{"id":"4abe38ef-221d-4bc6-9338-9a349ac6ee98","arxiv_id":"2605.30263","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"minWM supplies an end-to-end pipeline that fine-tunes bidirectional T2V/TI2V models with camera control then distills them via Causal Forcing into few-step autoregressive generators for low-latency rollout.","lead":"The paper presents minWM, an open-source framework that turns existing bidirectional video diffusion models into camera-controllable, few-step autoregressive models for real-time interactive video generation. A smart generalist might read it to see a practical, modular recipe for building low-latency world models from current foundation models.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Causal Forcing / Causal Forcing++ pipeline may not preserve quality/stability when distilling bidirectional models to causal few-step AR generators","rationale":"The reader's weakest_assumption is precisely the load-bearing step; the abstract-only review correctly flags it, and the described ablations do not directly test rollout stability, so UNVERDICTED remains appropriate even after noting the full pipeline description.","tokens_in":1853,"tokens_out":328,"duration_ms":16330,"concrete_test":"On the released Wan2.1-T2V-1.3B checkpoint, run 30-frame autoregressive rollouts (camera-conditioned) and compute FVD / CLIP-T against both ground-truth videos and the original bidirectional model under identical conditioning; if distilled FVD rises >25% or temporal artifacts appear in >30% of sequences, the no-major-loss assumption fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim requires that fine-tuning a bidirectional T2V/TI2V model with camera control, followed by AR diffusion training + causal ODE/consistency distillation + asymmetric DMD, yields low-latency autoregressive rollouts without major quality loss or instability. This is the least secure link: bidirectional models depend on non-causal context, and the listed distillation steps are known to introduce compounding artifacts and temporal drift in video; the abstract only cites ablations on camera trajectories, controllability steps, and batch size, with no reported metrics on long-horizon rollout fidelity (FVD, temporal consistency, or perceptual quality vs. base model).","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents minWM, an open-source full-stack framework that converts existing bidirectional T2V/TI2V video diffusion models (e.g., Wan2.1-T2V-1.3B, HY1.5-TI2V-8B) into camera-controllable few-step autoregressive world models. The pipeline consists of camera-controlled fine-tuning of the bidirectional model, followed by AR diffusion training, causal ODE/consistency distillation, and asymmetric DMD to enable low-latency rollout; the work also supports adaptation of existing world models and releases code, checkpoints, and ablations on camera trajectories, controllability steps, and batch size.","tokens_in":1979,"tokens_out":492,"duration_ms":20775,"significance":"If the described pipeline reliably yields stable low-latency rollouts with limited quality degradation, the contribution would be significant for lowering barriers to interactive video world models. The modular architecture support, open-source release of runnable scripts and checkpoints, and practical ablations constitute concrete strengths that aid reproducibility.","major_comments":[{"comment":"Abstract and Evaluation section: The central claims of low-latency, high-quality autoregressive rollouts without major quality loss rest on unshown quantitative results; no FVD, temporal consistency, perceptual quality, or long-horizon rollout metrics versus the base bidirectional models are reported, leaving the success of the Causal Forcing / Causal Forcing++ pipeline unverified.","section":"Abstract and Evaluation section"},{"comment":"§3 (Causal Forcing pipeline description): The assumption that AR diffusion training plus causal ODE/consistency distillation and asymmetric DMD can convert non-causal bidirectional models into stable causal few-step generators is load-bearing for the main claim, yet the manuscript provides no empirical evidence on compounding artifacts, temporal drift, or stability over extended rollouts.","section":"§3 (Causal Forcing pipeline description)"}],"minor_comments":[{"comment":"Ensure the GitHub repository link includes complete documentation, all referenced training/inference scripts, and the exact checkpoints used for the reported ablations.","section":"Project Page and Code Release"}],"recommendation":"major_revision","confidential_remarks":"This is primarily a systems/engineering contribution centered on pipeline release rather than a novel algorithmic advance; assess fit with the journal's scope on methodological novelty."