Pith. sign in

REVIEW 4 major objections 6 minor 12 cited by

Lavida-O claims that a single masked diffusion model—an 8B understanding branch paired with a 2.4B generation branch plus explicit planning and self-reflection—can simultaneously lead on text-to-image generation, image editing, and object g

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 15:22 UTC pith:55AZPDC5

load-bearing objection A real architectural contribution to unified multimodal MDMs, but the headline SOTA comes from test-set selection of inference hyperparameters and needs a fixed protocol before it's credible. the 4 major comments →

arxiv 2509.19244 v3 pith:55AZPDC5 submitted 2025-09-23 cs.CV

Lavida-O: Elastic Large Masked Diffusion Models for Unified Multimodal Understanding and Generation

Shufan Li , Jiuxiang Gu , Kangning Liu , Zhe Lin , Zijun Wei , Aditya Grover , Jason Kuen This is my paper
classification cs.CV
keywords masked diffusion modelsmultimodal understanding and generationtext-to-image generationimage editingobject groundingElastic-MoTplanning and reflectiondiscrete token diffusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that masked diffusion models, which generate discrete tokens by progressively unmasking them, can match or beat both autoregressive and continuous-diffusion rivals across the full multimodal spectrum: understanding images, locating objects, editing, and generating 1024px images from text. To do this it introduces an architecture that couples a small generation branch to a large understanding branch, plus sampling and conditioning tricks, and then uses the model's own understanding to plan and critique its generations. The reported results put Lavida-O ahead of specialist models on grounding and editing benchmarks, at up to 6.8x faster grounding latency. If correct, this would establish masked diffusion as a viable single-model path to generalist multimodal AI.

Core claim

The central discovery is that a single masked diffusion transformer can become state-of-the-art for both understanding and generation by separating capacities into an 8B understanding branch and a 2.4B generation branch that only interact in the early layers. This Elastic-MoT design keeps joint attention where modalities need to exchange information, but prevents text and image tokens from interfering later, and lets the model activate only the parameters a task needs. On top of this, the paper shows that the model's own understanding abilities—grounding and self-critique—can be explicitly scheduled as planning and reflection, yielding large gains on prompt following and editing without any

What carries the argument

The central mechanism is Elastic-MoT (Elastic Mixture-of-Transformers), a two-branch transformer where the generation branch is smaller than the understanding branch and joint attention is limited to the first 16 of 32 layers. This lets the model load only 6.4B parameters for text-to-image generation, 8B for understanding, and 10.4B for interleaved tasks. Supporting it are modality-aware masking, which decides at decode time which tokens route to which branch via a special expansion token; universal text conditioning, which feeds aesthetic and luminance conditions as plain text; and stratified random sampling, which unmasks tokens in a spatially balanced pattern to respect the independence a

Load-bearing premise

The fixed Meissonic VQ encoder is assumed to represent images well enough for high-fidelity generation and editing; the paper admits it cannot faithfully reconstruct small text, so any reconstruction ceiling from that tokenizer caps the entire model's generation quality.

What would settle it

Generate a 1024px image with a sign or poster containing a short sentence, using Lavida-O's standard pipeline, and check whether the text is legible. If small text comes out garbled, the tokenizer's known limitation directly limits the model's text rendering, contradicting the implied claim of overall text-to-image parity; alternatively, retraining the same architecture with a text-faithful tokenizer and showing GenEval/DPG scores stay flat would indicate the bottleneck lies elsewhere.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Text-to-image quality on GenEval reaches 0.89 with reflection, beating continuous diffusion models like SD3 and Flux-dev, with DPG 83.2 and FID 6.68 on MJHQ-30k (without reflection/planning active there).
  • Object grounding on RefCOCO/RefCOCO+/RefCOCOg reaches Precision@0.5 up to 90.6, surpassing Qwen2.5-VL-7B and InternVL3-8B while being up to 6.8x faster.
  • Image editing overall score on ImageEdit reaches 3.80 with planning, outperforming GPT-4o on replace and remove categories and beating specialist editors like FluxKontext-dev.
  • Training efficiency improves 3.17x over a standard MoT architecture due to the smaller generation branch and decoupled attention.
  • Explicit planning and self-reflection, both executed by the model itself, provide measurable improvements: planning adds +0.19 on object position in GenEval, and reflection adds +0.24 on position and +0.14 on counting over baseline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the tokenizer is the bottleneck, the same architecture with a better or trainable image tokenizer could push text rendering and fine details further without any architectural change.
  • The planning/reflection loop suggests a general pattern for unified MDMs: the understanding branch acts as an in-model verifier for the generation branch, giving inference-time scaling without external reward models; this could extend to other structured interleaved tasks like layout-conditioned graphic design or multi-turn editing.
  • The speed-quality tradeoff from parallel decoding of quantized coordinates could transfer to other fixed-format structured predictions, such as keypoint detection or 3D bounding box estimation.
  • Since the reflection gains saturate by N=8 rounds, a practical serving system could use a low fixed budget and still capture most of the benefit, while simple prompts would early-stop and save latency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. Lavida-O is a unified masked diffusion model (MDM) that combines an 8B understanding branch (initialized from LaViDa) with a smaller 2.4B generation branch in an Elastic Mixture-of-Transformers (Elastic-MoT) architecture. The paper introduces several techniques: token compression, modality-aware masking with a special [exp] token, universal text conditioning expressed as plain text, stratified random sampling, coordinate quantization for grounding, and planning/reflection loops that leverage the model's own understanding to improve generation. The model is trained in three stages and evaluated on image understanding, text-to-image generation (GenEval, DPG-Bench, MJHQ FID-30k), object grounding (RefCOCO/RefCOCO+/RefCOCOg), and image editing (Image-Edit Bench). The paper claims state-of-the-art performance for a multi-modal MDM on these tasks, citing, for example, GenEval 0.89 with reflection, RefCOCO P@0.5 up to 90.6, Image-Edit overall 3.80, and a 6.8x inference speedup on grounding.

