ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?
1. Paper overview
In one sentence: ImageWAM learns robot actions from image-editing caches, avoiding decoded future videos while making its strongest efficiency and robustness claims dependent on the comparison setup. e-motivatione-pipelinee-inferencee-efficiencye-reale-plus
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors question whether manipulation needs dense future-video prediction. They argue that video synthesis spends computation on appearance and temporal details that may not determine the next action, while inaccurate imagined contacts can mislead control. Image editing instead supplies a prior for the instruction-conditioned difference between a current observation and a useful endpoint. e-motivatione-diagnostic |
| Core mechanism | Source description The framework turns editing-transformer caches into an action-conditioning interface, retaining endpoint-image supervision during policy learning while omitting image decoding at deployment. e-pipelinee-inference |
| A key reported result | LIBERO-Plus robustness: 83.1 Reported average success rate (%). Original LIBERO demonstration training; seven perturbation dimensions across four suites; FLUX.2 4B. FastWAM 51.5; OpenVLA-OFT 69.6; ImageWAM OmniGen2 71.8 and Ovis-U1 71.2. A large reported improvement under this protocol; backbone and pretraining differences prevent attributing the entire gap to editing alone. e-datae-pluse-architecture |
| Reading caution | Reader analysis Attention overlays and selected video-artifact examples support a qualitative explanation, not causal proof that editing pretraining produces the gains. Backbone comparisons change components, despite the claim that the action expert is unchanged. Appendix action-head sizes differ across variants. e-diagnostice-backbone-claime-architecture |
Core contributions
- Source description
The framework turns editing-transformer caches into an action-conditioning interface, retaining endpoint-image supervision during policy learning while omitting image decoding at deployment. e-pipelinee-inference
Figure 2. Endpoint prediction trains the visual branch; its internal cache conditions robot actions. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the diagram from the instruction and observation toward the blue editing block, then follow the KV arrow into the green action expert. The state and action-noise tokens enter that expert separately. Snowflakes mark the frozen language and VAE modules; flames identify the two adapted branches. The lower VAE-decoder path depicts the endpoint observation at t+H+1. Figure 2's caption describes synthesis, whereas Section 3.4 specifies deployment without image decoding. Thus the decoder path should be read as the visual prediction branch rather than a required inference stage. Appendix 5.1 further describes separate visual/action projections and one-way action access through joint attention. e-pipelinee-inferencee-architecturee-formulation
What it supports. The policy's interface is a set of transformer features, not the pixels of a generated goal image. Endpoint supervision can shape those features during training, while the action expert consumes the cache directly. This distinction explains how the method can retain a predictive learning objective and still omit decoded future observations during execution.
Where the evidence stops. The diagram does not show a searched action plan or a control-feedback loop. Its drawn decoder must not override the explicit inference procedure. Joint training also does not imply that the visual backbone and action expert share all parameters.
2. Motivation
2.1 The problem and the proposed response
The authors question whether manipulation needs dense future-video prediction. They argue that video synthesis spends computation on appearance and temporal details that may not determine the next action, while inaccurate imagined contacts can mislead control. Image editing instead supplies a prior for the instruction-conditioned difference between a current observation and a useful endpoint. e-motivatione-diagnostic
2.2 What this reading follows
A robot asked to put shoes in a box needs a useful account of what should change in the scene. ImageWAM borrows that account from an image-editing model: train it to predict an endpoint, then let an action expert read the internal features. The distinction between training and deployment matters. The illustrated pipeline contains an image decoder, but the deployed policy can use one editing forward pass without decoding that image. This edition follows the cache interface, examines the reported simulation and physical-robot results, and separates the broad video-rollout speedup from the smaller gain against an already efficient cache-based baseline. e-motivatione-pipelinee-inferencee-efficiencye-reale-plus
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Supports the recorded classification
Dual-system × IDM is supported as a functional interpretation: separate visual and action transformers communicate through visual KV context, and actions are inferred from current observations plus endpoint-oriented features. Joint attention and training do not establish one shared-parameter model. IDM here means latent-context-conditioned action decoding, not an explicit observed current/goal-image interface at deployment. e-formulatione-pipelinee-inferencee-architecture
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Use an endpoint as supervision, then read its features
