MUNITE: Unified Multimodal Latent Inference for Any-to-Any Multimodal Generation

Kyeongmin Yeo; Minhyuk Sung

cs.LGOct 7, 2026

MUNITE: Unified Multimodal Latent Inference for Any-to-Any Multimodal Generation

Authors: Kyeongmin Yeo, Minhyuk Sung

Organizations: KAIST

Abstract

We introduce MUNITE, a latent-variable framework for flexible any-to-any multimodal generation that treats encoding and latent generation as the same inference problem under different amounts of observed evidence. Given any subset of modalities, MUNITE models the conditional distribution over the latent representation associated with the complete observation. Full observation recovers deterministic encoding, no observation recovers the latent marginal, and intermediate subsets define conditional latent inference, all within a single conditional flow model. A shared latent sample captures variation that must remain consistent across generated targets, while modality-specific generative decoders model the remaining uncertainty independently. To learn these conditional distributions from incomplete training examples, we extend conditional flow matching through self-distillation: predictions conditioned on richer available observations supervise the same model conditioned on smaller subsets at the same intermediate latent state. When the richer-evidence trajectory follows the exact conditional flow, this provides the same expected learning signal as full-target denoising. Across PolyMNIST-D-Q, FFHQ64, and image-text-audio, MUNITE achieves competitive or better generation quality and source-target alignment, with higher joint-generation coherence. In particular, it attains the highest coherence in all one-to-many and unconditional image-text-audio comparisons, showing the effectiveness of unified latent inference across diverse multimodal settings.

Figures & tables

Figure 1: Latent inference and decoding. Left: latent distributions under different observation subsets. Right: factorized decoding from a latent value.

Panel of Figure 2 — Figure 2: MUNITE training objectives. (a) Reconstruction learns a shared representation, using target detaching to discourage copying modality-private information. (b) Self-distillation uses detached predictions conditioned on more modalities to supervise subset-conditioned predictions at the same noisy latent state, with both branches using the same model. (c) An auxiliary contrastive loss encourages semantic alignment between complementary modality subsets using in-batch negatives.

Panel of Figure 3 — Figure 3: Latent inference architecture. A shared Transformer performs latent inference under arbitrary observation subsets by aggregating modality information in latent registers. Keeping modality streams separate allows reconstruction gradients to be stopped on the target’s key/value paths to the registers (slashes), while retaining its information in the forward pass.

PolyMNIST-D-Q
Route	Input $\to$ output	Metric	MUNI	CFM	DFM	MUNITE
Uncond.	All five modalities	FD $\downarrow$	3.0407	71.6626	19.9587	1.5230
		Digit Coh. (all) $\uparrow$	0.5799	0.0237	0.2973	0.9826
		Quad. Coh. (all) $\uparrow$	0.9041	0.9902	0.8222	1.0000
$1\to 1$	$d\to m_{i}$	Digit Acc. $\uparrow$	0.9843	0.2402	0.7224	0.9941
$1\to 1$	$q\to m_{i}$	Quad. Acc. $\uparrow$	1.0000	0.9936	0.9372	1.0000

Table 1: Results on PolyMNIST-D-Q (top) and FFHQ64 (bottom). Acc. and Coh. denote accuracy and coherence,

d

and

q

denote digit and quadrant, and scores for

m_{i}

are averaged over the three views. FD uses verifier features, and Normal Err. is

1-\cos

. Best results are in bold.

	$1\to N$ coherence			Unconditional coherence
Method	$T\to(I,A)$	$I\to(T,A)$	$A\to(T,I)$	CLIP	CLAP	AIS
	AIS	CLAP	CLIP	$(T,I)$	$(T,A)$	$(I,A)$
CoDi	63.866	7.418	23.414	—	—	—
OmniFlow	77.027	14.781	22.697	21.17	14.23	50.95
MUNI	81.558	32.237	25.850	26.949	28.467	81.989
CFM	74.702	21.695	24.034	23.622	19.896	74.208

Table 2: Image–text–audio coherence between jointly generated modalities.

T

I

, and

A

denote text, image, and audio. OmniFlow’s unconditional scores are from Yeo et al. (2026, Table 4) , and dashes denote unsupported unconditional generation. Higher is better, and best results are in bold.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Width	Depth	Time cond.
PolyMNIST-D-Q
MUNITE	256	7	FiLM
CFM	256	16	FiLM
DFM	256	24	N/A
FFHQ64
MUNITE	512	11	FiLM

Appendix

Table 3: Transformer backbones.

Observed subset	Rows	%
$m_{0},m_{1},m_{2}$	55,000	25.0
$m_{0},m_{1}$	18,333	8.3
$m_{0},m_{2}$	18,334	8.3
$m_{1},m_{2}$	18,333	8.3
$m_{0}$	36,667	16.7
$m_{1}$	36,666	16.7

Appendix

Table 4: Training observations grouped by available modality subset. For PolyMNIST-D-Q and FFHQ64, categorical labels are observed in all groups.

T

I

, and

A

denote text, image, and audio, respectively.

