PaperlayerView exact revision on arXiv

Source-linked AI summary

SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer

Haoyi Zhu, Haozhe Liu, Yuyang Zhao, Tian Ye, Junsong Chen, Jincheng Yu, Tong He, Song Han, Enze Xie

arXiv:2605.15178v1cs.CV

TL;DR

Minute-scale, camera-controlled 720p world modeling remains costly in data, training, and inference. SANA-WM natively trains an efficient open-source model and achieves stronger action following than open baselines, comparable visual quality, and up to 36× higher generation throughput.

  • Problem

    Existing minute-scale world models require costly models, data, training, and inference, while short-video distillation provides limited supervision for long-horizon persistence and trajectory following.

  • Method

    SANA-WM combines hybrid GDN–softmax long-context modeling, dual-branch camera control, progressive native minute-scale training, and long-video refinement.

  • Results

    SANA-WM achieves stronger action following than prior open-source baselines with comparable visual quality and up to 36× higher generation throughput.

  • Takeaways & Limitations

    SANA-WM supports practical minute-scale 720p world-model workflows using single-GPU rollout and refinement of promising trajectories.

  • Takeaways & Limitations

    SANA-WM remains scale-limited, lacks explicit 3D scene memory, and can drift in dynamic scenes, rare viewpoints, or longer rollouts.

Abstract

from arXiv · show

We introduce SANA-WM, an efficient 2.6B-parameter open-source world model natively trained for one-minute generation, synthesizing high-fidelity, 720p, minute-scale videos with precise camera control. SANA-WM achieves visual quality comparable to large-scale industrial baselines such as LingBot-World and HY-WorldPlay, while significantly improving efficiency. Four core designs drive our architecture: (1) Hybrid Linear Attention combines frame-wise Gated DeltaNet (GDN) with softmax attention for memory-efficient long-context modeling. (2) Dual-Branch Camera Control ensures precise 6-DoF trajectory adherence. (3) Two-Stage Generation Pipeline applies a long-video refiner to stage-1 outputs, improving quality and consistency across sequences. (4) Robust Annotation Pipeline extracts accurate metric-scale 6-DoF camera poses from public videos to yield high-quality, spatiotemporally consistent action labels. Driven by these designs, SANA-WMdemonstrates remarkable efficiency across data, training compute, and inference hardware: it uses only $\sim$213K public video clips with metric-scale pose supervision, completes training in 15 days on 64 H100s, and generates each 60s clip on a single GPU; its distilled variant can be deployed on a single RTX 5090 with NVFP4 quantization to denoise a 60s 720p clip in 34s. On our one-minute world-model benchmark, SANA-WM demonstrates stronger action-following accuracy than prior open-source baselines and achieves comparable visual quality at $36\times$ higher throughput for scalable world modeling.

1. Introduction

SANA-WM is a 2.6B-parameter open-source world model natively trained for one-minute, 720p video generation controlled by 6-DoF camera trajectories. It combines efficient long-context modeling, dual-branch camera control, long-video refinement, and pose-based data annotation to improve accessibility and action-following performance.

  • Contributions: SANA-WM uses 2.6B parameters, ∼213K public clips with metric-scale pose supervision, and 15 days of training on 64 H100 GPUs for native one-minute generation.It provides bidirectional, chunk-causal, and distilled autoregressive single-GPU inference variants.
  • Efficient Native One-Minute Backbone: The hybrid Linear DiT combines frame-wise Gated DeltaNet blocks with periodic softmax attention, keeping minute-scale context affordable while supporting scene persistence and long-range recall.A high-compression LTX2 tokenizer [10] reduces the token burden for one-minute 720p generation.
  • Dual-Branch Camera Control: Dual-rate camera conditioning combines a latent-rate UCPE branch [12] for global trajectory structure with raw-frame Plücker mixing for fine camera motion.The design targets faithful continuous trajectory following despite aggressive video compression.
  • Two-Stage Visual Refinement: A dedicated second-stage refiner operates on full-minute SANA-WM outputs to correct structural artifacts and sharpen details across the sequence.The refiner is applied after stage-1 generation as a quality-improvement pass.
  • Robust Data Annotation and Evaluation Benchmark: A robust annotation pipeline recovers metric-scale camera poses from public videos using pose and geometry estimators [13] [14] [15], enabling training and evaluation for long-horizon camera-controlled modeling.After filtering, the pipeline yields ∼213K clips with precise metric-scale pose annotations and supports a dedicated one-minute benchmark.
  • Results: Up to 36× higher generation throughput accompanies higher action-following accuracy than prior open-source baselines at comparable visual quality, with one-minute generation on a single GPU.The distilled variant generates a one-minute video in 34s on a single RTX 5090 using NVFP4 quantization.