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed feedback emphasizing the need for quantitative validation. We agree that the manuscript as submitted does not contain the requested metrics or stability analyses, and we will revise accordingly to address these gaps.","responses":[{"response":"We agree that direct quantitative comparisons are necessary to substantiate the claims. The current manuscript prioritizes the description of the modular pipeline, open-source release, and ablations on camera trajectories and training hyperparameters, but omits FVD, temporal consistency, perceptual quality, and long-horizon metrics against the base models. In the revised version we will add these evaluations in the Evaluation section, including comparisons that verify the Causal Forcing pipeline.","revision_made":"yes","referee_comment":"[Abstract and Evaluation section] Abstract and Evaluation section: The central claims of low-latency, high-quality autoregressive rollouts without major quality loss rest on unshown quantitative results; no FVD, temporal consistency, perceptual quality, or long-horizon rollout metrics versus the base bidirectional models are reported, leaving the success of the Causal Forcing / Causal Forcing++ pipeline unverified."},{"response":"We concur that empirical evidence on rollout stability is required. Section 3 currently describes the pipeline components without accompanying experiments on compounding artifacts, temporal drift, or long-horizon behavior. We will incorporate such analyses and any observed limitations into the revised manuscript.","revision_made":"yes","referee_comment":"[§3 (Causal Forcing pipeline description)] §3 (Causal Forcing pipeline description): The assumption that AR diffusion training plus causal ODE/consistency distillation and asymmetric DMD can convert non-causal bidirectional models into stable causal few-step generators is load-bearing for the main claim, yet the manuscript provides no empirical evidence on compounding artifacts, temporal drift, or stability over extended rollouts."}],"tokens_in":1534,"tokens_out":400,"duration_ms":26643,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"This paper's main point is that minWM wires together camera-controllable fine-tuning on bidirectional T2V/TI2V models with AR diffusion training, causal ODE/consistency distillation, and asymmetric DMD to produce few-step autoregressive generators, then releases the full stack on Wan2.1 and HY1.5 backbones plus adaptation support for models like HY-WorldPlay.\n\nWhat the work does well is the modularity and the actual release. Runnable scripts, checkpoints, documentation, and inference code lower the barrier for anyone who wants to try building interactive video systems. The reported ablations on camera trajectory quality, controllability training steps, and minimal batch size are the sort of concrete engineering details that help others get started.\n\nThe soft spot is the thin evidence on whether the distilled models hold up. The abstract describes the pipeline and claims low-latency controllable rollouts but only shows ablations on training choices, with no reported numbers on long-horizon metrics such as FVD, temporal consistency, or quality drop versus the base model. That is exactly where these distillation steps often introduce drift, so the central claim rests on implementation success that is not yet quantified in the provided text.\n\nThis is for researchers in generative video and robotics simulation who need a working recipe rather than a new theoretical result. A reader who plans to implement or extend real-time world models will get direct value from the released assets. It deserves serious referee time because the contribution is the integrated, reproducible pipeline and the open implementation, even if the evaluation section needs more rollout data.","headline":"minWM is a practical open-source engineering release that assembles camera fine-tuning and Causal Forcing distillation into a modular pipeline for turning bidirectional video models into low-latency autoregressive ones, with code and checkpoints shipped.","tokens_in":2501,"tokens_out":402,"would_cite":false,"duration_ms":27367,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"minWM framework converts bidirectional video diffusion models into camera-controllable autoregressive world models for real-time interaction.","keywords":["video world models","autoregressive diffusion","camera control","few-step distillation","causal forcing","real-time video generation","bidirectional to autoregressive","open-source framework"],"falsifier":"A side-by-side comparison on the same prompts showing that videos from the distilled few-step autoregressive models exhibit clear instability such as flickering, drifting camera paths, or loss of visual coherence over dozens of frames compared with the original bidirectional model.","tokens_in":2754,"feed_emoji":"🎥","tokens_out":755,"duration_ms":25010,"temperature":0.7,"pith_summary":"The