PaperlayerView exact revision on arXiv

Source-linked AI summary

HYDRA-X: Native Unified Multimodal Models with Holistic Visual Tokenizers

Guozhen Zhang, Xuerui Qiu, Yutao Cui, Tianhui Song, Changlin Li, Junzhe Li, Tao Huang, Xiao Zhang, Yang Li, Jianbing Wu, Miles Yang, Zhao Zhong, Liefeng Bo, Limin Wang

arXiv:2606.13289v1cs.CVcs.AI

TL;DR

Unified multimodal models have given less attention to a single representation for images and videos; HYDRA-X brings both into one ViT-based tokenizer. It reports strong performance across image and video understanding and generation, and surpasses dedicated 3D-convolution video VAEs in reconstruction fidelity.

  • Problem

    Unified tokenization has focused on static images, leaving less explored how one representation can handle both images and videos in UMMs.

    Source: §1 · p. 1

  • Method

    HYDRA-X uses one ViT-based tokenizer for images and videos, with a decompressor for joint image-video teacher supervision and latent-level source-target interaction for editing.

    Source: §1 · pp. 1–2, §4.2 · p. 5

  • Results

    HYDRA-X reports strong performance across image and video understanding and generation, and its design choices surpass dedicated 3D-convolution video VAEs in reconstruction fidelity.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The tokenizer is framed as a holistic image-and-video interface unifying five tasks under one shared backbone.

    Source: §7 · p. 10

    Resource constraints left long-video generation and video editing unexplored.

    Source: E Limitations · p. 18

Abstract

from arXiv · show

Holistic visual tokenizers are fundamental to unified multimodal models (UMMs) as they map diverse visual inputs into a unified representation space. In this paper, we present HYDRA-X, the first UMM that unifies image and video tokenization within a single Vision Transformer (ViT). Our design is driven by two core challenges: efficiently injecting spatiotemporal reconstruction capability into a native ViT, and embedding image- and video-level semantic awareness into the latent space. To address the first, comprehensive ablations reveal two key findings: (1) frame-level causal temporal attention suffices for visual reconstruction, whereas full spatiotemporal attention degrades it; and (2) hierarchical temporal compression substantially outperforms single-step alternatives. To tackle the second, we propose a lightweight decompressor that upsamples temporally compressed features under joint image-video teacher supervision, thereby enforcing complementary semantic structures within the compact latent space. Building on this holistic tokenizer, we further propose a principled improvement of the editing pipeline: source-target interaction should occur at the latent level inside the tokenizer rather than at the semantic level inside the LLM, substantially improving editing consistency and accelerating convergence. Instantiated at the 7B dense model, HYDRA-X achieves strong performance across image and video understanding and generation tasks, paving the way for future unified-tokenizer UMMs.

1 Introduction

HYDRA-X introduces HYDRA-XTOK, a single ViT tokenizer for images and videos, designed to preserve reconstruction fidelity and shared semantic structure. It also moves source–target editing interaction into the tokenizer’s latent space.

1 Introduction

  • HYDRA-XTOK unifies image and video encoding in one ViT, addressing gaps in frame-wise tokenizers and cascaded designs that lack temporal interaction or semantic constraints.Frame-wise tokenizers provide no cross-frame dynamics; cascaded designs use an isolated VAE that may discard information important for understanding.Source: §1 · p. 1
  • Frame-level causal temporal attention to the immediately preceding frame outperforms global temporal attention, while hierarchical patchify outperforms single-step patchify for reconstruction.Together, these choices enable reconstruction fidelity beyond dedicated 3D-convolution video VAEs such as Wan2.2-VAE (Wan et al., 2025).Source: §1 · pp. 1–2
  • HYDRA-XTOK jointly tokenizes editing sources and targets with cross-frame interaction, fusing structural details into the target before it reaches the LLM.This latent-level interaction improves editing consistency and accelerates convergence compared with semantic-level interaction in HYDRA and cascaded designs.Source: §1 · p. 2
  • At 7B scale, HYDRA-X achieves strong performance on image and video understanding and generation, while unifying five UMM tasks through a shared encoder.The five tasks are image and video generation, image and video understanding, and image editing.Source: §1 · p. 2

2 Related Work

Related work traces HYDRA-X’s design to unified visual tokenizers and native UMMs, while distinguishing its single-ViT image–video representation and tokenizer-stage source–target interaction for editing.