2. Related Work

Related work spans long-video generation, interactive action-conditioned world models, camera- and geometry-aware control, efficient long-context sequence modeling, and data, annotation, and evaluation for camera-controllable systems.

  • Long-video generation and interactive world models: Long-video generation uses diffusion or flow-transformer backbones over compressed spatiotemporal latents, while long-duration synthesis commonly relies on autoregressive or block-wise rollout, diffusion forcing, or streaming training.
  • Long-video generation and interactive world models: Interactive world models extend video generation to long-rollout action-conditioned simulation with keyboard, gamepad, camera, text, robot, or mixed controls.Related work also studies explicit memory, scene persistence, and geometry-aware state for revisits and long-horizon consistency, including BEV or occupancy-based driving simulators and camera-aware memories.
  • Camera control, geometry, and action spaces: Action-conditioned world models use diverse interfaces, including robot or embodied actions, keyboard or gamepad controls, and language, event, or mixed high-level commands.Camera-controlled generation is closely related to novel-view synthesis and geometric video generation, including camera-control modules added to pretrained video diffusion models.
  • Efficient sequence models for long visual horizons: Because softmax attention’s memory and compute grow with context length, long-horizon visual modeling increasingly explores linear, gated, state-space, convolutional, test-time-training, and delta-rule sequence models.Softmax attention remains effective and can be accelerated with kernels such as FlashAttention [68].
  • Data, annotation, and metrics: Camera-controllable world modeling depends on data with reliable geometry, diverse motion, and long-horizon scene coverage drawn from internet videos, spatial collections, 3D captures, embodied scenes, games, synthetic environments, and controlled benchmark pipelines.Prior work also applies filtering and enhancement procedures such as shot detection and video-quality assessment.

3. Method

SANA-WM addresses minute-scale, high-resolution world modeling under strict efficiency constraints through progressive training and four core architectural designs. Its method combines compressed latent modeling, hybrid recurrent-attention processing, dual-rate camera conditioning, and efficient chunk-causal deployment.

  • Progressive Training: SANA-WM progressively scales from short clips to minute-long videos through four stages: VAE adaptation, hybrid architecture adaptation, minute-scale camera conditioning, and chunk-causal distillation.The pipeline begins with LTX2-VAE adaptation, stabilizes the hybrid backbone on short clips, extends sequence length with metric 6-DoF control, and distills deployment to four denoising steps with constant per-chunk softmax cost.
  • Efficient VAE Adaptation: LTX2-VAE provides a representation 2.0× smaller than ST-DC-AE and 8.0× smaller than Wan2.1-VAE, improving training and inference efficiency.The baseline VAE is replaced and the patchify layer and final projection are re-initialized before 50k steps of full-model fine-tuning.
  • Chunk-Causal Deployment: Chunk-causal fine-tuning enables autoregressive rollout, while attention-sink tokens and local temporal windows keep softmax memory and per-chunk latency constant with rollout length.The reversed GDN scan resets at chunk boundaries to provide local future context without leakage, and self-forcing distillation reduces sampling to four denoising steps.
  • Hybrid GDN/Softmax Attention: Frame-wise GDN replaces token-wise recurrence by consuming all spatial tokens per latent frame while maintaining a D×D recurrent state, and alternating GDN with softmax attention improves long-video generation.Decay and per-token update gates control stale content, while key scaling stabilizes the transition under spatial-token aggregation.
  • Dual-Branch Camera Control: Dual-rate geometric conditioning combines latent-frame UCPE for global 6-DoF pose with raw-frame Plücker mixing to compensate motion within each VAE stride.The coarse camera branch uses ray-local pose transformations and shared frame-wise GDN gates, while the fine branch computes pixel-wise Plücker raymaps from raw-frame poses.

4. Data Pipeline

SANA-WM constructs a 213K-clip corpus by re-annotating seven open-source video sources with robust metric-scale camera poses. The pipeline combines improved depth and intrinsic optimization, synthetic trajectory augmentation, and camera-aware filtering and captioning.