paper presents minWM as an open-source framework that supplies a complete pipeline to transform existing bidirectional text-to-video and text-image-to-video diffusion models into interactive world models. It begins with fine-tuning for camera control, then applies autoregressive diffusion training followed by causal distillation and asymmetric DMD steps to achieve few-step causal generation. A sympathetic reader would care because this directly tackles the practical barriers to turning high-quality but non-causal video models into controllable, low-latency simulators. The work shows the pipeline on two different open backbones and includes ablations on training parameters such as batch size and steps. If the conversion holds, it would let users build and adapt real-time video world models from available checkpoints rather than training everything anew.","feed_headline":"Pipeline converts video models into interactive world models","feed_subtitle":"End-to-end method adds camera control and few-step autoregressive rollout to existing foundation models for low-latency use.","key_machinery":"The Causal Forcing / Causal Forcing++ pipeline, which performs autoregressive diffusion training, causal ODE or consistency distillation, and asymmetric DMD to turn bidirectional models into few-step autoregressive generators.","core_discovery":"minWM provides an end-to-end pipeline that converts existing bidirectional T2V/TI2V video foundation models into camera-controllable few-step autoregressive world models. Specifically, it first fine-tunes a bidirectional video diffusion model with camera control, and then applies the Causal Forcing / Causal Forcing++ pipeline, including AR diffusion training, causal ODE or causal consistency distillation, and asymmetric DMD, to distill it into a few-step autoregressive generator for low-latency rollout. The framework is modular and architecture-extensible across cross-attention-based and MMDiT-style models, and it also supports adapting existing video world models to new data distributions a","pith_inferences":["The modular design could support adding other control signals such as object motion or text instructions on top of camera trajectories.","The released scripts and checkpoints could serve as a starting point for testing the pipeline on newer or larger video backbones as they appear.","Community users might combine this conversion method with different distillation techniques to trade off speed and quality in new ways."],"forward_implications":["Existing bidirectional video models gain camera controllability through targeted fine-tuning.","Autoregressive rollouts at low latency become available after the distillation steps.","The same pipeline works on both cross-attention and MMDiT architectures.","Existing video world models can be adapted to new data and latency targets.","Practical minimums for batch size and training steps are established through ablations."],"fun_headline_variants":["minWM converts video models to camera-controllable autoregressive world models","minWM distills bidirectional models into few-step interactive video generators","Full-stack minWM adds camera control and causal distillation to video models","minWM pipeline turns T2V models into low-latency autoregressive world models","Open minWM framework enables real-time rollout from existing video diffusion models"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The Causal Forcing pipeline can reliably produce low-latency high-quality autoregressive rollouts from bidirectional models without major quality loss or instability.","fun_headline_variants_meta":{"raw":{"variants":["minWM converts video models to camera-controllable autoregressive world models","minWM distills bidirectional models into few-step interactive video generators","Full-stack minWM adds camera control and causal distillation to video models","minWM pipeline turns T2V models into low-latency autoregressive world models","Open minWM framework enables real-time rollout from existing video diffusion models"]},"model":"grok-4.3","cost_usd":0.004337,"raw_usage":{"total_tokens":2263,"prompt_tokens":841,"num_sources_used":0,"completion_tokens":83,"cost_in_usd_ticks":43374500,"prompt_tokens_details":{"text_tokens":841,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1339,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":841,"tokens_out":83,"duration_ms":11816,"temperature":1.0,"reasoning_tokens":1339,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T08:20:33.563844+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A side-by-side comparison on the same prompts showing that videos from the distilled few-step autoregressive models exhibit clear instability such as flickering, drifting camera paths, or loss of visual coherence over dozens of frames compared with the original bidirectional model.","supporting_citations":[],"review_version":1}