Hy-Nomas: Towards Free-form Interactive World Generation

Abstract: Interactive world generation requires models to follow successive commands while preserving relevant history and action consequences. We present Hy-Nomas, a video world model with an 8B diffusion backbone that supports free-form text instructions and continuous camera control across multiple interaction turns. It generates variable-length segments conditioned on current inputs and cached history. Our central design aligns structured video captions with action-aware causal attention: descriptive context remains accessible across segments, while action instructions are restricted to their intended temporal windows. A vision-language Prompt Enhancer (PE) translates user requests and multimodal interaction histories into structured conditions for the next segment. A multi-stage curriculum progressively introduces causal generation, text interaction, camera control, and training on model-generated histories, followed by diffusion distillation. In human evaluation, Hy-Nomas shows superior performance over existing world models. We release model weights, code and benchmarks, together with a lightweight adaptation workflow, to support reproducible research on interactive world generation.

• The paper proposes Hy-Nomas, an 8B-parameter diffusion transformer designed to generate interactive video worlds from free-form text and continuous camera trajectories, while preserving visual and action consistency over multiple turns.
• The system shows significant improvements in multi-turn text interaction, achieving 75.0% and 73.3% win rates for prompt following and physical plausibility against LingBot-World v2, respectively, but struggles with spatial memory compared to Genie 3 during camera revisitation.
• Hy-Nomas implements action-aware causal attention which ensures actions are correctly time-segmented in generated video, enabling more coherent interactive and robust training and inference pipeline.

Hy-Nomas addresses interactive video world generation as a causal, multi-turn prediction problem rather than as independent short-form video synthesis. The model is designed to accept successive free-form text instructions and continuous camera trajectories while preserving relevant visual history and action consequences. Its central claim is that interactive generation requires coordinated design across data representation, attention structure, autoregressive training, diffusion acceleration, and inference-time memory management. The paper presents an 8B-parameter diffusion transformer, a structured captioning pipeline, the Hy-Nomas-Bench evaluation suite, open model weights and code, and a LoRA-based adaptation workflow (2610.36729).

Problem formulation and system objectives

The paper defines a video world model as an interactive environment that renders future visual observations conditioned on the current world state, user commands, and prior interaction history. This formulation differs materially from conventional text-to-video generation. A fixed-prompt generator can optimize short-term visual quality without representing persistent state, action consequences, or causal ordering. By contrast, an interactive model must support a closed loop: the user observes a generated segment, issues a new command, and receives a continuation that remains compatible with all preceding events.

Hy-Nomas targets four related capabilities. First, it must follow free-form natural-language commands over multiple turns. Second, it must respond to continuous camera motion rather than only discrete navigation actions. Third, it must interleave semantic actions and camera control without resetting the session or switching between specialized models. Fourth, it must preserve spatial and object-level consistency when previously observed content re-enters the field of view.

The paper’s main architectural premise is that these requirements cannot be obtained merely by applying a temporal causal mask to a pretrained diffusion transformer. Causality must be aligned with the temporal scope of language supervision. Persistent scene descriptions should remain available to later segments, whereas an action instruction should not remain active indefinitely. This leads to the separation between descriptive context and localized action conditioning that underpins Hy-Nomas.

Data representation and supervision

Hy-Nomas is trained on a heterogeneous corpus containing real and virtual-environment videos with human and object motion, physical interactions, camera movement, and longer-range temporal dependencies. The data pipeline includes shot segmentation, transcoding, quality filtering, semantic deduplication, taxonomy labeling, camera-trajectory estimation, and structured captioning. Videos are clustered using visual embeddings, deduplicated within clusters, and labeled according to semantic and visual attributes. The resulting taxonomy permits adaptive sampling when evaluation reveals weaknesses in particular interaction types or camera regimes.