2.1 Visual Tokenizers for Unified Multimodal Models

  • Unlike AToken (Lu et al., 2025), which emits task-specific outputs, HYDRA-X unifies image and video in one ViT tokenizer with temporal causality, hierarchical patchify, and semantic decompression.Earlier image tokenizers combine reconstruction and semantics, while video VAEs often lack semantic structure; HYDRA-XTOK inherits HYDRA’s Generation–Semantic Bottleneck approach.Source: §2.1 · pp. 2–3

2.2 Native Unified Multimodal Models

  • Native UMMs range from quantized-token and decoupled pathways to unified encoders; HYDRA-X extends unified-encoder systems from images to video and editing interaction.Composite systems preserve specialized models but have shallow task synergy; quantized designs face VQ fidelity limits, while decoupled designs duplicate visual pathways and require downstream realignment.Source: §2.2 · p. 3

2.3 Image Editing in Unified Multimodal Models

  • Unlike adapter-based or independently encoded editing pipelines, HYDRA-XTOK places source and target in one temporal window so source structure enters the target latent before the LLM.This reuses the video-trained temporal pathway and avoids an extra cross-image attention module; Show-o2 (Xie et al., 2025a) and TUNA (Liu et al., 2025b) concatenate post-encoder features instead.Source: §2.3 · p. 3

3 Preliminaries: Representation-Harmonized Tokenization

HYDRA-X follows the image-only UMM framework HYDRA (Qiu et al., 2026), using a single ViT with generation and semantic pathways. It extends that design to video through explicit temporal causality, hierarchical patchify, and a Decompressor.

3 Preliminaries: Representation-Harmonized Tokenization

  • HYDRA (Qiu et al., 2026) uses a single ViT split into Gen-ViT and Sem-ViT, linked by a Generation–Semantic Bottleneck for generation and semantic perception.The Gen-ViT maps image features through the bottleneck into compact latent z; the Sem-ViT un-projects z into semantic feature s, aligned to a pretrained teacher by distillation.Source: §3 · p. 3
  • The downstream LLM uses Sem-ViT output s for understanding and generation, while the pixel decoder reconstructs images from z only during tokenizer training.Source: §3 · p. 3
  • HYDRA-X extends the image-only design to videos through explicit temporal causality, hierarchical patchify, and a Decompressor introduced in Section 4.Source: §3 · p. 4

4 HYDRA-XTOK: Holistic Visual Tokenization in a Single ViT

HYDRA-XTOK is HYDRA-X’s visual interface, with latents designed to support compact generation, faithful reconstruction, and semantic understanding. Its ablations favor 2-frame tubelet attention and hierarchical temporal compression, while a Decompressor enables video-level semantic supervision.

4 HYDRA-XTOK: Holistic Visual Tokenization in a Single ViT

  • HYDRA-XTOK is HYDRA-X’s visual interface, trained with reconstruction and semantic-distillation losses to make pre-LLM latents compact, reconstructable, and semantically useful.The reconstruction term preserves pixel fidelity, while distillation aligns Sem-ViT features with semantic features.Source: §4 · p. 4

4.1 Spatiotemporal Reconstruction in a ViT

  • Two-frame tubelet attention gives the best reconstruction, while two-stage 2× temporal patchification consistently outperforms single-step 4× patchification at the same compression ratio.Wider attention receptive fields degrade reconstruction, and progressive temporal compression preserves reconstruction better than single-step compression.Source: §4.1 · p. 5

4.2 Spatiotemporal Semantic Distillation via the Decompressor

  • Because compressed video latents do not match frame-rate teacher features, a lightweight training-only Decompressor restores native temporal length for video-teacher distillation and is discarded afterward.The compact Sem-ViT output remains the representation used by the LLM; uncompressed image latents are distilled against an image teacher, while the Decompressor output supports video-teacher alignment.Source: §4.2 · p. 5, Figure 3 · p. 5
  • Decompressor-enabled video-teacher distillation yields the strongest video understanding while preserving image performance and the best image generation and editing scores; removing semantic distillation collapses understanding, and bidirectional attention degrades every metric.Source: §4.2 · pp. 5–6

5 HYDRA-X: Advancing Unified Multimodal Models with Holistic Tokenizers

HYDRA-X uses one holistic tokenizer and shared LLM backbone for five image and video tasks. For editing, it couples source and target latents inside the tokenizer, improving editing consistency.

