Recent advancements in diffusion models have significantly improved text-to-image generation, enabling their application to visual storytelling, where coherent image sequences are generated from narrative prompts. However, existing methods primarily focus on identity preservation and often fail to adapt to diverse scene variations, resulting in semantic misalignment and structural inflexibility. We propose EOS-Story, a training-free framework that harmonizes identity consistency with scene-specific adaptation across text and vision spaces. In the text space, we decouple the narrative into identity and scene sub-prompts to prevent information dilution, but this leaves the scene without access to appearance cues. To bridge this gap, we introduce Cross-Contextual Bridging, which softly injects a phrase-level appearance residual into the scene's <EOS> embedding. In the vision space, we find that scale mismatch between identity and scene subjects causes a few identity tokens to be repeatedly transferred, producing structural artifacts. We address this with Adaptive Identity Preservation, which anchors identity features at the scene-specific scale (Identity Anchoring) and integrates them via Identity-Aware Self-Attention. To evaluate performance under varied spatial compositions, we extend the Consistory+ benchmark with diverse framing scenarios (Consistory+_+). Extensive experiments show that EOS-Story achieves competitive or superior performance on both benchmarks.
Figures & tables
Figure 1: Limitations in scene-level adaptation of previous methods. Existing approaches often fail to account for scene-specific variance : (a) semantic misalignment in simple cases, (b) loss of fine-grained prompt details in complex scenarios, and (c) inability to adapt to diverse viewpoints and scales . These issues often lead to repetitive subject structures and weakened alignment with scene-specific prompts. In contrast, our method achieves better scene-level fidelity by adaptively harmonizing identity consistency with the unique context of each scene.
Figure 2: Effect of EOS modification. (a) Injecting “black” into the ⟨ EOS ⟩ token most effectively reflects the intended attribute. (b) Editing ⟨ EOS ⟩ embedding induces coherent semantic changes in images.
Figure 3: EOS-Story pipeline. Our framework contains two complementary modules: (1) Cross-Contextual Bridging for representation-level semantic recomposition ( Sec. 3.2 ), and (2) Adaptive Identity Preservation (AIP) for spatially-aware structural alignment ( Sec. 3.3 ). AIP consists of two subcomponents: Identity Anchoring (IA), which establishes scale- and context-aware correspondence, and Identity-Aware Self-Attention (IASA), which refines context-dependent feature integration. During the denoising process over diffusion timesteps T , we switch from explicit structural anchoring (IA) to attention-based refinement (IASA) at timestep t within the unet.
Figure 4: Identity-token coverage. Unique identity-token coverage—the fraction of distinct identity tokens referenced by scene tokens—drops sharply as scale mismatch ∣logr∣ grows without size matching (orange), while size matching keeps it stable (blue). Low coverage means few tokens are repeatedly transferred.
Consistory+
Consistory+_+
CLIP-T ↑
CLIP-I ↑
DreamSim ↓
Aesthetic ↑
ImageReward ↑
CLIP-T ↑
CLIP-I ↑
DreamSim ↓
Aesthetic ↑
ImageReward ↑
SDXL (Base)
0.8971
0.8430
0.3124
6.1227
0.8289
0.9071
0.8376
0.3246
6.1508
1.1835
Training-Based
PhotoMaker [ 13 ]
0.8871
0.8342
0.3479
5.6051
0.8289
0.8776
0.8300
0.3594
5.6222
0.8717
IP-Adapter [ 36 ]
0.8771
0.9398
0.1396
5.8172
0.8283
0.8351
0.9410
0.1329
5.8367
0.5114
Training-Free
Table 1: Quantitative results on Consistory+ [ 17 ] and our extended Consistory+_+. Our method achieves better text alignment without compromising image quality. For StoryDiffusion [ 40 ] , we generate images at a resolution of 768×768 .
Figure 5: Top-1 framing classification accuracy on Consistory+_+. Evaluated with ShotVL-7B [ 15 ] , EOS-Story effectively captures the intended framing phrases.
Figure 6: Image comparison with previous methods. Our method faithfully reflects the given scenario while showing a flexible subject in each image.