Table 8

Task	FlowBind	MUNI	CFM	DFM	MUNITE
PolyMNIST-D-Q	—	32.09	20.09	20.57	19.33
FFHQ64	—	130.99	105.11	107.27	106.64
Image–text–audio	567.97	328.90	281.21	281.16	278.72

Appendix

Table 7: Trainable parameters in millions, including learned encoders, priors, and decoders. Frozen foundation models and the frozen VQ-VAEs of DFM, which have 13.27M parameters on PolyMNIST-D-Q and FFHQ64, are excluded.

Method	Epochs	Updates	Batch	LR
PolyMNIST-D-Q
MUNI, MUNITE	100	85,600	256	$1.5\times 10^{-4}$
CFM, DFM	200	171,200	256	$1.5\times 10^{-4}$
FFHQ64
MUNI, MUNITE	500	252,000	128	$1.5\times 10^{-4}$
CFM, DFM	1,000	504,000	128	$1.5\times 10^{-4}$

Appendix

Table 8: Training budgets. Batch sizes are global, and all methods use a constant learning rate. FlowBind’s epoch count assumes 702 updates per epoch.

PolyMNIST-D-Q
Method	Conditional	Unconditional	Decoding
MUNITE	Euler 50	Euler 50	Image Euler 50, CFG 1.5
MUNI	Posterior sampling	Euler 50	Image Euler 50, CFG 1.5
CFM	Euler 150	Euler 150	—
DFM	Random-order AR	Random-order AR	VQ decoder
FFHQ64

Appendix

Table 9: Inference and decoding methods. Euler step counts equal the number of function evaluations, CFG denotes classifier-free guidance, and AR denotes autoregressive sampling. Rendering of image–text–audio features is not included.

	TV $\downarrow$			Coherence $\uparrow$
Route	MUNI	MUNITE	Ref.	MUNI	MUNITE	Indep.	Chance
$d\to(m_{0},m_{1},m_{2})$	0.029	0.029	0.022	0.2528	1.0000	0.2500	0.25
$q\to(m_{0},m_{1},m_{2})$	0.109	0.034	0.024	0.1074	0.9904	0.1014	0.10

Appendix

Table 10: Missing attribute in PolyMNIST-D-Q one-to-many generation. TV is the total variation distance between the missing-label distribution of a generated view and the uniform data conditional, averaged over input labels and views; Ref. is its expected value for the same number of samples drawn from that conditional. Indep. is MUNITE with

m_{1}

and

m_{2}

decoded from separately drawn latents, keeping

m_{0}

and all decoder noise fixed. Chance is the coherence of independent uniform views.

	Image: FID $\downarrow$		Audio: FAD $\downarrow$		Text: CIDEr $\uparrow$
Method	$T\to I$	$A\to I$	$T\to A$	$I\to A$	$I\to T$	$A\to T$
CoDi	24.97	53.60	8.67	13.92	11.97	7.70
OmniFlow	20.52	97.56	4.08	4.87	26.60	31.42
MUNI	15.615	22.786	3.507	2.246	44.400	51.794
CFM	13.455	28.971	3.564	2.105	13.085	19.729
DFM	14.400	28.374	3.670	1.991	11.263	20.909

Appendix

Table 11: Image–text–audio one-to-one fidelity on the evaluation sets of Table 6 , grouped by target modality. CoDi and OmniFlow scores are taken from Yeo et al. (2026, Table 8) . Bold marks the best score in each column.

	Text–image: CLIP		Text–audio: CLAP		Image–audio: AIS
Method	$T\to I$	$I\to T$	$T\to A$	$A\to T$	$I\to A$	$A\to I$
CoDi	29.70	26.05	16.72	32.72	59.45	84.47
OmniFlow	30.76	27.48	30.07	44.74	76.52	64.22
MUNI	29.718	28.373	30.596	40.102	90.390	94.773
CFM	27.710	24.918	25.127	32.094	82.888	80.842
DFM	27.321	24.353	24.883	32.262	79.903	79.933

Appendix

Table 12: Image–text–audio one-to-one alignment. CoDi and OmniFlow scores are taken from Yeo et al. (2026, Table 9) . Higher is better, and bold marks the best score in each column.

	$(I,A)\to T$		$(T,A)\to I$		$(T,I)\to A$
Method	CLIP	CLAP	CLIP	AIS	CLAP	AIS
CoDi	24.05	33.72	24.98	85.52	11.06	65.31
OmniFlow	24.73	36.26	26.41	81.51	13.50	63.55
MUNI	27.922	36.695	27.056	85.168	30.524	82.621
CFM	25.221	31.889	25.712	80.305	24.907	76.098
DFM	24.855	33.259	25.398	80.347	25.141	75.846

Appendix

Table 13: Image–text–audio many-to-one alignment on 975 triplets, with each generated target scored against both source modalities. CoDi and OmniFlow scores are taken from Yeo et al. (2026, Table 3) . Higher is better, and bold marks the best score in each column.