The method begins with a language-conditioned policy over action chunks. Its conceptual visual intermediate is the observation after the action horizon, rather than every intervening frame. In the implemented training configurations, the paper pairs an observation with a future image 16 frames ahead and an action chunk of length 16. The editing branch learns that supervised visual transformation. The action expert receives layer-wise keys and values after the visual tokens interact with the instruction, so it can use internal predictive features without receiving a decoded target image. At inference, a single editing forward step supplies the cache. Figure 2 includes the decoder because the architecture supports image prediction; Section 3.4 explicitly removes the need to decode it during deployment. This is the central training–inference distinction. e-formulatione-pipelinee-inferencee-training
5.2 Understand why the action expert remains a separate system
The action expert learns a velocity field that turns noisy action chunks into demonstration-like actions. Equation (8) interpolates between the expert chunk and Gaussian noise, while Equation (9) trains a predictor of noise minus expert action. Its conditioning includes the editing cache, sampled at different editing timesteps during training. The image branch has its own prediction loss, and the paper sums the two objectives. Appendix 5.1 nevertheless describes separate visual and action transformers with their own QKV projections; joint attention is their communication mechanism. Reader interpretation: shared training does not erase this architectural separation, which supports the catalog's dual-system assessment. The IDM label is best understood functionally as action decoding from predictive visual context. Deployment does not require an explicit pair of current and generated goal images. e-objectivese-architecturee-pipelinee-inference
5.3 Separate predictive usefulness from a universal robustness claim
Three questions organize the experiments: whether policies succeed, whether they survive distribution shifts, and how much inference costs. Standard LIBERO is already near ceiling, making its small average differences less informative than the larger LIBERO-Plus differences. The physical-robot table supplies execution evidence, but conflicting trial counts limit uncertainty analysis. RoboTwin's randomized-training evaluation must also remain separate from Table 6's Clean2Hard transfer column, where ImageWAM reports only 18.3%. On efficiency, the video-rollout baseline explains the largest speedup; the one-step cache baseline gives a smaller difference. Reader interpretation: together these results show that editing caches can be useful action context, but they do not isolate why. Attention maps support a hypothesis, and optimized prefix-only latency remains incomplete evidence until paired with matched success measurements. e-pluse-main-tablese-reale-real-protocole-unifiede-efficiencye-optimizatione-diagnostic
5.4 Training and inference
During training
LIBERO provides 500 demonstrations per suite across four suites; training merges the suites. LIBERO-Plus uses these original demonstrations without augmented training data. RoboTwin uses 2,500 clean and 25,000 randomized trajectories. Real-world training combines four tasks with 100 demonstrations each. e-datae-real-protocol
Appendix settings use eight NVIDIA H20 GPUs, bf16, AdamW at learning rate 0.0001, and ZeRO-1, except ZeRO-2 for FLUX.2 9B. LIBERO concatenates two views to 224×448; RoboTwin arranges three views at 288×256. Both use a future observation 16 frames ahead and a 16-action chunk. e-training
During inference
Choose a fixed editing timestep τ* and obtain caches from one editing-branch forward pass. Denoise actions conditioned on that cache, without decoding an edited frame or rolling out video. Appendix efficiency variants additionally use prefix-only training/inference. Table 11 uses three action denoising steps. Cache-timestep selection and feedback/execution cadence are not fully specified. e-inferencee-optimization
5.5 Implementation flow
- Encode instruction and observation
Frozen language/vision understanding components provide context, and a frozen VAE encodes images. The trainable editing branch combines instruction context, observation latents and noisy prediction latents. Figure 2 also draws a decoder and endpoint image; these depict the visual prediction branch, not a mandatory deployment stage. e-pipelinee-inference
- Expose editing features to a separate action transformer
At a sampled editing timestep, collect layer-wise keys and values after language–visual interaction. Appendix 5.1 describes separate visual and action QKV projections within joint attention, with one-way action access to context. Action-head weights are copied/interpolated from the editing transformer, with projections added for action input and output. e-pipelinee-architecture
- Learn endpoint prediction and action denoising
Expert action chunks and future observation latents supply separate flow-matching targets. Understanding components stay frozen; the image diffusion branch and action expert are jointly updated. The intermediate is learned predictive context, not a decoded goal that an independent planner searches against. e-objectivese-pipeline
6. Experiments & results
ImageWAM adapts pretrained image-editing models into robot policies: future endpoint supervision shapes internal visual features, and a separate flow-matching action expert reads their key/value caches. Deployment uses one editing forward pass without decoding the endpoint image. Reported robustness and efficiency gains depend strongly on the backbone and comparison protocol.