Significance. If the empirical claims hold, Lavida-O is a meaningful architectural advance: it is the first unified MDM to demonstrate competitive object grounding, instruction-based editing, and 1024px text-to-image generation in one framework, and it introduces reusable components (Elastic-MoT, modality-aware masking, stratified sampling) that could benefit future work. The paper is unusually transparent about training data, hyperparameters, and compute (53k GPU hours), and it contains extensive ablations (Tables 8-15) plus an honest limitations section. The planning/reflection paradigm is a promising direction for leveraging understanding to improve generation. However, the central SOTA claims are currently undermined by test-set selection of inference hyperparameters and by the lack of released code, weights, or error bars, so the significance is conditional on these issues being resolved.

major comments (4)
  1. [Section 4.2, Table 3; Appendix B.3, Table 10] The reported FID-30k of 6.68 is a post-hoc selection: Table 10 evaluates confidence (11.42), uniform (8.22), Halton (7.38), and stratified (6.68) samplers on the same MJHQ-30k set and the best value is then used in Table 3. Under confidence sampling, the same model's FID (11.42) is worse than Flux-dev's (10.15). Thus the claim that Lavida-O surpasses continuous diffusion models on FID is not a property of a fixed inference protocol. Please pre-register the sampler on a validation split or report the full sampler comparison in Table 3; otherwise the FID comparison is not valid.
  2. [Section 4.2, Table 3; Appendix B.5, Table 15] The GenEval score of 0.89 is the maximum of a sweep over reflection rounds N (0.848 at N=1, 0.890 at N=12, 0.886 at N=20), and the +Reflection row uses an inference-time scaling procedure with up to ~53 s/image. Baselines such as Flux-dev and SD3-Medium are single-pass, so the comparison conflates test-set-selected compute budget with model quality. The paper should specify a fixed inference budget (including N and the critic stopping criterion) and report the GenEval-vs-N curve with variance; otherwise the 'SOTA text-to-image' claim is not established.
  3. [Section D (Limitations)] The paper itself states that the Meissonic VQ tokenizer 'cannot faithfully reconstruct small texts' and that Lavida-O's text rendering is 'very limited.' Since every image output passes through this fixed tokenizer, this caps text rendering and fine-grained fidelity independent of the diffusion model. This is load-bearing for the text-to-image claim: the paper should quantify tokenizer reconstruction (e.g., rFID, text rendering accuracy) and, if possible, report GenEval/DPG subsets involving text. The limitation is not disqualifying but needs to be contextualized in the main results.
  4. [General (all tables)] All benchmark numbers are single runs without error bars, significance tests, or released code/weights/evaluation scripts. For a paper whose central contribution is 'state-of-the-art' empirical performance, this makes it impossible to determine whether the reported gaps (e.g., GenEval 0.89 vs. 0.82 for BAGEL; FID 6.68 vs. 10.15 for Flux-dev) are reproducible or within run-to-run noise. Please release the model and evaluation code, and report mean±std over at least a few seeds for the core tables.
minor comments (6)
  1. [Section 3.1.1] The name 'Elastic-MoT' is used inconsistently (Elastic-MoT, ElasticMOT, Elastic-MoT in Figure 3). Please unify the spelling.
  2. [Table 8] The claim that the 2B generation branch 'achieves the best balance' rests on differences as small as 0.01 in GenEval (0.57 vs. 0.56) with no error bars; the ranking is fragile. Please add confidence intervals or soften the claim.
  3. [Section 3.2] Typo: 'introudced' should be 'introduced'.
  4. [Appendix B.3] The sampler comparison uses only FID on MJHQ-30k. Since FID is widely acknowledged to be insensitive to prompt fidelity, consider reporting GenEval or DPG under the different samplers as well.
  5. [Section 4.2 / Table 4] The abstract claims 'state-of-the-art' grounding, but Table 4 shows InternVL3-8B is higher on RefCOCO val (92.5 vs. 92.3). Please either qualify the claim or explain why the difference is not material.
  6. [Appendix B.2] The Stage-3 ablation for M (Table 9) is run from a checkpoint trained with M=16, so comparisons across M in Stage 3 are confounded; the paper acknowledges this, but the caveat should appear near the main Elastic-MoT discussion.