5.1 Overall Architecture

  • The shared LLM processes interleaved text and HYDRA-XTOK visual tokens for image generation, understanding, video generation, understanding, and editing, using language and vision heads.The same Gen-ViT serves all five tasks; training is end-to-end with next-token prediction and flow-matching losses, each weighted 1 by default.Source: §5.1 · p. 6

5.3 Tokenizer-Stage Source-Target Interaction

  • Compared with HYDRA-X-Indep, HYDRA-X-STI raises Recon-PSNR on ImgEdit by nearly 7 dB, lifts ImgEdit by 0.4, and improves GenEval by 1.46.STI also yields consistent gains on most non-editing benchmarks; the Recon-PSNR increase supports the claim that editing consistency failures stem from latent-level isolation inside the tokenizer.Source: Table 3 · p. 7, §5.3 · p. 7

5.2 Independent Encoding Bypasses the Latent

  • Conventional pipelines independently tokenize editing sources and targets, leaving their latents disconnected and the LLM to discover alignment, which fails on detail-faithful edits.Independent encoding remains sufficient for high-level semantic edits.Source: §5.2 · p. 7

5.3 Tokenizer-Stage Source-Target Interaction

  • STI routes editing pairs through HYDRA-XTOK as a length-2 clip, letting Sem-ViT apply video-style tubelet causal attention to source and target latents.Gen-ViT encodes the images independently, and its cross-frame tubelet attention is disabled for editing; this preserves reconstruction while semantic-stage interaction supports alignment, without added parameters or a separate cross-image module.Source: §5.3 · p. 7

6 Main Results

At 7B, HYDRA-X leads unified baselines on visual generation and editing benchmarks, matches or exceeds 7B native UMM baselines on most reported image-understanding metrics, and improves over comparable unified video baselines. Its reported implementation uses a shared ViT tokenizer with a lightweight temporal decompressor and joint image-video teacher supervision.

6 Main Results

  • The reported HYDRA-X uses Qwen2.5-7B-Instruct as its LLM backbone, while a matched 1.5B variant is used for methodological ablations.Its symmetric ViT encoder-decoder uses 3D RoPE, a bottleneck dimension of C = 64, and a lightweight 4× temporal decompressor built from two upsample-and-transformer stages; SigLIP and InternVideo-Next-L serve as image and video teachers.Source: §6 · p. 7

6.1 Multimodal Understanding

  • HYDRA-X matches or exceeds 7B native UMM baselines on most reported image-understanding metrics, including OCR- and chart-heavy tasks.The benchmarks include AI2D, MME, MMMU, OCRBench, MMBench, RealWorldQA, ChartQA, DocVQA, and InfoVQA.Source: §6.1 · p. 8, Table 4 · p. 8, §6 · p. 8
  • HYDRA-X improves over reported 1.5B and 7B unified baselines where comparable video-understanding numbers are available, while remaining below the strongest dedicated or proprietary video LMMs on several metrics.It uses one ViT tokenizer across understanding, generation, and editing; the reported results are consistent with dual-teacher distillation providing image- and video-level semantics in the compressed latent.Source: §6.1 · p. 8, Table 5 · p. 8

6.2 Visual Generation

  • HYDRA-X is strongest among 7B unified baselines on every GenEval and WISE column, and leads unified VBench entries by +1.87 QS, +3.26 SS, and +2.15 Total.On VBench, these are improvements over the closest unified competitor, Show-o2-1.5B. HYDRA-X also leads in semantic-heavy dimensions including Object Class, Human Action, and Scene.Source: §6.2 · pp. 8–9, Table 6 · p. 9

6.3 Image Editing

  • Among 7B unified models, HYDRA-X scores 4.04 Ext. (+1.77), 4.38 Rm. (+1.14), and 4.34 ImgEdit Over. (+0.90), and leads GEdit G-SC/G-Over. (7.80/7.17).It also beats BAGEL-14B on every column and trails Qwen-Image-20B by only 0.20/0.39 on G-SC/G-Over.; the largest gains are on Ext. and Rm., which require identity-faithful source preservation.Source: §6.3 · p. 9, Table 7 · p. 9

7 Conclusion … A.2 Tokenizer Pre-training

HYDRA-X unifies image and video tokenization in one ViT, combining reconstruction and semantic supervision with latent-level source–target interaction for editing. Its tokenizer is trained through a reconstruction-and-distillation objective and three progressive pre-training stages.