6.1 Read the original evidence
Table 3. Robustness improves substantially in the reported protocol, with remaining weaknesses across perturbations. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with P.T., which records additional embodied policy pretraining rather than the absence of foundation-model pretraining. Then compare Avg across the three ImageWAM rows before inspecting individual perturbations. The evaluation reuses original LIBERO demonstrations and does not train on augmented LIBERO-Plus data. FLUX.2 4B has the strongest ImageWAM average, but its Robot column is weaker than Ovis-U1's. Background is another useful check: OpenVLA-OFT exceeds the FLUX.2 row there. These exceptions help distinguish broad average improvement from universal robustness. Keep the table's reported averages as reported; do not treat one favorable column as representative of all shifts. e-pluse-datae-architecture
What it supports. FLUX.2 4B reports 83.1% average success versus 51.5% for FastWAM and 69.6% for OpenVLA-OFT. Its Robot result is 50.3%, compared with Ovis-U1's 58.4%; its Background result is 85.5%, below OpenVLA-OFT's 93.3%. The aggregate advantage therefore coexists with substantial, specific vulnerabilities.
Where the evidence stops. This is an end-to-end comparison across differently pretrained backbones, not a controlled isolation of image-editing pretraining. The source supplies no seed-level uncertainty here, and Appendix 5.1 lists different action-head sizes across ImageWAM variants.
Table 4. Physical execution improves on four reported tasks, with an unresolved evaluation denominator. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Map T1 to stacking three bowls, T2 to folding a towel, T3 to opening a drawer, storing a marker and closing the drawer, and T4 to hanging a cup on a rack. These are execution success percentages on the Dobot XTrainer, rather than scores for generated images. Compare ImageWAM with FastWAM column by column: the largest absolute gain is on towel folding, while the drawer task changes little. Training pools roughly 100 demonstrations per task. Before interpreting the precision of these percentages, consult both protocol descriptions: the main text says 100 evaluation trials, while the appendix specifies 50 trials per task. e-reale-real-protocole-datae-cost
What it supports. The reported average rises from 79.0% for FastWAM to 84.5% for ImageWAM, a gain of 5.5 percentage points. Task gains are 6, 9, 1 and 6 points. The towel result is consistent with usefulness for deformable-object manipulation, but does not by itself establish that editing-specific features caused that gain.
Where the evidence stops. Main-text and appendix trial counts conflict, so this table cannot support reconstructed confidence intervals without clarification. The appendix's training-cost table also names only OmniGen2 for real-world training, whereas the main experiment setup names FLUX.2 4B.
Table 5. The reference baseline determines how large the efficiency gain appears. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the rightmost Interm. column before comparing latency: the first row uses a Video intermediate, while the other two use Cache. The A6000 measurements compare ImageWAM's 263 ms and 9.72 TFLOPs with two substantially different alternatives. Relative to FastWAM-IDM, 1081 ms and 63.65 TFLOPs, the result supports the paper's approximate quarter-latency and one-sixth-compute headline. The closer inference-interface comparison is the middle row: 302 ms and 13.21 TFLOPs for one-step FastWAM. Keep those baselines separate when assessing the benefit of changing the generative prior rather than removing a costly rollout. e-efficiencye-inferencee-optimization
What it supports. ImageWAM's reported latency is about 4.11 times lower than the video-intermediate baseline, but only about 1.15 times lower than one-step FastWAM. Equivalently, the latter comparison yields a calculated 12.9% latency reduction. Both comparisons are useful; they answer different questions about rollout cost and the remaining difference between cache-based methods.