Circularity Check

2 steps flagged

T2I headline numbers are selected on the evaluation benchmarks (sampler and reflection-budget sweeps), making the SOTA claim partially a fitted result; otherwise no self-definitional circularity.

specific steps
  1. fitted input called prediction [Appendix B.3 (Table 10) vs. Section 4.2 (Table 3)]
    "Table 10: Performance of Different Samplers in Text-to-Image Generation Tasks. ... Method FID-30k↓ Confidence 11.42 Uniform 8.22 Halton 7.38 Stratified6.68"

    The FID-30k value 6.68 reported as Lavida-O's text-to-image result in Table 3 is the winning entry of a four-way sampler comparison computed on the same MJHQ-30k test set (Table 10). Under confidence sampling the same model yields 11.42, which is worse than Flux-dev's 10.15 in Table 3. The sampler is an inference component of the method, so the reported comparison to continuous diffusion models is not a prediction under a fixed protocol but a selection of the best FID on the target benchmark. The paper presents the selected value as method performance without a validation-based choice.

  2. fitted input called prediction [Appendix B.5 (Table 15) vs. Section 4.2 (Table 3)]
    "Num. of Reflection RroundsN=1 N=2 N=4 N=8 N=12 N=16 N=20 / GenEval Score↑0.848 0.864 0.875 0.882 0.890 0.886 0.886"

    The headline GenEval 0.89 for '+Reflection' in Table 3 is exactly the maximum of a sweep over the reflection budget N run directly on GenEval (N=12 in Table 15). The paper does not specify N as a pre-registered setting; it reports the best value in the sweep. At N=1 the same model with planning scores 0.848, and at N=20 it scores 0.886. Thus the central text-to-image SOTA claim reduces to selecting an inference-time hyperparameter on the same benchmark used to support the claim, rather than being an independent prediction.

full rationale

The core architecture is not circular: Elastic-MoT is initialized from the authors' prior LaViDa weights, but that is a concrete model transfer, not an unverified citation; grounding, editing, and understanding results are evaluated on external benchmarks and do not derive from fitted quantities. The acknowledged VQ-tokenizer text-rendering limitation (Appendix D) is a genuine cap, not a circular step. The circularity is confined to the T2I headline numbers: both the FID-30k (6.68) and GenEval with reflection (0.89) are selected as best values over inference-setting sweeps performed on the same evaluation sets (Tables 10 and 15). Reporting these selected maxima as the method's state-of-the-art performance makes the 'outperforms continuous diffusion models' claim partially a fitted-input result. This does not invalidate the model design or the grounding/edit results, but it does mean the central T2I claim should be re-examined under a fixed protocol.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The central claims rest on a substantial body of prior work (LaViDa base, Meissonic tokenizer, MDM theory from Sahoo et al.) and on several choices tuned on the evaluated benchmarks rather than derived from first principles. No formal proofs or machine-checked components are provided. The most consequential free parameters are the architecture sizes and the undisclosed evaluation-time conditioning values.

free parameters (6)
  • Generation branch size and hidden dimension = 2.4B parameters, hidden size 2048
    Selected by ablation in Table 8; no theoretical justification.
  • Number of joint attention layers M = 16 of 32 layers
    Selected by ablation in Table 9; M=16 gives best speed-quality balance.
  • Token compression factor = 4x reduction in VQ tokens
    Chosen by design for efficiency; no ablation shown.
  • Universal text conditioning values at evaluation = Unspecified
    The paper does not state the SCORE, HPS, luminance, contrast values used for Table 3 results; these could materially affect the numbers.
  • Maximum reflection rounds N = 20 with early stopping
    Chosen empirically in Table 15; gains saturate around N=8-12.
  • Dataset mix scheduler ratio = New:Old from 3:1 to 1:3
    Chosen by ablation in Table 11.
axioms (4)
  • domain assumption The MDM reverse process factorizes over token positions, p_theta(X_{t_{k-1}} | X_{t_k}) = product_i p_theta(X^i_{t_{k-1}} | X^i_{t_k})
    Invoked in Section A.1 and used throughout; stratified sampling is designed to compensate for violations of this independence.
  • domain assumption The Meissonic VQ encoder provides a token space sufficient for high-fidelity 1024px image generation and editing
    The tokenizer is used as a fixed frontend for all generation tasks; the paper acknowledges it cannot reconstruct small text (Section D).
  • domain assumption A 2.4B generation branch initialized from truncated weights can match much larger continuous diffusion models
    Core of Elastic-MoT; only supported by the authors' ablations (Table 8, Figure 14).
  • ad hoc to paper Planning and reflection data produced by GroundingDINO and the authors' reflection datasets are valid supervision for improving generation
    The data pipeline in Section B.1 constructs layouts and reflection pairs internally without external validation of their quality.
invented entities (1)
  • [exp] token no independent evidence
    purpose: Signals where an image should be generated in the modality-aware masking process; at inference it expands into a sequence of image VQ mask tokens
    A new special token whose meaning is defined solely by the paper's training procedure; no external handle.

pith-pipeline@v1.3.0-alltime-deepseek · 27625 in / 11765 out tokens · 88179 ms · 2026-08-04T15:22:45.676893+00:00 · methodology