7 Conclusion

  • HYDRA-X is the first native UMM framework to unify image and video tokenization in one ViT using frame-level causal tubelet attention, hierarchical temporal patchify, and dual image-video teacher supervision.These three design choices transform an image tokenizer into a video-and-image tokenizer.Source: §7 · p. 9
  • HYDRA-X processes source and target images as length-2 clips with its temporal-causal mechanism, restoring fine-grained latent-level coupling lost in prior independent-encoding pipelines.The approach moves source-target interaction into the tokenizer rather than treating editing as purely an LLM-side problem.Source: §7 · pp. 9–10
  • The shared tokenizer backbone serves five tasks and acts as a holistic image-and-video interface rather than only a static image encoder.Source: §7 · p. 10

A.1 Tokenizer Training Loss

  • HYDRA-XTOK initializes Gen-ViT and Sem-ViT from SigLIP 2 (Tschannen et al., 2025) and trains with reconstruction and semantic-distillation terms aligning Sem-ViT to image and video teachers.The tokenizer is designed to provide compact, reconstructive, and semantic visual tokens before they reach the LLM.Source: §A.1 · p. 16
  • The reconstruction objective combines L1 pixel recovery, LPIPS perceptual fidelity, GAN texture refinement, and KL regularization toward a standard normal latent prior.Source: §A.1 · p. 16

A.2 Tokenizer Pre-training

  • Pre-training proceeds in three stages, beginning with SigLIP-2-initialized ImageNet-1.2M training at 256 × 256, then mixed-resolution training on 256 × 256 videos and 256–2048-pixel images.Stage 1 uses 300k AdamW iterations at a peak learning rate of 2 × 10−4 with a hybrid SigLIP-2 / InternVideo distillation teacher.Source: §A.2 · p. 16
  • Stage 2 freezes the encoder and fine-tunes the 27-layer ViT decoder with GAN loss, while Stage 3 freezes Gen-ViT and the decoder and updates Sem-ViT after normalizing Gen-ViT features.The final-stage normalization reduces feature heterogeneity between the heads and establishes a semantic-aware latent space for faithful reconstruction and downstream UMM tasks.Source: §A.2 · p. 16

A.3 Native Unified Multimodal Models Pre-training · A.4 Ablation Study Training Details

HYDRA-X pre-training uses three progressive stages to unify multimodal representations, jointly train understanding and generation, and refine instruction-following quality. The ablation studies evaluate multimodal understanding, image generation, and image reconstruction under separate training setups.

A.3 Native Unified Multimodal Models Pre-training

  • HYDRA-X uses a three-stage progressive training strategy to cultivate a harmonized unified multimodal model.The paper says detailed configurations and computational cost are summarized in Table 8.Source: §A.3 · p. 16
  • Stage 1 freezes Qwen2.5-7B-Instruct and tunes vision components on 100M image–text pairs to align visual latents with language.The tuned components are the projector, time-step embedding, and flow head.Source: §A.3 · p. 16
  • Stage 2 unlocks all parameters and jointly trains on 30M understanding samples, 30M generative samples, approximately 2M editing samples, and 10M video samples.The generative samples are strategically filtered from Stage 1.Source: §A.3 · p. 16
  • Stage 3 uses 6M MMU instruction samples from LLaVA-OneVision (Li et al., 2024a) and Pixmo (Deitke et al., 2025), plus 1.2M video samples from LLaVA-Video (Zhang et al., 2025).Generation training uses 10M aesthetic-filtered images and 6M high-fidelity synthetic images, while high-quality editing data continues to improve precise control.Source: §A.3 · p. 17

A.4 Ablation Study Training Details

  • The multimodal-understanding ablation trains on LLaVA-1.5 combined with LLaVA-Video SFT (Zhang et al., 2025).This is the stated dataset setup for evaluating multimodal understanding.Source: §A.4 · p. 17
  • The image-generation ablation uses Qwen2.5-1.5B (Yang et al., 2024a), trains first on 20M image-caption pairs, then fine-tunes with ImgEdit for editing.The staged setup adds image-editing capability after image-caption training.Source: §A.4 · p. 17
  • The image-reconstruction ablation trains on ImageNet-1k (1.2M) for 150k iterations and evaluates reconstruction quality with rFID.Source: §A.4 · p. 17

B Visual Reconstruction · C Evaluation Details of Multi-modal Understanding Benchmarks · D Tokenizer-Stage Source–Target Interaction: Visual Evidence

HYDRA-XTOK delivers strong image and video reconstruction, while HYDRA-X matches or exceeds 7B native UMM baselines on most reported understanding metrics. For editing, tokenizer-stage source–target interaction improves source fidelity by recovering nearly 7 dB of reconstruction PSNR and preserving identity-sensitive details.