Where the evidence stops. Appendix Table 11 reaches 69 ms using prefix-only inference and additional optimizations, but supplies no paired success result there. Its ImageWAM 'Vid. Denoise' label is inconsistent with the method description and should not be interpreted as actual video generation.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO-Plus robustness Original LIBERO demonstration training; seven perturbation dimensions across four suites; FLUX.2 4B. | 83.1 Reported average success rate (%) | FastWAM 51.5; OpenVLA-OFT 69.6; ImageWAM OmniGen2 71.8 and Ovis-U1 71.2. A large reported improvement under this protocol; backbone and pretraining differences prevent attributing the entire gap to editing alone. e-datae-pluse-architecture |
| RoboTwin 2.0 multi-task manipulation Clean and randomized training/evaluation; main protocol specifies 100 trials per task. | 93.20 clean; 93.56 randomized; 93.38 average Success rate (%) | Table 1: FastWAM 91.83 average; LingBot-VA 92.20. Table 12 identifies these ImageWAM numbers with FLUX.2 4B. Randomized training is included; this is not clean-only transfer to randomized scenes. e-datae-main-tablese-per-task |
| Standard LIBERO manipulation Spatial, Object, Goal and Long suites. | 98.4 Average success rate (%) | Fast-WAM 97.6; LingBot-VA 98.5. Competitive near-ceiling performance, not the best reported average in Table 2. e-main-tables |
| Real-world dual-arm manipulation Dobot XTrainer; bowls, towel, drawer/marker and hanging-cup tasks. Trial denominator differs between main text and appendix. | 94, 84, 78, 82; average 84.5 Execution success rate (%) | FastWAM 79.0 average; π0.5 72.3; π0 55.8. Reported gain over FastWAM: 5.5 percentage points. These are physical execution results; uncertain trial counts prevent reconstructing reliable uncertainty estimates. e-reale-real-protocol |
| Inference efficiency Table 5, A6000 GPU. | 263 ms; 9.72 TFLOPs Latency (ms) and TFLOPs; lower is better | FastWAM-IDM: 1081 ms, 63.65 TFLOPs; FastWAM (1 Step): 302 ms, 13.21 TFLOPs. The roughly quarter-latency headline uses the video-intermediate baseline. Against one-step FastWAM the latency reduction is approximately 12.9%, calculated from the table. e-efficiency |
| Editing-backbone scaling LIBERO-Plus, FLUX.2 4B versus 9B. | 83.1 → 85.2 Reported average success rate (%) | Robot perturbation: 50.3 → 58.7; Camera: 80.8 → 79.8. Scaling helps unevenly. Prose gives 85.21%, while Table 7 rounds to 85.2%; the action head also grows from 642M to 952M parameters. e-scalinge-architecture |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 7. A larger configuration improves the average without improving every perturbation. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the two rows one perturbation at a time before reading Avg. Robot, Language, Background and Layout improve in the 9B configuration; Camera, Light and Noise decrease. The Robot column changes from 50.3 to 58.7, giving a concrete example of a favorable shift. The table rounds the final average to 85.2, while the neighboring prose gives 85.21%. Appendix 5.1 is essential to interpreting this ablation: scaling the editing model also changes the associated action expert from 642M to 952M parameters. The row difference consequently describes a larger system configuration, not an isolated change to one fixed component. e-scalinge-architecturee-cost
What it supports. The reported average rises by 2.1 percentage points when moving from FLUX.2 4B to 9B. Robot robustness improves by 8.4 points, but Camera drops from 80.8% to 79.8%. The result supports uneven benefits from scaling, rather than a monotonic improvement in all forms of distribution shift.
Where the evidence stops. The action expert scales too, preventing attribution solely to the editing backbone. Training costs also differ: Table 10 reports 18 hours for LIBERO 4B and 1.6 days for 9B. No uncertainty or matched-compute scaling control is supplied.
Figure 4. Attention overlays offer a qualitative account of the editing-cache hypothesis. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each task column begins with the observation, followed by FastWAM and ImageWAM overlays. Read the instruction underneath before deciding which visible regions are relevant: the left task moves a bowl to a stove, while the right involves a mug, plate and pudding placement. Compare the same camera panels vertically, using the uncolored observation to identify objects. In these examples ImageWAM shows concentrated hotspots around some manipulated objects and targets. The source interprets this as attention to task-relevant change. The figure includes no numerical color scale or layer/head aggregation specification, so the colors should be read qualitatively rather than as calibrated attention magnitudes. e-diagnostice-objectives
What it supports. The examples make the authors' representation argument visually plausible: some ImageWAM hotspots align with objects and destination regions named by the instruction. They help explain what a useful editing cache might contain. They do not measure whether those highlighted features are necessary for successful action prediction or whether the pattern persists across the full benchmark.