0 comments
read the original abstract

We propose Lavida-O, a unified Masked Diffusion Model (MDM) for multimodal understanding and generation. Unlike existing multimodal MDMs such as MMaDa and Muddit which only support simple image-level understanding tasks and low-resolution image generation, Lavida-O presents a single framework that enables image-level understanding, object grounding, image editing, and high-resolution (1024px) text-to-image synthesis. Lavida-O incorporates a novel Elastic Mixture-of-Transformers (Elastic-MoT) architecture that couples a lightweight generation branch with a larger understanding branch, supported by token compression, universal text conditioning and stratified sampling for efficient and high-quality generation. Lavida-O further incorporates planning and iterative self-reflection in image generation and editing tasks, seamlessly boosting generation quality with its understanding capabilities. Lavida-O achieves state-of-the-art performance on a wide range of benchmarks including RefCOCO object grounding, GenEval text-to-image generation, and ImgEdit image editing, outperforming existing autoregressive models and continuous diffusion models such as Qwen2.5-VL and FluxKontext-dev, while offering considerable speedup at inference. These advances establish Lavida-O as a new paradigm for scalable multimodal reasoning and generation.

Figures

Figures reproduced from arXiv: 2509.19244 by Aditya Grover, Jason Kuen, Jiuxiang Gu, Kangning Liu, Shufan Li, Zhe Lin, Zijun Wei.

Figure 1
Figure 1. Figure 1: We propose Lavida-O, a unified masked diffusion model capable of multi-modal under￾standing and generation. Abstract We propose Lavida-O, a unified Masked Diffusion Model (MDM) for multi￾modal understanding and generation. Unlike existing multimodal MDMs such as MMaDa and Muddit which only support simple image-level understanding tasks and low-resolution image generation, Lavida-O presents a single frame￾w… view at source ↗
Figure 2
Figure 2. Figure 2: Overall Pipeline of Lavida-O. Given an input image and text prompt, we first concate￾nate the image semantic embedding Ci , image VQ embedding Cv, and text prompt embedding Cp to form the conditioning embedding C. The combined embedding is then passed to the model along￾side the partially masked sequence Xt. The model then predicts the fully-unmasked sequence X0. the conditional embedding C = Concat(Ci , C… view at source ↗
Figure 3
Figure 3. Figure 3: Design of Elastic MoT. Elastic-MoT introduces two major modifications to standard MoT. First, the generation branch has a smaller hidden size. Second, given an N-layer model, we only allow text and image modalities to interact in the first M layers. These two designs allow us to flexibly load only a portion of parameters depending on tasks, improving the efficiency. 3.1.2 Modality-aware Masking One of the … view at source ↗
Figure 4
Figure 4. Figure 4: Design choices of Lavida-O. (a) Forward diffusion process with modality-aware masking. (b) Visualization of the unmasking order in the proposed stratified random sampling process. unmasked tokens. This negatively affecting image quality because adjacent tokens are highly corre￾lated, which contradicts the independence assumption of MDMs. To mitigate this, we introudced a stratified sampling process. Starti… view at source ↗
Figure 5
Figure 5. Figure 5: Training and Inference Speed of Lavida-O. We compare the end-to-end inference la￾tency of Lavida-O on three tasks, as well as pretraining efficiency measured by per-step latency. 4.4 Additional Qualitative Results Finally, we provide additional qualitative examples demonstrating Lavida-O’s capabilities on diverse prompts and editing instructions [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative examples of text-to-image generation. We provide additional examples of text-to-image generation outputs on diverse prompts. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative examples of image editing. We provide additional examples of image editing outputs on diverse instructions. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Activated parameters of Lavida-O under different task settings. Elastic-MoT Design allow Lavida-O to dynamically loads its parameters depending on the tasks. For understanding-only tasks, we only load the 8B generation branch. For text-to-image generation tasks, we load the first M = 16 layers of the understanding branch, which consists of 4B parameters, and the full 2.4B generation branch. For interleaved… view at source ↗
Figure 9
Figure 9. Figure 9: Training and inference with modality-aware masking. We visualize the sampling process with modality-aware masking on the left and the training process on the right. During the training, the loss is applied on either X0 or X′ 0 depending on the value of t with respect to texp. A.3 Modality-Aware Masking In this section, we provide details of the changes to the training and sampling process introduced by mod… view at source ↗
Figure 10
Figure 10. Figure 10: Effect of Universal Text Conditioning. On the left side, we visualize the text format used in Universal Text Conditioning. On the right side, we visualize generation results under differ￾ent choices of universal text conditioning. at which a sequence of VQ image tokens is collapsed into a single [exp] text token. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of different sampling processes. We compare the unmasking order of the stratified sampler, Halton sampler, and uniform random sampler. Uniform random sampler produces the least desirable spatial pattern, with many unmasked tokens clustered together. Halton sampler is less ideal than stratified sampler because it does not guarantee perfectly stratified coverage. For example, when the number o… view at source ↗
Figure 12
Figure 12. Figure 12: Coordinate Quantization. We normalize bounding box coordinates into the range [0,1] and discretize them into 1025 bins. This ensures that each bounding box is represented by exactly 4 tokens, allowing efficient parallel decoding of multiple bounding boxes in a single step. Our design is inspired by the stratified sampling process commonly used in numerical integration and computer graphics. It also follow… view at source ↗
Figure 13
Figure 13. Figure 13: Interleaved Generation with Planning and Reflection. We provide visual examples of interleaved generation, including text-to-image generation with planning (Top), text-to-image generation with reflection (Middle), and image editing with planning (Bottom). We always enable planning during the reflection process. The layout traces is omitted in the middle figure for clarity and better presentation. simple j… view at source ↗
Figure 14
Figure 14. Figure 14: Effect of truncated initialization. Validation loss comparison of truncated initialization vs. training from scratch during Stage 2. Truncated initialization converges faster and achieves lower loss [PITH_FULL_IMAGE:figures/full_fig_p028_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Speed–quality tradeoff on generation, grounding, and reasoning. Latency (s/sample) and benchmark scores are shown. For MJHQ: FID (lower is better). For RefCOCO: Precision@0.5 (higher is better). For MathVista: accuracy (higher is better). On MathVista, the maximum genera￾tion length is capped at 256 tokens. MJHQ-30k text-to-image generation (Li et al., 2024b), RefCOCO grounding (Kazemzadeh et al., 2014), … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 12 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Language Generation as Optimal Control: Closed-Loop Diffusion in Latent Control Space

