∘
63.1%
24.2 ∘
77.7%
18.8 ∘
w/o swap pref.
66.8%
22.7 ∘
57.0%
26.5 ∘
65.6%
20.4 ∘
w/o SLR
20.3%
69.5 ∘
16.8%
73.1 ∘
23.5%
71.1 ∘
w/o attn. local.
9.4%
53.5 ∘
8.9%
51.5 ∘
10.2%
50.1 ∘
post-enc. injection
0.0%
97.3 ∘
0.0%
90.2 ∘
0.5%
88.3 ∘
Table 4: Clock ablation (Acc@ 6∘ : both hands within 6∘ ; averaged over train and unseen-style prompts). Avg. MAE denotes the mean angular error across both hands. Unless otherwise specified, ablations use the sinusoidal encoder and SLR. IAT: ItsAboutTime ResNet50 reader.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Distribution of generated analog-clock times for pretrained text-to-image models. Several models exhibit strong modes near canonical configurations, particularly 10:10, despite variation in the requested target time.
Model
N
Top time
Frequency
SD 3.5
1000
10:10
33.7%
Flux.1
1000
10:10
26.2%
Flux.2
1000
10:10
48.0%
Gemini Nano Banana
100
10:10
47.0%
Appendix
Table 5: Canonical clock-time bias in pretrained text-to-image models. We report the most frequent readable clock configuration and its frequency among readable generations.
#
Template
Training styles
1
PLACEHOLDER, a realistic analog wall clock
2
PLACEHOLDER, a decorative analog clock
3
PLACEHOLDER, a black analog clock
4
PLACEHOLDER, a vintage analog clock
5
PLACEHOLDER, a minimalist white clock
Appendix
Table 6: The 20 clock style templates. Styles 1–12 are training styles; 13–20 are held out and used for the unseen-style evaluation.
Model
Steps
Guidance
Resolution
SD 2.1
50
7.5
5122
SDXL
50
7.5
5122
SD 3.5
28
4.5
5122
Flux.1
28
3.5
5122
Ours
50
7.5
5122
Appendix
Table 7: Sampler configuration. Each model uses its corresponding step count and guidance scale; all main-table images are generated at 512×512 .
Target
Row
Prompt
7:15
Baselines
an analog clock showing 7:15, a realistic analog wall clock in a bright kitchen with coffee mugs and fruit on the counter, high detail, natural lighting
Ours
a ⟨eh⟩ hour hand and a ⟨em⟩ minute hand, a realistic analog wall clock in a bright kitchen with coffee mugs and fruit on the counter, high detail, natural lighting
3:00
Baselines
an analog clock showing 3:00, a large analog clock on a cafe wall above small tables and chairs, high detail, natural lighting
Ours
a ⟨eh⟩ hour hand and a ⟨em⟩ minute hand, a large analog clock on a cafe wall above small tables and chairs, high detail, natural lighting
7:00
Baselines
a round analog clock in a greenhouse, showing 7:00, with plants and clay pots nearby, high detail, natural lighting
Ours
a round analog clock in a greenhouse, showing a ⟨eh⟩ hour hand and a ⟨em⟩ minute hand, with plants and clay pots nearby, high detail, natural lighting
Appendix
Table 8: Clock prompts. ⟨eh⟩ and ⟨em⟩ carry the target hand angles; no textual time appears in our prompts.
Target
Row
Prompt
2/5
Baselines
a top-down realistic food scene with a giant round chocolate chip cookie showing only 40% of its area as a filled arc sector starting at 12 o’clock going clockwise, next to a glass of milk and chocolate pieces, on a kitchen counter
Ours
a top-down realistic food scene with a giant round chocolate chip cookie with a ⟨es⟩ start boundary and a ⟨ee⟩ end boundary, next to a glass of milk and chocolate pieces, on a kitchen counter
4/5
Baselines
a top-down realistic food scene with a round pumpkin pie showing only 80% of its area as a filled arc sector starting at 12 o’clock going clockwise, next to cinnamon sticks and a tea cup, on a warm wooden table
Ours
a top-down realistic food scene with a round pumpkin pie with a ⟨es⟩ start boundary and a ⟨ee⟩ end boundary, next to cinnamon sticks and a tea cup, on a warm wooden table
3/4
Baselines
a top-down realistic food scene with a round red velvet cake showing only 75% of its area as a filled arc sector starting at 12 o’clock going clockwise, next to a cup of tea and berries, on a terrazzo table
Ours
a top-down realistic food scene with a round red velvet cake with a ⟨es⟩ start boundary and a ⟨ee⟩ end boundary, next to a cup of tea and berries, on a terrazzo table
Appendix
Table 9: Fraction prompts. ⟨es⟩ and ⟨ee⟩ encode the start and end boundaries of the sector; the fraction never appears as text.