Where the evidence stops. Attention localization is observational evidence. Neither these selected overlays nor Figure 5's selected artifact examples isolate the cause of policy improvement. A controlled intervention on endpoint supervision or conditioning features would be needed to test that interpretation.
7. Analysis & limitations
7.1 What the evidence leaves open
Attention overlays and selected video-artifact examples support a qualitative explanation, not causal proof that editing pretraining produces the gains. Backbone comparisons change components, despite the claim that the action expert is unchanged. Appendix action-head sizes differ across variants. e-diagnostice-backbone-claime-architecture
Robustness is protocol-specific: Table 6 reports 84.4% on RoboTwin Clean Only but 18.3% on Clean2Hard, below keyframe BagelVLA's 20.9%. These columns cannot be merged with Table 1's randomized-training results. Aggregate success tables provide no seed-level uncertainty. e-unifiede-datae-main-tablese-pluse-real
Source inconsistencies remain unresolved: main real-world text says 100 trials, appendix says 50 per task; main training uses 30k steps while appendix gives epoch budgets. Equation (6) has a stray endpoint-index suffix, (7) switches to undefined K, and inclusive chunk notation is not reconciled with implemented length 16. e-real-protocole-datae-traininge-formulatione-objectives
7.2 Questions for discussion
- Would removing endpoint supervision while holding the backbone and action head fixed preserve robustness?
- Does prefix-only inference preserve success on contact-sensitive tasks under matched compilation settings?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs pretrained editing/LLM checkpoints, copied action-head initialization, attention masks, camera composites and endpoint/action pairing. Table 8 specifies optimizer betas (0.9, 0.95), weight decay 0.01, warmup-cosine scheduling and gradient clipping 1.0; exact software versions and the fixed cache timestep are not supplied. e-architecturee-traininge-inference
Resolve configuration provenance before comparing speed or success: Table 10 lists real-world training cost only for OmniGen2, while the main setup names FLUX.2 4B. Table 11's optimized 69 ms prefix-only result has no paired success measurement there. Proposed checks should measure matched success and latency and isolate endpoint supervision from capacity. e-datae-coste-optimizatione-objectives
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Does endpoint supervision cause the robustness gain?
Reader-proposed experiment, not performed: start all runs from the same FLUX.2 4B checkpoint and the same 642M action-head initialization. Train one with the paper's summed image/action losses, one with the image loss removed, and a diagnostic control with endpoint targets shuffled within tasks. Hold demonstrations, camera preprocessing, action horizon, optimizer, updates and inference timestep fixed; use repeated seeds and identical LIBERO/LIBERO-Plus evaluation instances. Report success by perturbation as well as average, and record endpoint-prediction quality separately. If full supervision consistently improves shifted-task success over both controls, that would support a role for correctly paired future supervision. Similar success despite degraded prediction would weaken that explanation. Resolve the printed image-loss indices before implementation. e-objectivese-inferencee-datae-pluse-traininge-architecture
Check 2: Does prefix-only speed preserve useful action context?
