Multimodality translation (e.g., text-to-image) is a core generative AI task. However, existing approaches (1) follow generative paths that do not directly represent the source modality, limiting the flexibility of some sampling algorithms; and (2) are unidirectional, preventing inversion (e.g., image-to-text). We propose BIT: Bidirectional Image-Text Diffusion Bridges. In contrast to previous approaches, BIT starts directly from text and interpolates into images, providing (1) a source-aware generative path that enables diverse and flexible sampling algorithms; and (2) an endpoint-conditioned process that can be traversed from image to text, providing a unified, bidirectional generative framework. BIT is derived through stochastic calculus, yielding SDE forms amenable to simulation and tractable loss functions that scale to high dimensions. Our experiments show that BIT is competitive with denoising-diffusion and deterministic-flow baselines, and outperforms them on several vision--language and natural-science evaluations.
Figures & tables
Figure 1: BIT Schematic: Forward and reverse-time SDEs (with drift parameterized by neural net) both traverse the same generative path, allowing text-image translation. Text is represented as invertible token embeddings (blue-ish texture at x0 ), with zero-padding (brown bottom at x0 ).
Figure 2: Sample Generations: Using our DiT-XL/2 model (1,054,621,840 parameters) trained for 122,500 iterations at batch size 1792. CFG scale ω=0.5 , 500 SDE steps.
Method
FID ( ↓ )
CLIP T2I ( ↑ )
Gen PPL ( ↓ )
CLIP I2T ( ↑ )
BIT
6.23
27.10
123.4
26.99
Diffusion (ablate data end)
5.73
27.13
178.2
26.42
Flow (ablate stochasticity)
279.13
20.13
169.2
27.25
Table 1: Quality comparison between BIT , noise-to-data diffusion ( Song et al., 2020 ) , and cross-modality flow matching ( Albergo et al., 2025 ) . We use 500 SDE steps, with ω=0 . T2I measured at 50k samples, I2T measured at 10k samples.
Figure 3: Cross-Modal Round-Trip Stochastic Variation: We want to generate semantically related variants of a given data sample.We run Alg. 1 with dt=0.004 with Euler-Maruyama simulation. We compare BIT to denoising diffusion and data-to-data flow matching. App. T shows qualitative examples.
Method
FID ( ↓ )
CLIP T2I ( ↑ )
Gen PPL ( ↓ )
CLIP I2T ( ↑ )
BIT ω=0.0
4.39
27.75
78.1
29.4
BIT ω=0.5
2.54
28.80
61.1
29.5
Qwen-3-VL-4B-Instruct
n/a
n/a
37.7
31.9
Stable Diffusion 1.5
9.92
33.15
n/a
n/a
Table 2: Impact of CFG, and Comparison to Foundation Models. We use 500 SDE steps.
Method
Fwd MSE ↓
Rev MSE ↓
Fwd MMD ↓
Fate kNN ↑
Clone kNN ↑
Cycle MSE ↓
Avg. Rank ↓
BIT , uniform shared
1.388
1.365
0.0446
0.404
0.182
1.307 / 1.418
3.33
BIT , uniform separate
1.415
1.394
0.0352
0.377
0.209
1.230 / 1.217
3.42
BIT , cosine shared
1.388
1.262
0.0431
0.445
0.160
1.042 / 1.136
2.67
Noise-to-data diffusion
1.329
1.381
0.0517
0.377
0.176
1.270 / 1.474
4.00
Rectified flow
2.444
1.694
0.0406
0.398
0.141
0.621 / 1.020
3.83
Endpoint regression
1.195
1.176
0.0847
0.336
0.047
1.076 / 1.102
3.75
Table 3: State-fate bridge results on LARRY day-2 to day-6 clone-paired endpoints. Average rank is computed per metric column (ties broken by midranks) and averaged over the six metrics; Cycle MSE sub-columns ranked separately and averaged so cycle consistency contributes a single metric.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Image Editing via Text Bridge.
Method
Gen PPL ( ↓ )
CLIP ( ↑ )
BIT (without REPA)
112.7
22.7
BIT (REPA λ=0.5 )
123.4
27.0
Appendix
Table 4: Impact of REPA on Image-to-Text: Same exact training and inference settings, except that one was trained with and one was trained without REPA.
Figure 5: Image Stochastic Variation: Qualitative results corresponding to Figure 3 . Because the text endpoint retains source-related information, BIT can recover semantically related variations of the source image after reversing the forward corruption process to t=0 .
Figure 6: Text Stochastic Variation: Qualitative text results, corresponding to Fig 3 . Corruption fraction 90% . Diffusion baseline goes off-topic, while BIT generates semantically related captions.