    cs.CL 2026-05 unverdicted novelty 7.0

    The paper introduces Manta-LM, which approximates the Hamilton-Jacobi-Bellman optimal policy via Flow Matching in a rectified latent control space to enable high-fidelity parallel language generation.

  2. UniEditBench: A Unified and Cost-Effective Benchmark for Image and Video Editing via Distilled MLLMs

    cs.CV 2026-04 unverdicted novelty 7.0

    UniEditBench unifies image and video editing evaluation with a nine-plus-eight operation taxonomy and cost-effective 4B/8B distilled MLLM evaluators that align with human judgments.

  3. Nemotron-Labs-Diffusion-Image: Advancing Masked Discrete Diffusion for High-Resolution Image Synthesis

    cs.CV 2026-06 unverdicted novelty 6.0

    A masked discrete diffusion model adds token editing at inference and grouped cross-entropy training to reach 0.90 GenEval, 86.9 DPG, and 10.76 HPSv3 scores.

  4. Nemotron-Labs-Diffusion-Image: Advancing Masked Discrete Diffusion for High-Resolution Image Synthesis

    cs.CV 2026-06 conditional novelty 6.0

    Masked discrete diffusion with token editing and grouped cross-entropy reaches strong text-to-image generation scores in an 8B decoder-only model, reporting GenEval 0.90, DPG 86.9, HPSv3 10.76.

  5. Guidance Contrastive Token Credit Assignment for Discrete Policy Optimization

    cs.CV 2026-05 unverdicted novelty 6.0

    GCPO performs per-token credit assignment in discrete policy optimization by setting token advantages proportional to the difference in model predictions under positive versus negative prompts, outperforming GRPO and ...

  6. Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models

    cs.AI 2026-05 unverdicted novelty 6.0

    Proposes HT-GRPO with sketch-then-paint staged updates, prompt-conditioned importance ratios, and hierarchical credit assignment for dMLLMs, reporting gains on GenEval and DPG plus quality metrics.

  7. Language Generation as Optimal Control: Closed-Loop Diffusion in Latent Control Space

    cs.CL 2026-05 unverdicted novelty 6.0

    Language generation is recast as optimal control and solved approximately with flow matching in rectified latent control space to enable high-fidelity parallel text generation.

  8. Language Generation as Optimal Control: Closed-Loop Diffusion in Latent Control Space

    cs.CL 2026-05 unverdicted novelty 6.0

    Manta-LM approximates the HJB equation via flow matching in latent control space to realize closed-loop optimal control for language generation.

  9. Dataset-Level Metrics Attenuate Non-Determinism: A Fine-Grained Non-Determinism Evaluation in Diffusion Language Models

    cs.LG 2026-04 unverdicted novelty 6.0

    Dataset-level metrics in diffusion language models mask substantial sample-level non-determinism that varies with model and system factors, which a new Factor Variance Attribution metric can decompose.

  10. Sparse-LaViDa: Sparse Multimodal Discrete Diffusion Language Models

    cs.CV 2025-12 conditional novelty 6.0

    Sparse-LaViDa speeds up masked diffusion LLMs by truncating masked tokens, using register tokens and a step-causal attention mask, reaching ~2-2.8x speedup over LaViDa-O with comparable generation quality.

  11. DataEvolver: Let Your Data Build and Improve Itself via Goal-Driven Loop Agents

    cs.AI 2026-05 unverdicted novelty 5.0

    DataEvolver introduces a reusable framework with generation-time self-correction and validation-time self-expansion loops that improves visual datasets, shown to outperform baselines on an object-rotation task.