B Visual Reconstruction

  • With ImageNet-1.2M-matched training, HYDRA-XTOK† outperforms RAE† and VAVAE† on every ImageNet metric, including +5.26 dB PSNR over VAVAE†.Despite twice the compression ratio of dedicated 8× image VAEs and less training data, it exceeds FLUX.1 on ImageNet PSNR (32.96 vs. 32.86) and rFID (0.154 vs. 0.176).Source: B Visual Reconstruction · pp. 17–18
  • HYDRA-XTOK is the strongest unified tokenizer on every video metric, improving over AToken-So/C by +1.59 dB DAVIS PSNR and +2.22 dB UCF PSNR while more than halving rFVD.It also outperforms the dedicated 16× video VAE Wan 2.2 by +0.55 dB DAVIS PSNR and +0.77 dB UCF PSNR, reducing rFVD by 21% and 25%, respectively. Table 9 reports reconstruction metrics under a unified protocol with inputs resized and centre-cropped to 256×256.Source: B Visual Reconstruction · pp. 17–18

C Evaluation Details of Multi-modal Understanding Benchmarks

  • HYDRA-X matches or exceeds 7B native UMM baselines on most reported metrics across nine benchmarks spanning general understanding, expert knowledge, document/chart comprehension, and fine-grained visual perception.OCR- and chart-heavy tasks require both fine visual details and semantic structure, which HYDRA-XTOK’s compact latent is designed to preserve.Source: C Evaluation Details of Multi-modal… · p. 18

D Tokenizer-Stage Source–Target Interaction: Visual Evidence

  • Tokenizer-stage source–target interaction recovers nearly 7 dB of source-reconstruction PSNR without changing the rest of the architecture or parameter count.The comparison is between independent source/target encoding and routing both through a shared Sem-ViT with tubelet causal attention.Source: D Tokenizer-Stage Source–Target… · p. 18
  • HYDRA-X-Indep encodes source and target through separate Sem-ViT branches, whereas HYDRA-X-STI routes them jointly through one shared Sem-ViT as a length-2 clip.Both variants retain the same Gen-ViT, LLM, flow-matching head, parameter count, and training schedule.Source: D Tokenizer-Stage Source–Target… · p. 18
  • In the visual examples, HYDRA-X-STI preserves the still-life and car scenes, while HYDRA-X-Indep fragments the still life and reimagines the car, removing its occupants and on-screen text.The STI result preserves the people inside the car and the visible plate.Source: D Tokenizer-Stage Source–Target… · p. 18
  • Independent encoding loses identity-sensitive information in the latent before the LLM reads it, whereas STI fixes this bottleneck inside the tokenizer at zero parameter cost.The passage links this mechanism to consistent margins on identity-sensitive editing dimensions reported in Tables 7 and 12.Source: D Tokenizer-Stage Source–Target… · p. 18

E Limitations · F Broader Impacts · HYDRA-X-Indep HYDRA-X-STI Input

The authors identify training scale, untested long-video tasks, and backbone choice as limitations, acknowledge misuse risks from text-to-image generation, and report that tokenizer-stage interaction preserves identity-sensitive details lost by independent encoding.

E Limitations

  • The current training-data and parameter scale may limit HYDRA-X’s ability to capture the full complexity of high-dimensional video distributions.The authors identify both data scale and model parameter scale as bottlenecks.Source: E Limitations · p. 18
  • Resource constraints prevented exploration of long-video generation and video editing, which the authors describe as natural extensions of the holistic encoder.Source: E Limitations · p. 18
  • HYDRA-X was instantiated only with a 7B dense LLM for fair comparison; pairing its tokenizer with MoE or MoT backbones is proposed as a path to amplify cross-task gains.The cited examples are MoE (Li et al., 2024b) and MoT (Liang et al., 2024).Source: E Limitations · p. 18

F Broader Impacts

  • The text-to-image capabilities could enable misleading or fabricated visuals, including deepfakes, for disinformation or impersonation.Source: F Broader Impacts · p. 18

HYDRA-X-Indep HYDRA-X-STI Input

  • HYDRA-X-STI preserves object layout, characters, and on-screen text that HYDRA-X-Indep loses, despite both using the same LLM and parameter count.The variants differ in source–target encoding: STI jointly encodes through tubelet causal attention, while Indep encodes source and target independently.Source: Figure 5 · p. 19

F Broader Impacts

G Additional Main Results · H Qualitative Comparisons · H.1 Image Reconstruction at 512×512

The additional results expand benchmark summaries into per-category and per-dimension breakdowns, while the qualitative comparisons cover HYDRA-X’s supported tasks and examine image reconstruction at 512×512.