Figure 8Table 9
Figure 9: Attention sample visualization. eosinf attends more strongly to semantically relevant regions.
Figure 11
Figure 11: Empirical analysis of ⟨EOS⟩ linearity. (Parallelism) Δapp from the same attribute stays aligned across subjects but separates across attributes, indicating a subject-agnostic direction. (Transferability) A transferred residual matches the explicit-prompt embedding ( transfer target ), unlike the base or a random attribute. (Bottom) Increasing λ in progressively strengthens the target appearance (round glasses), confirming controllable influence.
Figure 13
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 14
CLIP-T
(A) Baselines / naive combinations
eosorg
0.8935
eosscorg+eosapporg+eosstyorg
0.8625
eossc+eosid
0.8925
(B) Mixing-weight ablation (hyperparameter)
0.5eossc+0.5eosid
0.8864
Appendix
Table 6: eosinf construction ablation. We analyze text-alignment (CLIP-T) by varying (A) base EOS choice, (B) scaling factor of the injected direction, and (C) choice of appearance direction to inject.
Shot Size
Wide
Lens Size
Long lens
Shot Framing
Single
Camera Movement
Push in
Close up
Wide
Insert
Pull out
Extreme wide
Ultra wide (fisheye)
2 shot
Boom up
Medium close up
Medium
Group shot
Pan left
Medium wide
Lighting Type
Daylight
Establishing shot
Pan right
Medium
Artificial light
Over the shoulder
Tilt down
Extreme close up
Mixed light
3 shot
Tilt up
Appendix
Table 7: Framing phrases in ShotBench
Figure 16: Statistical analysis across scene structure categories. Each graph presents the mean and standard deviation of top-1 and top-3 accuracies for each option within a category. Results are averaged over 10 random seeds.
Figure 17: Effectiveness of adaptive thresholding. While different scenes require distinct α values (red), our adaptive threshold ( μ+σ ) provides reliable results by dynamically adjusting to each scene (blue).
Figure 19Figure 20Figure 21Figure 22
Figure 24: Free-form prompt generation results. Our model can handle free-form user prompts by parsing them into the intended EOS-Story format with an LLM.
Method
CLIP-T ↑
CLIP-I ↑
DreamSim ↓
Aesthetic ↑
ImageReward ↑
SDXL (Base)
0.8971
0.8430
0.3124
6.1227
0.8289
EOS-Story (Ours)
0.9042
0.8850
0.2005
6.1764
1.1680
EOS-Story (Free-form)
0.9040
0.8848
0.2001
6.1867
1.1712
Appendix
Table 11: Quantitative results on free-form story generation.
Figure 25: EOS-Story results with SD3 [ 5 ] .
CLIP-T ↑
CLIP-I ↑
DreamSim ↓
ImageReward ↑
Aesthetic ↑
FLUX (baseline)
0.8696
0.8555
0.2445
1.2988
6.0466
Large Pretrained
FLUX-Kontext
0.8544
0.9032
0.1602
1.2246
5.9489
Training-Free
CharaConsist
0.8697
0.8636
0.2358
1.2774
6.0572
EOS-Story (FLUX)
0.8722
0.8712
0.2052
1.2673
6.0943
Appendix
Table 12: Quantitative comparison under the FLUX backbone. Bold and underline denote the best and second-best.
Figure 26: EOS-Story on FLUX. Although the large pre-trained model (Flux-Kontext) better preserves the subject identity, it still struggles to reflect the given text (e.g., computer) and produces rigidly similar pose.
Figure 27: Image results on multi-characters.
Figure 29
Figure 30: Failure cases. There are cases where the baseline model (SDXL) struggles to generate certain concepts (e.g., “faun”). Additionally, since our method—like prior works—leverages the cross-attention map as a mask, the resulting masks sometimes capture irrelevant regions.
Figure 31: Additional qualitative comparison.
Figure 32: Additional qualitative comparison on framing phrase adaptation.
Figure 33
Figure 35: Long story generation. Our model accurately reflects detailed scene descriptions while preserving subject identity, even in extended scenarios.