Reader-proposed experiment, not performed: compare the one-editing-step policy with the appendix's separately trained prefix-only variant on the same task split and initial states. Fix backbone, action chunk length, three action denoising steps and A6000 hardware; evaluate each with and without the same compilation/static-graph settings. Measure end-to-end latency after warmup, its distribution, and execution success, while reporting prefill and action-loop time separately. Include contact-sensitive tasks so a faster but less useful cache can fail visibly. The hypothesis that prediction-token removal is practically free would be contradicted by a reproducible success decrease despite improved latency. This comparison supplies the missing success–latency pairing for Table 11 rather than assuming its 69 ms row preserves Table 4's performance. e-inferencee-optimizatione-efficiencye-traininge-real
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version page, all five figures, all twelve tables, method equations, main evaluation protocols and appendix architecture/training/efficiency/task details on the listed PDF pages. Every final crop was separately viewed and checked for readable labels and faithful boundaries. Figure 2's decoder was reconciled with Section 3.4; image-loss index inconsistencies and Table 11's denoising label are preserved as source issues. Reference pages 10–15 were read as text, not independently sourced or used as evidence for the cited works. No separate supplement, code or project page was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 16, 17, 18, 19. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title, authors and abstract (p. 1)
- 1 Introduction (pp. 1–3)
- 2 Related Works, 2.1–2.2 (p. 3)
- 3 Method, 3.1–3.4 (pp. 3–5)
- 4 Experiments, 4.1–4.4 (pp. 6–9)
- 5 Conclusion (p. 9)
- References (pp. 10–15)
- Appendix 5.1.1–5.1.3: all three architectures (p. 16)
- Appendix 5.2 Training Details (pp. 16–17)
- 6 Efficiency Optimization (pp. 17–18)
- 7 Real-World Experiments Detail, 7.1 (p. 18)
- 8 RoboTwin Evaluation Results (pp. 18–19)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction's image limitation was addressed by inspecting the original PDF figures, tables, equations and final crops.
- Only the supplied PDF was reviewed; linked code and project pages were not inspected, and no experiments were reproduced.
- Identity notes: exact title and all ten authors match the catalog. The inspected artifact says arXiv:2606.19531v1, 17 Jun 2026, while its title block says June 19, 2026. No other edition or revision chain was supplied or compared.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title/author/affiliation block, date and arXiv margin
Exact title and ten-author identity match. Margin: 2606.19531v1, 17 Jun 2026; title block: June 19, 2026. Five affiliations are printed.
Go to primary source ↓e-motivationPDF pp. 1–3, Introduction and Figure 1
Authors motivate editing through instruction-to-change alignment, a single endpoint proxy and compact inference instead of dense video synthesis.
Go to primary source ↓e-formulationPDF pp. 3–4, Section 3.1, Equations (1–4)
Actions are conditioned on current observation and instruction. The conceptual visual intermediate becomes an endpoint at t+H+1. Equation (1) lists actions inclusively through t+H.
Go to primary source ↓e-pipelinePDF p. 4, Figure 2 and Section 3.2, Equation (5); p. 5, opening paragraph
Layer-wise editing KV features feed an action expert with robot state and action noise. Figure 2 marks LLM/VAE frozen and editing/action branches tunable; text freezes understanding modules.
Go to primary source ↓e-objectivesPDF p. 5, Section 3.3, Equations (6–9)
Image and action flow-matching losses are summed; training samples editing timesteps. The image target is defined at t+H+1, but (6) prints an extra s in the index and (7) uses t+K without defining K. Action interpolation and velocity targets are explicit.
Go to primary source ↓e-inferencePDF p. 5, Section 3.4, Equations (10–11)
One editing forward step at fixed τ* supplies caches for action denoising without edited-image decoding; numerical timestep and selection rule are unspecified.
Go to primary source ↓e-architecturePDF p. 16, Appendix Sections 5.1.1–5.1.3
Separate visual/action QKV and directional attention form a MoT architecture. Weight-copy initialization is described. Action DiT sizes: approximately 760M for OmniGen2, 642M/952M for FLUX.2 4B/9B, and 1.1B for Ovis-U1.
Go to primary source ↓e-dataPDF p. 6, Section 4.1, opening and benchmark protocol paragraphs
No extra embodied policy pretraining. Main setup names FLUX.2 4B. LIBERO: 500 demonstrations per suite; LIBERO-Plus adds no augmented training data. RoboTwin: 2,500 clean and 25,000 randomized trajectories, 30k training steps and 100 trials per task.
Go to primary source ↓e-main-tablesPDF p. 6, Tables 1–2, ImageWAM and comparison rows
RoboTwin: 93.20/93.56/93.38 clean/randomized/average, versus FastWAM average 91.83 and LingBot-VA 92.20. LIBERO averages: ImageWAM 98.4, Fast-WAM 97.6, LingBot-VA 98.5.
Go to primary source ↓e-plusPDF p. 7, Table 3, all rows and perturbation columns
Averages: FLUX.2 4B 83.1, OmniGen2 71.8, Ovis-U1 71.2, FastWAM 51.5, OpenVLA-OFT 69.6. FLUX.2 4B Robot 50.3, Camera 80.8, Background 85.5; Ovis-U1 Robot 58.4; OpenVLA-OFT Background 93.3.