Mixture
Validation
Test
55 / 30 / 15
78.6%
77.7%
70 / 20 / 10
76.2%
72.5%
40 / 40 / 20
73.9%
63.4%
33 / 33 / 33
72.9%
64.8%
20 / 40 / 40
70.4%
62.4%
Appendix
Table 10: SLR mixture comparison using IAT Acc@ 6∘ . Mixtures are ordered as component/object/unconstrained. All results use training seed 1908.
Mixture
Validation
Test
55 / 30 / 15
74.4±11.0
72.4±10.7
70 / 20 / 10
65.1±14.1
63.8±9.7
Appendix
Table 11: Training-seed variability across three runs (seeds 1908, 2027, and 3141), evaluated with IAT. Values are mean ± standard deviation of Acc@ 6∘ across training seeds.
Table 12: Fraction ablation on n=994 samples (14 training object categories, 71 angles). Acc@ k : fraction error ≤k .
Figure 8: Qualitative comparison of the original periodic encoding and the augmented encoding with the scalar fraction f near the angular boundaries. The periodic encoding produces small sectors near 0∘ but may generate narrow slices even for complete-sector targets ( 360∘ ), reflecting its endpoint ambiguity. Adding f distinguishes the endpoints and enables complete-object generation, but can produce oversized sectors for small target angles ( 1∘ – 5∘ ). Results are shown for pizza, pie, and cake.
Figure 9: In-scene generation with no bounding-box supervision at inference. (a) Clock time control across diverse room scenes. (b) Food-fraction control in top-down food settings. In both domains the model accurately tracks the requested geometric quantity (time / fraction) while maintaining scene coherence and object placement.
Current text-to-image models struggle to provide precise camera control using natural language alone. In this work, we present a framework for precise camera control with global scene understanding in text-to-image generation by learning parametric camera tokens. We fine-tune image generation models for viewpoint-conditioned text-to-image generation on a curated dataset that combines 3D-rendered images for geometric supervision and photorealistic augmentations for appearance and background diversity. Qualitative and quantitative experiments demonstrate that our method achieves state-of-the-art accuracy while preserving image quality and prompt fidelity. Unlike prior methods that overfit to object-specific appearance correlations, our viewpoint tokens learn factorized geometric representations that transfer to unseen object categories. Our work shows that text-vision latent spaces can be endowed with explicit 3D camera structure, offering a pathway toward geometrically-aware prompts for text-to-image generation. Project page: https://randdl.github.io/viewtoken_control/
Text-to-Image (T2I) diffusion/flow models have recently achieved remarkable progress in visual fidelity and text alignment. However, they remain limited when users need to precisely control image layouts, something that natural language alone cannot reliably express. Controllable generation methods augment the initial T2I model with additional conditions that more easily describe the scene. Prior works straightforwardly train the augmented network with the same loss as the initial network. Although natural at first glance, this can lead to very long training times in some cases before convergence. In this work, we revisit the training objective of controllable diffusion models through a detailed analysis of their denoising dynamics. We show that direct supervision on the clean target image, dubbed x0-supervision, or an equivalent re-weighting of the diffusion loss, yields faster convergence. Experiments on multiple control settings demonstrate that our formulation accelerates convergence by up to 2× according to our novel metric (mean Area Under the Convergence Curve - mAUCC), while also improving both visual quality and conditioning accuracy. Our code is available at https://github.com/CEA-LIST/x0-supervision
Amadou S. Sangare, Adrien Maglo, Mohamed Chaouch +1
Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France
Text-to-image diffusion models have two major drawbacks that severely limit their practical utility: (1) standard models lack an intrinsic mechanism for continuous, concept-specific guidance (e.g., for precisely controlling how aesthetically pleasing an image looks), and (2) they lack reliability for tasks requiring high local coherence (e.g., generating text or human hands). To tackle these issues, we introduce a novel notion of concept-wise mutual information and find large, concept-dependent differences between individual layers, demonstrating that the generation of specific structures is localized in distinct parts of the network. We exploit this insight by reinforcing the impact of concept-relevant layers in Concept Guidance (CoG), a precise, target-specific guidance method that works for models out-of-the-box without additional training, external models, gradients, or prompt engineering. CoG first quantifies each layer's concept-specific impact and then guides denoising using a weighted combination of predictions generated with concept-relevant layers skipped. We demonstrate performance increases across various targets and popular models like PixArt-alpha, SD3, SD3.5, and FLUX.1-dev. Code is available at https://github.com/CompVis/concept_guidance
Nikolai Röhrich, Isabell Hans, Felix Krause +1
LMU Munich, Germany · Konrad Zuse School of Excellence in Reliable AI (relAI), Germany · Munich Center for Machine Learning (MCML), Germany