Camera supervision is derived primarily from estimated camera-to-world trajectories. The pipeline uses ViPE, optical-flow and point-tracking diagnostics, camera calibration, distortion checks, and trajectory audits to remove unreliable samples. A separate inverse-dynamics model predicts discrete camera-motion categories from video and is used for filtering and categorization. This combination is technically important because explicit camera conditioning is only useful when the motion labels are geometrically credible. The paper nevertheless does not provide a comprehensive estimate of trajectory-label precision or quantify how annotation errors affect downstream camera control.

The captioning scheme is more consequential than ordinary dense video description. Each video receives structured JSON containing visual style, camera perspective, persistent subject identities, static subject profiles, segment descriptions, timestamped actions, and visible action outcomes. The first segment establishes the scene and subject identities. Later segments describe changes relative to the previous state rather than repeatedly restating unchanged content. Actions are assigned non-overlapping temporal intervals bounded by their originating segment. Camera descriptions are marked independently with explicit tags so that they can either be removed when trajectory supervision is available or retained as textual supervision when it is not.

This representation supplies supervision for three distinctions that are often conflated in video generation: what remains true about the environment, what changes during a segment, and when a particular action is active. The paper argues that this decomposition reduces the burden on the diffusion model to infer temporal structure from appearance alone. That interpretation is plausible, but the paper does not report ablations isolating caption quality, subject identifiers, timestamps, and camera tags individually.

Architecture and action-aware causal attention

Hy-Nomas is initialized from HunyuanVideo 1.5 and uses an 8B-parameter dual-stream MMDiT backbone with 54 transformer blocks. Video is encoded by DCAE-V with $32\times$ spatial and $4\times$ temporal compression. The latent representation has 32 channels, and each latent position is treated as a token. Qwen2.5-VL-7B provides frozen text representations, which are projected and refined through timestep-conditioned modules. The model predicts latent flow velocity under a flow-matching objective.

During generation, each segment is denoised autoregressively from Gaussian noise. Keys and values from preceding text and video tokens are cached and supplied to subsequent segments. The architecture therefore retains a direct interface between long-range history and current-segment synthesis, although the size of this cache creates a substantial memory-management problem for long rollouts.

Segment-wise causal attention

The initial causal mechanism partitions video and text into aligned temporal segments. Tokens in segment $j$ can attend bidirectionally within that segment and causally to earlier segments, but cannot access future segments. This preserves intra-segment modeling capacity while preventing explicit future leakage.

The paper identifies two failures of this relatively coarse structure. First, unrestricted access to all action text within a segment can cause actions to occur in the wrong temporal order. Second, later segments can continue attending to earlier action descriptions, causing previously completed actions to recur. These observations motivate action-aware causal attention.

Under the refined mask, each segment contains a persistent descriptive head and action-specific text tails. Video tokens within an action interval can attend to the corresponding action text and the segment’s descriptive head. Descriptive heads remain available to later segments, while action tails are restricted to their original segment. The design therefore preserves environmental context but prevents both premature action execution and cross-segment action repetition.

This is one of the paper’s strongest conceptual contributions: causal attention is defined not only over video time but also over the temporal validity of language instructions. The approach assumes, however, that action intervals and segment boundaries are sufficiently accurate. If caption timestamps are wrong, the mask can enforce an incorrect causal structure rather than merely fail to provide useful supervision.

Continuous camera conditioning

Camera motion is represented through frame-to-frame relative camera-to-world transformations. Translation is normalized using a video-level 99th-percentile scale, and groups of four transformations are flattened into 64-dimensional tokens aligned with latent video frames. Camera tokens are injected through a dedicated cross-attention module in every MMDiT block, with visual tokens as queries.

Unlike systems restricted to discrete directional actions, Hy-Nomas supports continuous translation and rotation, allowing variation in direction, speed, and scale. The model is trained with per-segment camera dropout: some segments receive explicit trajectory conditioning, while others rely on text. Camera descriptions are removed when trajectories are provided to prevent textual leakage. This mixed-control regime is intended to support transitions between text-driven and camera-driven interaction within a single rollout.