G Additional Main Results

  • GenEval’s per-category breakdown covers single-object, two-object, counting, color, position, and color-attribute prompts, locating models’ compositional strengths.Models with at least 14B parameters are excluded from the ranking; daggered methods use LLM rewriters.Source: G Additional Main Results · p. 20, Table 10 · p. 20
  • WISE’s breakdown probes world knowledge across culture, time, space, biology, physics, and chemistry, complementing GenEval’s geometric and compositional probes.Models with at least 14B parameters are excluded from the ranking.Source: G Additional Main Results · p. 20, Table 11 · p. 20
  • ImgEdit-Bench reports results across nine instruction-guided editing operations, from object addition and removal to background replacement, style transfer, and compositional edits.The dimensions also include altering, extracting, replacing, and action edits; models with at least 14B parameters are excluded from the ranking.Source: G Additional Main Results · p. 20, Table 12 · p. 21
  • VBench expands the main-paper QS/SS/Total summary across fourteen dimensions, probing visual quality, motion, semantic correctness, and compositional reasoning.The table covers video-generation results, including quality score, semantic score, subject and background consistency, motion smoothness, dynamic degree, and other dimensions.Source: G Additional Main Results · p. 20, Table 13 · p. 21

H Qualitative Comparisons

  • Qualitative comparisons span HYDRA-X’s five supported tasks, using representative unified multimodal and task-specialised baselines and organizing comparisons by task and resolution.The supplied passage does not enumerate the five tasks.Source: H Qualitative Comparisons · p. 22

H.1 Image Reconstruction at 512×512

  • At 512×512, reconstruction comparisons include HYDRA-X, dedicated image VAEs, unified-model tokenizers, and RAE, making texture, fine edges, and small text visually comparable.The compared baselines are FLUX, MingTok, AToken, and RAE.Source: §H.1 · p. 22, Figure 6 · p. 22

H.2 Image Reconstruction at 1280×768 … H.6 Image Editing

These sections present qualitative comparisons of high-resolution image and video reconstruction, alongside image and video generation samples and instruction-guided editing comparisons. They describe the settings and systems shown without reporting quantitative results or a comparative winner.

H.2 Image Reconstruction at 1280×768

  • At 1280×768, HYDRA-X is qualitatively compared with Wan 2.2 (Wan et al., 2025), AToken (Lu et al., 2025), and FLUX (Labs et al., 2025) to stress-test reconstruction beyond training resolution.The setting probes how tokenizers handle dense details such as text, foliage, and small structures under a stretched spatial token budget.Source: Figure 7 · p. 23, §H.2 · p. 23

H.3 Video Reconstruction at 512×512

  • HYDRA-X’s temporally consecutive-frame reconstructions are compared with Wan 2.2 (Wan et al., 2025) and AToken (Lu et al., 2025) to assess motion-sensitive cues, including object boundaries and inter-frame consistency.Source: §H.3 · p. 24, Figure 8 · p. 24

H.4 Image Generation

  • HYDRA-X text-to-image samples span realistic photography, stylised illustration, and compositional and knowledge-driven scenes.The samples characterise prompt coverage and aesthetic quality.Source: §H.4 · p. 25, Figure 9 · p. 25

H.5 Video Generation

  • HYDRA-X text-to-video samples cover varied subjects, scenes, and motion patterns, illustrating temporally coherent synthesis under the same UMM backbone.The section presents qualitative samples rather than quantitative comparisons.Source: §H.5 · p. 26, Figure 10 · p. 26

H.6 Image Editing

  • Instruction-guided edits are compared with unified multimodal models BAGEL and OmniGen2 and editing-specialised generators Qwen-Image-Edit and Step1X-Edit.The side-by-side comparison is intended to let readers gauge identity preservation, instruction adherence, and visual quality.Source: §H.6 · p. 27
  • Figure 11 presents the qualitative editing comparison of HYDRA-X with BAGEL (Deng et al., 2025), Qwen-Image-Edit (Wu et al., 2025a), Step1X-Edit (Liu et al., 2025a), and OmniGen2 (Wu et al., 2025c).The passage lists these systems as the visual comparison set; it does not establish a winner.Source: Figure 11 · p. 27

Related papers

Loading 2606.13289v1…