  12. Illuminating Unified Multimodal Model for Free-form Interleaved Text-Image Generation

    cs.CV 2026-06 unverdicted novelty 4.0

    ILLUME-X is a unified multimodal model that generates free-form interleaved text-image sequences via an expanded data pipeline, progressive self-adaptive training, and ILScore evaluation, claiming outperformance over ...

Reference graph

Works this paper leans on

83 extracted references · 40 linked inside Pith · cited by 9 Pith papers

  1. [1]

    Structured denoising diffusion models in discrete state-spaces

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in neural information processing systems, 34: 0 17981--17993, 2021

    2021

  2. [2]

    Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis

    Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Halton scheduler for masked generative image transformer

    Victor Besnier, Mickael Chen, David Hurych, Eduardo Valle, and Matthieu Cord. Halton scheduler for masked generative image transformer. arXiv preprint arXiv:2503.17076, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Instructpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 18392--18402, 2023

    2023

  6. [6]

    Coyo-700m: Image-text pair dataset

    Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022

    2022

  7. [7]

    Maskgit: Masked generative image transformer

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 11315--11325, 2022

    2022

  8. [8]

    Muse: Text-to-image generation via masked generative transformers

    Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset

    Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, et al. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset. arXiv preprint arXiv:2505.09568, 2025 a

    Pith/arXiv arXiv 2025

  10. [10]

    Sharegpt-4o-image: Aligning multimodal models with gpt-4o-level image generation

    Junying Chen, Zhenyang Cai, Pengcheng Chen, Shunian Chen, Ke Ji, Xidong Wang, Yunjin Yang, and Benyou Wang. Sharegpt-4o-image: Aligning multimodal models with gpt-4o-level image generation. arXiv preprint arXiv:2506.18095, 2025 b

    Pith/arXiv arXiv 2025

  11. [11]

    Open-llava-next: An open-source implementation of llava-next series for facilitating the large multi-modal model community

    Lin Chen and Long Xing. Open-llava-next: An open-source implementation of llava-next series for facilitating the large multi-modal model community. https://github.com/xiaoachen98/Open-LLaVA-NeXT, 2024

    2024

  12. [12]

    Janus-pro: Unified multimodal understanding and generation with data and model scaling

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-pro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025 c

    Pith/arXiv arXiv 2025

  13. [13]

    Emerging properties in unified multimodal pretraining

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp.\ 4171--4186, 2019

    2019

  15. [15]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 12873--12883, 2021

    2021

  16. [16]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M \"u ller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024

    2024

  17. [17]

    Mme: A comprehensive evaluation benchmark for multimodal large language models

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Geneval: An object-focused framework for evaluating text-to-image alignment

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36: 0 52132--52152, 2023

    2023

  19. [19]

    Mammoth-vl: Eliciting multimodal reasoning with instruction tuning at scale

    Jarvis Guo, Tuney Zheng, Yuelin Bai, Bo Li, Yubo Wang, King Zhu, Yizhi Li, Graham Neubig, Wenhu Chen, and Xiang Yue. Mammoth-vl: Eliciting multimodal reasoning with instruction tuning at scale. arXiv preprint arXiv:2412.05237, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Unified discrete diffusion for simultaneous vision-language generation

    Minghui Hu, Chuanxia Zheng, Heliang Zheng, Tat-Jen Cham, Chaoyue Wang, Zuopeng Yang, Dacheng Tao, and Ponnuthurai N Suganthan. Unified discrete diffusion for simultaneous vision-language generation. arXiv, 2022

    2022

  21. [21]

    Equip diffusion models with llm for enhanced semantic alignment

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu Ella. Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 5 0 (7): 0 16, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Visualwebinstruct: Scaling up multimodal instruction data through web search

    Yiming Jia, Jiachen Li, Xiang Yue, Bo Li, Ping Nie, Kai Zou, and Wenhu Chen. Visualwebinstruct: Scaling up multimodal instruction data through web search. arXiv preprint arXiv:2503.10582, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Referitgame: Referring to objects in photographs of natural scenes

    Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp.\ 787--798, 2014

    2014

  24. [24]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 4015--4026, 2023

    2023

  25. [25]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024

    2024

  26. [26]

    Flux.1 kontext: Flow matching for in-context image generation and editing in latent space, 2025

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Müller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context image ...

    Pith/arXiv arXiv 2025

  27. [27]

    Llava-onevision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024 a

    Pith/arXiv arXiv 2024

  28. [28]

    mplug: Effective and efficient vision-language learning by cross-modal skip-connections

    Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022

    Pith/arXiv arXiv 2022

  29. [29]

    Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024 b

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024 b

    2024

  30. [30]

    Mask dino: Towards a unified transformer-based framework for object detection and segmentation

    Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3041--3050, 2023 a

    2023

  31. [31]

    Instructany2pix: Flexible visual editing via multimodal instruction following

    Shufan Li, Harkanwar Singh, and Aditya Grover. Instructany2pix: Flexible visual editing via multimodal instruction following. arXiv preprint arXiv:2312.06738, 2023 b

    Pith/arXiv arXiv 2023

  32. [32]

    Omniflow: Any-to-any generation with multi-modal rectified flows

    Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Zichun Liao, Yusuke Kato, Kazuki Kozuka, and Aditya Grover. Omniflow: Any-to-any generation with multi-modal rectified flows. arXiv preprint arXiv:2412.01169, 2024 c