Route	Input $\to$ output	Metric	Control	No target detach	No contrastive	Full-only distillation
$1\to 1$	$T\to I$	FID $\downarrow$	15.939	16.700	17.308	15.230
		CLIP $\uparrow$	29.007	28.543	27.750	29.593
	$I\to T$	CIDEr $\uparrow$	37.409	35.414	35.741	39.667
		CLIP $\uparrow$	27.526	27.261	26.919	27.882
	$T\to A$	FAD $\downarrow$	3.059	3.145	3.321	3.234
		CLAP $\uparrow$	30.695	30.458	31.461	30.540

Appendix

Table 14: Component ablations on image–text–audio. Control is the full MUNITE model of Tables 2 and 11 – 13 . Each variant removes one component and otherwise keeps the control’s architecture, optimization, and evaluation protocol. The

1\to N

and unconditional rows report coherence among jointly generated modalities. Best results in each row are in bold.

Figure 4: PolyMNIST-D-Q single-target generation. Icons show the observed labels: a digit, the conditioned quadrant shaded in a $2\times 2$ grid, or the digit placed in that quadrant. Each column generates the view listed under its input.

Figure 5: Unconditional PolyMNIST-D-Q co-generation. Each group is one joint sample of the three views and both labels, and its icon shows the generated digit in the generated quadrant.

Figure 6: PolyMNIST-D-Q one-to-many generation. Each row generates all three views from a digit (a) or a quadrant (b), so the views must agree on the unobserved attribute.

Figure 7: FFHQ64 RGB generation from age (a), gender (b), or both attributes (c).

Figure 8: Unconditional FFHQ64 co-generation of all five modalities. The generated gender and age labels of each joint sample appear below its maps.

Figure 9: FFHQ64 many-to-many generation of RGB, segmentation, and surface normals from age and gender. The three maps of each group come from one joint sample.

Figure 10: Image–text–audio one-to-one and many-to-one generation of images. Audio inputs are shown as waveforms.

Figure 11: Image–text–audio one-to-one and many-to-one generation of text. Audio inputs are shown as waveforms.

Figure 12: Unconditional image–text–audio co-generation by MUNITE. Each column decodes one latent sample into text, image, and audio.

Figure 13: One-to-many image–text–audio generation by MUNITE. In each column, the two outputs are decoded independently from one latent sample inferred from the input.

Explore similar work

MUNI: Multimodal Unified Latent Diffusion for Coherent Any-to-Any Generation
Jun 15, 2026Kyeongmin Yeo, Yunhong Min, Minhyuk SungMultimodal Generation Multimodal Diffusion Models
KAIST
LatentUMM: Dual Latent Alignment for Unified Multimodal Models
May 18, 2026Yinyi Luo, Wenwen Wang, Hayes Bai +2Multimodal Model Latent Representation Alignment
Carnegie Mellon University · William & Mary
Lance: Unified Multimodal Modeling by Multi-Task Synergy
May 18, 2026Fengyi Fu, Mengqi Huang, Shaojin Wu +10Multimodal Understanding Multi--Task Learning
Intelligent Creation Lab, ByteDance

Jun 15, 2026, cs.LGJ/K move · Enter open · S save

MUNI: Multimodal Unified Latent Diffusion for Coherent Any-to-Any Generation

Kyeongmin Yeo, Yunhong Min, Minhyuk Sung

KAIST

Multimodal Generation Multimodal Diffusion Models Latent Diffusion Model

Open paper AskPDF arXiv

We introduce MUNI, an end-to-end multimodal latent diffusion framework for any-to-any generation that unifies subset-conditioned cross-modal generation and unconditional joint sampling through a shared stochastic latent. Existing multimodal generative models are largely LLM-based, which limits leveraging modality-specific generators and requires text-paired data for training. Recent diffusion- and flow-based any-to-any extensions take a different direction but still rely on text-aligned embeddings, fully-paired training, or matched-dimensionality deterministic mappings. MUNI rests on two complementary contributions, one architectural and one in the training objective. First, we extend latent diffusion to multimodal any-to-any generation end-to-end: instead of the standard two-stage recipe that precomputes a frozen latent space and then fits a prior over it, MUNI jointly trains modality-specific encoders, expressive decoders, and a single shared flow-based prior under one objective. Second, we identify that the standard aggregation rules of multimodal variational inference are insufficient once coupled with a learned prior and expressive decoders. A suitable shared latent must simultaneously satisfy coherence across generated modalities, predictive sufficiency of subset latents, and minimality of the latent content. We propose a routed training objective whose structural choices align the latent with these criteria and admit a minimal-sufficiency characterization in the realizable setting. Experiments on PolyMNIST-Quadrant-Labels and a large-scale image-text-audio benchmark show MUNI matching or exceeding the strongest baselines on conditional generation while opening its largest margins on unconditional coherence. Project page: https://muni-proj.github.io/.

Abstract

Explore similar work

MUNI: Multimodal Unified Latent Diffusion for Coherent Any-to-Any Generation

LatentUMM: Dual Latent Alignment for Unified Multimodal Models

Lance: Unified Multimodal Modeling by Multi-Task Synergy