  • Sources and pose annotation: Re-annotating seven open-source video sources yields a 213K-clip corpus with metric-scale camera poses.Figure 3 summarizes the construction flow, while Table 1 lists the resulting data.
  • Sources and pose annotation: Replacing VIPE’s unstable long-video depth backend with Pi3X and MoGe-2, plus per-frame intrinsic optimization, produces robust metric-scale poses for videos in the wild.Pi3X supplies long-sequence-consistent depth, while MoGe-2 provides accurate per-frame metric scale.
  • 3DGS augmentation: Static DL3DV captures are converted into one-minute videos by fitting FCGS reconstructions, rendering diverse camera paths with known intrinsics and extrinsics, and refining artifacts with DiFix3D.The procedure augments static 3D scene data with rendered camera trajectories.
  • Filtering and captioning: Filtering removes low-quality, inconsistent, or cut scenes using visual criteria and additional camera checks on field of view, focal consistency, pose smoothness, and scale variation.The basic visual filters follow SANA-Video [25].

5. Experiments

SANA-WM is evaluated on a 60-second, four-category benchmark and achieves strong camera-following accuracy, near-leading visual quality, and substantially improved inference efficiency. Ablations attribute these results to hybrid GDN attention, dual camera conditioning, and stable scaling strategies.

  • Benchmark results: SANA-WM achieves the strongest action following, with refined RotErr 4.50°/8.34° and CamMC 1.41/1.44, while reaching 80.62/81.89 VBench Overall close to LingBot-World’s 81.82/81.89.The comparison uses Simple-/Hard-Trajectory splits; the refined model improves over both 480p and 720p baselines.
  • Benchmark results: The refined pipeline reaches 74.7 GB and 22.0 videos/hour, remaining within an 80 GB H100 budget and running 3.7× faster than the fastest visible 480p baseline.SANA-WM without refinement uses 51.1 GB and reaches 24.1 videos/hour; minute-long 720p LingBot-World inference is unaffordable under the evaluation budget.
  • Benchmark results: Refinement yields 14.46/14.80 dB revisit PSNR and reduces ΔIQ from 3.79/3.09 to 1.17/0.31, ranking first on Hard and limiting long-horizon drift.SANA-WM ranks second on Simple and first on Hard among visible methods; HY-WorldPlay shows severe degradation with ΔIQ 23.59/25.88.
  • Progressive-training ablation: The hybrid 15-GDN/5-softmax backbone raises VBench-I2V Total to 0.853 while LTX2 reduces peak memory from 8.9 to 5.4 GB and latency by 3.4×.The progressive-training ablation uses identical data and inference settings at a 5 s horizon.
  • Training stability and efficiency scaling: D_S scaling is the only stable GDN variant, whereas L2 and no-scale baselines trigger NaNs at steps 16 and 1; recurrent variants also avoid the all-softmax model’s 60 s out-of-memory failure.Matched scaling experiments show compact recurrent/linear states while all-softmax memory grows with its KV cache.
  • Camera conditioning: Dual UCPE + Plücker mixing gives the lowest Pi3X errors with competitive FVD, outperforming input-only Plücker and single attention-level conditioning variants.The ablation is conducted on held-out OmniWorld validation with 5 s clips after 10k fine-tuning steps.

6. Conclusion

SANA-WM combines efficient modeling, camera conditioning, metric-pose annotation, and long-video refinement for practical minute-scale 720p generation with 6-DoF control. It remains limited by scale, scene memory, dynamic-scene drift, rare viewpoints, and longer rollouts, motivating broader scaling and more robust controls.

  • Conclusion: SANA-WM combines hybrid GDN–softmax modeling, dual UCPE + Plücker conditioning, metric-pose annotation, and a long-video refiner for practical minute-scale 720p generation with 6-DoF control.The system targets limited-compute training and single-GPU rollout, while experiments on revisit-heavy 60s trajectories show stronger camera following than open baselines.
  • Limitations, social impact, and future work: SANA-WM remains scale-limited, lacks explicit 3D scene memory, and can drift in dynamic scenes, rare viewpoints, or longer rollouts.These limitations constrain robustness beyond the evaluated setting.
  • Limitations, social impact, and future work: Its efficiency broadens access to simulation, embodied AI, and robotics research, while deployment should document provenance, model scope, and evaluation settings clearly.Future work includes scaling models and data, exploring robot-action or point-tracking controls, strengthening persistent scene memory, and developing robust methods.