Go to primary source ↓e-realPDF p. 8, Table 4, T1–T4 and Avg columns
ImageWAM execution success: 94/84/78/82 (84.5 average); FastWAM: 88/75/77/76 (79.0); π0.5 average 72.3; π0 average 55.8.
Go to primary source ↓e-real-protocolPDF p. 7, Section 4.1 Real-world Experiments; p. 18, Section 7.1
Four Dobot XTrainer tasks use about 100 demonstrations each. Main text: 30k training steps and 100 evaluation trials; appendix: 50 trials per task. Tasks: stacking three bowls, folding a towel, opening/storing/closing a drawer, hanging a cup.
Go to primary source ↓e-efficiencyPDF p. 8, Table 5 and Section 4.3 Latency and FLOPs
A6000: FastWAM-IDM 1081 ms/63.65 TFLOPs, FastWAM (1 Step) 302 ms/13.21 TFLOPs, ImageWAM 263 ms/9.72 TFLOPs. Intermediates: Video, Cache, Cache.
Go to primary source ↓e-diagnosticPDF p. 6, Figure 4; p. 8, Section 4.3; p. 9, Figure 5 and caption
Attention overlays show two instructions and localized ImageWAM hotspots; authors interpret these as task-relevant focus. Figure 5 marks selected video artifacts with red boxes and associates them with unreliable action conditioning. Evidence is qualitative.
Go to primary source ↓e-backbone-claimPDF p. 8, Section 4.4 Q1
Text says backbones are replaced with action expert and training data unchanged; cited averages correspond to Table 3 although the paragraph references Table 7.
Go to primary source ↓e-unifiedPDF p. 8, Table 6 and Section 4.4 Q2
ImageWAM: 84.4 RoboTwin Clean Only, 18.3 Clean2Hard. BagelVLA with keyframes: 75.3/20.9; without keyframes: 56.7/15.9. Authors argue for frozen understanding and adapted generation.
Go to primary source ↓e-scalingPDF p. 9, Section 4.4 Q3 and Table 7
FLUX.2 4B/9B averages: 83.1/85.2; prose: 85.21. Robot, Language, Background and Layout improve; Camera, Light and Noise decrease.
Go to primary source ↓e-trainingPDF pp. 16–17, Appendix 5.2, Tables 8–9 and adjoining text
Eight H20 GPUs, bf16, AdamW, learning rate 1e-4, betas 0.9/0.95, weight decay 0.01, warmup cosine, clipping 1.0; ZeRO-2 only for FLUX.2 9B. LIBERO: 224×448 two-view input, 10 epochs; RoboTwin: 288×256 three-view input, 5 epochs. Future horizon/action length: 16. Real-world: RoboTwin preprocessing, 10 epochs.
Go to primary source ↓e-costPDF p. 17, Table 10 and footnote
LIBERO FLUX.2 4B: 18 hours, batch/GPU 10; 9B: 1.6 days, 12. RoboTwin: five days, effective batch/GPU 48 with three-step accumulation. Real-world cost row lists only OmniGen2: 18 hours, batch/GPU 16.
Go to primary source ↓e-optimizationPDF pp. 17–18, Section 6 and Table 11
Prefix-only attention training and image-denoising-free inference are evaluated with compilation/static graphs; all rows use three action denoising steps. ImageWAM prefix-only: 198 ms, then 85/77/69 ms with listed optimizations. FastWAM prefix-only: 194 ms, 80 ms compiled. Original ImageWAM row is labeled '1× Vid. Denoise' despite its image-editing method. No success column.
Go to primary source ↓e-per-taskPDF p. 19, Table 12, headers and Average row
FLUX.2 4B reproduces 93.20/93.56. OmniGen2, explicitly labeled 50 trials, reports 92.48/92.80. Baselines include Fast-WAM-IDM and Fast-WAM without co-training; they must not be silently equated with Table 1's FastWAM.
Go to primary source ↓8.5 Primary sources
ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing? ↗
PDF · 9,356 extracted words
Source fingerprint
0bf354364a18a2e504b46dcefd895ea923a63f04bc2f4ac12d9e8ea57ea80e2a