    Pith/arXiv arXiv 2024

  33. [33]

    Lavida: A large diffusion language model for multimodal understanding

    Shufan Li, Konstantinos Kallidromitis, Hritik Bansal, Akash Gokul, Yusuke Kato, Kazuki Kozuka, Jason Kuen, Zhe Lin, Kai-Wei Chang, and Aditya Grover. Lavida: A large diffusion language model for multimodal understanding. arXiv preprint arXiv:2505.16839, 2025 a

    Pith/arXiv arXiv 2025

  34. [34]

    Reflect-dit: Inference-time scaling for text-to-image diffusion transformers via in-context reflection

    Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Arsh Koneru, Yusuke Kato, Kazuki Kozuka, and Aditya Grover. Reflect-dit: Inference-time scaling for text-to-image diffusion transformers via in-context reflection. arXiv preprint arXiv:2503.12271, 2025 b

    Pith/arXiv arXiv 2025

  35. [35]

    Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models

    Weixin Liang, Lili Yu, Liang Luo, Srinivasan Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, et al. Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models. arXiv preprint arXiv:2411.04996, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Uniworld: High-resolution semantic encoders for unified visual understanding and generation

    Bin Lin, Zongjian Li, Xinhua Cheng, Yuwei Niu, Yang Ye, Xianyi He, Shenghai Yuan, Wangbo Yu, Shaodong Wang, Yunyang Ge, et al. Uniworld: High-resolution semantic encoders for unified visual understanding and generation. arXiv preprint arXiv:2506.03147, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Llava-next: Improved reasoning, ocr, and world knowledge, January 2024 a

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-next: Improved reasoning, ocr, and world knowledge, January 2024 a . URL https://llava-vl.github.io/blog/2024-01-30-llava-next/

    2024

  38. [38]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European conference on computer vision, pp.\ 38--55. Springer, 2024 b

    2024

  39. [39]

    Step1x-edit: A practical framework for general image editing

    Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, et al. Step1x-edit: A practical framework for general image editing. arXiv preprint arXiv:2504.17761, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pp.\ 216--233

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pp.\ 216--233. Springer, 2024 c

    2024

  41. [41]

    Discrete diffusion modeling by estimating the ratios of the data distribution

    Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834, 2023

    Pith/arXiv arXiv 2023

  42. [42]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022

    2022

  43. [43]

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023

    Pith/arXiv arXiv 2023

  44. [44]

    C hart QA : A benchmark for question answering about charts with visual and logical reasoning

    Ahmed Masry, Do Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. C hart QA : A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pp.\ 2263--2279, Dublin, Ireland, May 2022. Association for Computational Linguistics. doi:10.18653/v1/2022.findings-acl.17...

    doi:10.18653/v1/2022.findings-acl.177 2022

  45. [45]

    Docvqa: A dataset for vqa on document images

    Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp.\ 2200--2209, 2021

    2021

  46. [46]

    Large language diffusion models

    Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Large language diffusion models. arXiv preprint arXiv:2502.09992, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Dall·e 3

    OpenAI. Dall·e 3. https://openai.com/index/dall-e-3/, 2023

    2023

  48. [48]

    Gpt-4o system card

    OpenAI. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. URL https://arxiv.org/abs/2410.21276

    Pith/arXiv arXiv 2024

  49. [49]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M \"u ller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    Pith/arXiv arXiv 2023

  50. [50]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PmLR, 2021

    2021

  51. [51]

    Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S

    Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M. Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S. Khan. Glamm: Pixel grounding large multimodal model. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  52. [52]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  53. [53]

    Simple and effective masked diffusion language models

    Subham Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin Chiu, Alexander Rush, and Volodymyr Kuleshov. Simple and effective masked diffusion language models. Advances in Neural Information Processing Systems, 37: 0 130136--130184, 2024

    2024

  54. [54]

    Laion-aesthetics

    Christoph Schuhmann. Laion-aesthetics. https://laion.ai/blog/laion-aesthetics/, 2022. Accessed: 2024 - 03 - 06

    2022

  55. [55]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35: 0 25278--25294, 2022

    2022

  56. [56]

    Muddit: Liberating generation beyond text-to-image with a unified discrete diffusion model

    Qingyu Shi, Jinbin Bai, Zhuoran Zhao, Wenhao Chai, Kaidong Yu, Jianzong Wu, Shuangyong Song, Yunhai Tong, Xiangtai Li, Xuelong Li, et al. Muddit: Liberating generation beyond text-to-image with a unified discrete diffusion model. arXiv preprint arXiv:2505.23606, 2025

    Pith/arXiv arXiv 2025

  57. [57]

    Journeydb: A benchmark for generative image understanding

    Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. Advances in neural information processing systems, 36: 0 49659--49678, 2023

    2023

  58. [58]

    Segllm: Multi-round reasoning segmentation with large language models

    XuDong Wang, Shaolun Zhang, Shufan Li, Kehan Li, Konstantinos Kallidromitis, Yusuke Kato, Kazuki Kozuka, and Trevor Darrell. Segllm: Multi-round reasoning segmentation with large language models. In The Thirteenth International Conference on Learning Representations, 2025 a