A. Long-Video Refiner

The long-video refiner uses truncated-σflow matching to transform noisy Stage-1 latents toward high-fidelity targets while preserving identity through reference conditioning. Long-video adaptation is necessary because the original short-video refiner substantially degrades perceptual quality, late-window imaging quality, and camera-control accuracy on 60-second sequences.

  • Truncated-σflow matching: Truncated-σflow matching trains the refiner to predict the target velocity along interpolations between a Gaussian-noised Stage-1 latent and its high-fidelity target.Noise levels are sampled from a shifted-logit-normal distribution truncated to (0, σ_start], and training states remain on the segment connecting the source and target.
  • Reference conditioning: Reference conditioning prepends a clean target-latent slice as fixed key-value anchors, preserving identity and appearance consistency while excluding the reference tokens from the flow-matching loss.A block-wise attention mask provides the anchoring behavior.
  • Implementation details: The lightweight adaptation uses rank-384 LoRA on attention and feed-forward projections, trained on the multi-step LTX-2 base model before zero-shot merging into the distilled few-step model.This avoids unstable direct finetuning of the distilled refiner while preserving its distilled inference schedule.
  • Ablation: The adapted refiner outperforms the original LTX-2.3 short-video refiner on full 60-second latent sequences, which substantially reduces perceptual quality, late-window imaging quality, and camera-control accuracy.This ablation uses the same Stage-1 SANA-WM latents and applies each refiner directly to the full 60-second latent sequence.

B. Data Pipeline Details · B.1. Modified VIPE Pose Engine · B.2. 3DGS Augmentation Pipeline

The data pipeline adapts VIPE for metric-scale, internet-video pose annotation and augments static scenes with constrained, diverse 3DGS-rendered trajectories. It further improves robustness through dataset-specific supervision modes and multi-stage coverage and quality filtering.

  • B.1. Modified VIPE Pose Engine: The modified VIPE retains SLAM front-end components while replacing its depth and bundle-adjustment stages [13].Feature tracking and keyframe selection are preserved, while depth estimation and bundle adjustment are modified.
  • B.1. Modified VIPE Pose Engine: Pi3X [14] and MoGe-2 [15] replace Metric3D-Small, combining multi-frame 3D structure with metric-scale anchoring through temporally smoothed per-frame scale estimation.The scale factor uses inverse-depth weighting and exponential moving averaging with momentum 0.99.
  • B.1. Modified VIPE Pose Engine: Per-frame bundle adjustment independently optimizes focal lengths and principal points, supporting internet videos with non-square pixels and varying focal lengths.The intrinsics are represented as an (N, V, D) tensor across frames, views, and intrinsic dimensions.
  • B.1. Modified VIPE Pose Engine: Three annotation modes support internet video, GT-depth OmniWorld data, and GT-pose Sekai Game and DL3DV data with metric-scale recovery tailored to each source.GT-pose processing uses Umeyama Sim(3) alignment and 80th-percentile inlier filtering.
  • B.2. 3DGS Augmentation Pipeline: Static DL3DV scenes are augmented by rendering one-minute videos from fitted FCGS [94] Gaussian Splats while constraining cameras to the observed capture region.The pipeline computes scene and camera statistics including centroid, median radius, height range, and PCA directions.
  • B.2. 3DGS Augmentation Pipeline: Each scene receives 40 trajectories: 10 interpolate near training views, while 30 use diverse orbit, spiral, dolly, fly-through, crane, pendulum, random-walk, and compound motions.Trajectories are scaled by median camera distance, anchored near training cameras, oriented toward the Gaussian centroid when appropriate, and clamped to coverage bounds.
  • B.2. 3DGS Augmentation Pipeline: Coverage tests reject poorly observed views before rendering, while near-blank clips are removed afterward and surviving frames are refined with DiFix3D [95].The checks require sufficient splat projection in at least 70% of sampled frames, reject views with more than 65% empty tiles, and discard clips exceeding 30% near-blank frames.

B.3. Per-Dataset Quality Filter Thresholds · C. Implementation Details · C.1. Detailed Implementation and Training Efficiency

The implementation combines unified quality and geometry filtering with a hybrid GDN–softmax backbone and communication-efficient context-parallel training. Exact recurrent-state recovery, boundary handling, and fused Triton kernels support long-sequence training and improve efficiency.