    2025

  59. [59]

    Gpt-image-edit-1.5 m: A million-scale, gpt-generated image dataset

    Yuhan Wang, Siwei Yang, Bingchen Zhao, Letian Zhang, Qing Liu, Yuyin Zhou, and Cihang Xie. Gpt-image-edit-1.5 m: A million-scale, gpt-generated image dataset. arXiv preprint arXiv:2507.21033, 2025 b

    Pith/arXiv arXiv 2025

  60. [60]

    Fast-dllm: Training-free acceleration of diffusion llm by enabling kv cache and parallel decoding

    Chengyue Wu, Hao Zhang, Shuchen Xue, Zhijian Liu, Shizhe Diao, Ligeng Zhu, Ping Luo, Song Han, and Enze Xie. Fast-dllm: Training-free acceleration of diffusion llm by enabling kv cache and parallel decoding. arXiv preprint arXiv:2505.22618, 2025 a

    Pith/arXiv arXiv 2025

  61. [61]

    Omnigen2: Exploration to advanced multimodal generation

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation. arXiv preprint arXiv:2506.18871, 2025 b

    Pith/arXiv arXiv 2025

  62. [62]

    Lotlip: Improving language-image pre-training for long text understanding

    Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zha Zheng-Jun. Lotlip: Improving language-image pre-training for long text understanding. In arXiv, 2024

    2024

  63. [63]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023

    Pith/arXiv arXiv 2023

  64. [64]

    Omnigen: Unified image generation

    Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp.\ 13294--13304, 2025

    2025

  65. [65]

    Sana: Efficient high-resolution text-to-image synthesis with linear diffusion transformers

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution text-to-image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025 a

    2025

  66. [66]

    Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer, 2025 b

    Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, et al. Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer, 2025 b . URL https://arxiv.org/abs/2501.18427

    Pith/arXiv arXiv 2025

  67. [67]

    Show-o: One single transformer to unify multimodal understanding and generation

    Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024

    Pith/arXiv arXiv 2024

  68. [68]

    Multimodal large diffusion language models

    Ling Yang, Ye Tian, Bowen Li, Xinchen Zhang, Ke Shen, Yunhai Tong, and Mengdi Wang. Multimodal large diffusion language models. arXiv preprint arXiv:2505.15809, 2025

    Pith/arXiv arXiv 2025

  69. [69]

    Dream 7b, 2025 a

    Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. Dream 7b, 2025 a . URL https://hkunlp.github.io/blog/2025/dream

    2025

  70. [70]

    Imgedit: A unified image editing dataset and benchmark

    Yang Ye, Xianyi He, Zongjian Li, Bin Lin, Shenghai Yuan, Zhiyuan Yan, Bohan Hou, and Li Yuan. Imgedit: A unified image editing dataset and benchmark. arXiv preprint arXiv:2505.20275, 2025 b

    Pith/arXiv arXiv 2025

  71. [71]

    Llada-v: Large language diffusion models with visual instruction tuning

    Zebin You, Shen Nie, Xiaolu Zhang, Jun Hu, Jun Zhou, Zhiwu Lu, Ji-Rong Wen, and Chongxuan Li. Llada-v: Large language diffusion models with visual instruction tuning. arXiv preprint arXiv:2505.16933, 2025

    Pith/arXiv arXiv 2025

  72. [72]

    Anyedit: Mastering unified high-quality image editing for any idea

    Qifan Yu, Wei Chow, Zhongqi Yue, Kaihang Pan, Yang Wu, Xiaoyang Wan, Juncheng Li, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. Anyedit: Mastering unified high-quality image editing for any idea. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp.\ 26125--26135, 2025 a

    2025

  73. [73]

    Dimple: Discrete diffusion multimodal large language model with parallel decoding

    Runpeng Yu, Xinyin Ma, and Xinchao Wang. Dimple: Discrete diffusion multimodal large language model with parallel decoding. arXiv preprint arXiv:2505.16990, 2025 b

    Pith/arXiv arXiv 2025

  74. [74]

    Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for exp...

    2024

  75. [75]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 11975--11986, 2023

    2023

  76. [76]

    Magicbrush: A manually annotated dataset for instruction-guided image editing

    Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36: 0 31428--31449, 2023

    2023

  77. [77]

    Ultraedit: Instruction-based fine-grained image editing at scale

    Haozhe Zhao, Xiaojian Shawn Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. Advances in Neural Information Processing Systems, 37: 0 3058--3093, 2024

    2024

  78. [78]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025

    Pith/arXiv arXiv 2025

  79. [79]

    From reflection to perfection: Scaling inference-time optimization for text-to-image diffusion models via reflection tuning

    Le Zhuo, Liangbing Zhao, Sayak Paul, Yue Liao, Renrui Zhang, Yi Xin, Peng Gao, Mohamed Elhoseiny, and Hongsheng Li. From reflection to perfection: Scaling inference-time optimization for text-to-image diffusion models via reflection tuning. arXiv preprint arXiv:2504.16080, 2025

    Pith/arXiv arXiv 2025

  80. [80]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

Showing first 80 references.