  • B.3. Per-Dataset Quality Filter Thresholds: The unified data-selection pipeline scores saturation, VMAF motion, optical-flow magnitude, DOVER quality, and scene cuts before training.For long videos, optical flow samples frame pairs every 0.5s across the first 60s, while DOVER scores are averaged over…
  • B.3. Per-Dataset Quality Filter Thresholds: Camera geometry filtering requires horizontal and vertical fields of view between 25° and 120° and measures symmetric normalized focal mismatch.The fields of view are computed from frame resolution and camera intrinsics.
  • B.3. Per-Dataset Quality Filter Thresholds: Clips are rejected when the coefficient of variation of annotated per-frame scale factors exceeds 2.0.The coefficient is std(s_t)/(mean(s_t) + ϵ).
  • B.3. Per-Dataset Quality Filter Thresholds: Table 6 summarizes per-dataset acceptable quality-filter ranges, with “—” marking filters that are not applied.Ranges are reported as [min, max].
  • C.1. Detailed Implementation and Training Efficiency: The 20-block backbone uses head dimension D=112, interleaving 15 frame-wise GDN blocks with softmax blocks at layers {3, 7, 11, 15, 19}.Dual UCPE + Plücker camera conditioning is applied in every block; softmax anchors long-term spatial consistency while GDN supports efficient frame evolution.
  • C.1. Detailed Implementation and Training Efficiency: Context-parallel training shards the 961-frame latent sequence across P GPUs, with each rank computing compact GDN transition and input composites.The shard covers its assigned frame interval and avoids replicating the full sequence on every GPU.
  • C.1. Detailed Implementation and Training Efficiency: All-gathered composite summaries are composed as an exclusive prefix to recover each rank’s mathematically exact initial GDN state with minimal communication overhead.This communicates compact summaries rather than full activations.
  • C.1. Detailed Implementation and Training Efficiency: Fused OpenAI Triton kernels combine normalization, activation, scaling, positional preparation, and recurrent scanning, yielding about 1.5× to 2× efficiency gain.Halo exchange supplies convolutional boundary context and preserves equivalence to the unsharded computation, with causal zero-padding at global boundaries where applicable.

C.2. Training Hyperparameters … D.2. Benchmark Examples

SANA-WM’s training uses a progressive DiT pipeline requiring about 15 days on 64 H100 GPUs, while its benchmark evaluates first-frame-conditioned generation across diverse scenes and controlled camera trajectories. The benchmark includes scene-scale motion constraints, collision checks, reproducibility metadata, and representative examples spanning four scene categories.

  • C.2. Training Hyperparameters: 15 days on 64 H100 GPUs are required for the main DiT training stages, following roughly 3.5 days of LTX2 VAE adaptation on 64 H100 GPUs.During minute-scale training, precomputed VAE latents remove online VAE encoding cost.
  • C.2. Training Hyperparameters: All training stages use AdamW, BF16 mixed precision, and gradient clipping at 0.5.
  • D. Benchmark Details: The benchmark begins with 80 1280×720 first-frame images balanced across game-style, indoor, outdoor-city, and outdoor-nature scenes.Each scene receives a scene-static first-person prompt describing layout, materials, lighting, and autonomous world dynamics without camera-motion language.
  • D.1. Benchmark Construction: Two trajectory splits test camera control with 10 smooth navigation templates in Simple and 10 stress-test templates in Hard.Templates include waypoint-defined arcs, S-curves, backtracking, figure-eight returns, large yaw changes, vertical motion, extreme pitch, whip-pans, loops, and crane-like moves.
  • D.1. Benchmark Construction: Camera trajectories are adapted to each scene using Pi3X [14] intrinsics, metric depth, and point clouds, limiting travel to 60% of median depth and speed to 0.4 m/s.The benchmark uses these estimates for scene-scale motion and boundary checking; collisions trigger speed reduction and trajectory regeneration.
  • D.2. Benchmark Examples: Figure 9 presents representative first-frame conditioning images across four scene categories, illustrating diversity in geometry, lighting, visual style, and fine-grained scene detail.

D.3. Evaluation Protocol … E.2. Progressive-training ablation: VBench-I2V per-dimension breakdown

The evaluation protocol standardizes video sampling, camera-control inputs, visual-quality and camera metrics, revisit memory, and minute-long temporal degradation. Additional results show competitive loop-closure memory, reduced long-horizon drift from refinement, and per-dimension ablations isolating efficiency and consistency effects.

  • D.3. Evaluation Protocol: All methods use one generated video per scene under split-specific protocols, with timestamp-remapped frame sampling and shared trajectory-derived camera-control inputs.This accommodates different native frame rates while keeping discrete actions, pose tracks, camera-control tensors, and calibrated intrinsics consistent.
  • D.3. Evaluation Protocol: Visual quality is evaluated with custom-input VBench across eight valid dimensions and the aggregated VBench Overall score.The reported dimensions include subject and background consistency, temporal flickering, motion smoothness, aesthetic and imaging quality, dynamic degree, and overall consistency.
  • D.3. Evaluation Protocol: Camera accuracy uses Pi3X pose recovery aligned to benchmark trajectories with Umeyama Sim(3), reporting rotation and translation errors where lower is better.Ground-truth and recovered trajectories are relativized and sampled consistently with the evaluated video.
  • D.3. Evaluation Protocol: The benchmark includes Simple and Hard trajectory templates, with the Hard split adding vertical motion, loop closures, and pitch-heavy viewpoints over 60-second height profiles.These templates expose out-of-plane motion beyond ground-plane projections alone.
  • D.3. Evaluation Protocol: Revisit memory compares up to five closest near-identical-viewpoint frame pairs per scene using PSNR, SSIM, and LPIPS, averaged over valid pairs and scenes.Higher PSNR/SSIM and lower LPIPS indicate better scene memory under loop closure.
  • E.1. Revisit memory and temporal stability (full table): SANA-WM remains competitive on 720p loop-closure memory, while the second-stage refiner consistently reduces minute-long visual drift, especially on the Hard-Trajectory split.Revisit metrics assess scene identity at nearby returning viewpoints, whereas ΔIQ measures late-window imaging-quality degradation.
  • E.2. Progressive-training ablation: VBench-I2V per-dimension breakdown: The per-dimension VBench-I2V breakdown separates three I2V-specific axes from six general video-quality axes, showing that the LTX2 VAE swap preserves quality while improving efficiency and the hybrid backbone improves long-range consistency and I2V dimensions.The analysis expands rolled-up scores into camera motion, subject, background, consistency, smoothness, dynamic degree, aesthetic quality, and imaging quality.

E.3. Additional Qualitative Results … G. Existing Assets and Tool Terms

Additional qualitative results show that SANA-WM preserves layout, object identity, and coherent 3D cues during minute-long camera-controlled rollouts, while its efficiency broadens access to world modeling. The paper also emphasizes deployment risks, provenance and licensing transparency, and limitations from data and tool coverage.

  • E.4. 3D Reconstruction Visualization: Recovered structures and camera paths from Pi3X [14] indicate coherent 3D cues in SANA-WM videos generated from Nano Banana Pro first frames.Figure 12 uses three 60-second benchmark rollouts, extending evaluation beyond frame-wise image quality.
  • F. Broader Impact: SANA-WM reduces the training corpus, training compute, and inference hardware needed for one-minute 720p rollouts, lowering barriers for academic and smaller research groups.The intended applications include simulation, embodied AI, robotics research, and interactive content prototyping.
  • F. Broader Impact: Generated videos may be mistaken for real observations, and simulated rollouts may be over-interpreted as faithful predictions in safety-critical robotics, autonomous-driving, or physical-planning settings.The paper identifies provenance documentation as an important safeguard.
  • F. Broader Impact: Performance may not transfer uniformly across cultures, environments, rare viewpoints, or sensitive human-centered scenes because the model inherits biases and coverage limits from its data and tools.These limitations arise from public video sources, generated benchmark images, and filtering or captioning systems.
  • G. Existing Assets and Tool Terms: The authors document public license or terms status for existing assets and external tools used in data construction, annotation, evaluation, and refinement.They cite original projects, follow their release terms, and treat benchmark images and prompts as internal evaluation inputs rather than a separately released dataset.
  • E.3. Additional Qualitative Results: SANA-WM better preserves global layout and object identity across minute-long Hard-Trajectory rollouts, while baselines blur details, alter scene structure, or collapse toward weak motion.These examples complement aggregate pose, VBench, and revisit-memory results rather than introducing a separate evaluation protocol.
Loading 2605.15178v1…