The approach is technically flexible, but it does not guarantee metric geometric consistency. Relative poses are computed independently within each segment, and the generated video remains a learned visual continuation rather than a renderer tied to an explicit persistent 3D representation. The later spatial-memory results confirm that continuous camera control and long-term geometric memory are distinct capabilities.

Training curriculum and inference efficiency

The training procedure progressively converts a bidirectional video diffusion model into an autoregressive interactive model. It begins with latent-space adaptation to DCAE-V, followed by segment-causal continual pretraining, action-aware continual pretraining, camera-conditioned training, high-quality annealing, segment forcing, and diffusion distillation.

The context budget expands from 180K to 360K tokens during action-aware training. The authors state that the resulting 360K-token context is sufficient for minute-scale generation at 720p. Training such sequences requires Ulysses sequence parallelism, FSDP2 sharding, FLOPs-aware sequence packing, structured-mask attention kernels, activation checkpointing, CPU offloading, and asynchronous checkpointing. The systems contribution is substantial: computational cost depends not only on token count but also on the number of unmasked query–key pairs induced by the causal mask. Packing is therefore balanced by estimated FLOPs rather than by sample count or raw sequence length.

Segment forcing

A central training issue is exposure bias. If the model is trained only on ground-truth histories but deployed on its own imperfect predictions, small visual and state errors can compound over multiple segments. Segment forcing addresses this mismatch by sequentially processing segments, generating a one-step low-noise estimate for the current segment, and inserting that estimate into the cache used for subsequent segments. The current segment receives a flow-matching loss, while streaming backward releases earlier computation graphs and limits memory usage.

This procedure combines direct supervision of every segment with model-generated context. It is an important bridge between ordinary diffusion training and autoregressive rollout training. The paper does not, however, quantify the relative contribution of segment forcing versus action-aware attention or high-quality annealing through controlled ablations.

Distillation to four function evaluations

The base model uses 100 Euler steps per segment. A two-stage causal distillation procedure reduces this to four network function evaluations. Phased Consistency Models provide the initial few-step student, after which Decoupled Distribution Matching Distillation and an adversarial hinge objective refine visual fidelity, motion stability, and semantic alignment under hybrid forcing.

The student is trained partly on its own generated histories, and the final generated segment receives the primary distillation supervision. At inference, each generated segment is perturbed with a specified flow noise level before entering the causal cache. This design targets both latency and long-horizon stability, but the paper does not report a complete speed–quality–consistency curve across different numbers of function evaluations.

Inference is further optimized for a single NVIDIA RTX 4090 through INT8 weight and KV-cache storage, sliding-window attention, CPU offloading, sparse attention, and a one-step PixelDiT-based decoder. The paper reports a peak GPU memory requirement of 20.53 GiB for generating a 100-second video at 1280 × 704 resolution and 24 FPS across seven segments. This is a strong deployment result for an 8B model, although the authors explicitly acknowledge that quantization, sparse attention, and bounded history can introduce differences that become more visible over successive segments.

Prompt Enhancer and interaction interface

Hy-Nomas separates user-facing interaction from generator-facing conditioning through a VLM-based Prompt Enhancer. The PE receives the current user request, the complete dialogue history, previously generated video segments, and optional camera controls. It produces a structured caption and an appropriate duration for the next segment.

On the first turn, the PE establishes video-level metadata, visual description, and initial actions. On later text turns, it emits scene and camera changes relative to the established context together with new actions. On camera-control turns, it emits only scene and camera changes, leaving action fields empty while the explicit trajectory is passed separately to the generator. This interface makes the structured training representation available at inference without requiring users to formulate JSON captions themselves.

The PE is trained by supervised fine-tuning on synthetic and real user prompts, including multimodal multi-turn histories. According to the paper, it achieves nearly 100% adherence to the required response format, compared with approximately 50% for the original Qwen3.8-27B and 70% for Kimi K3. It also improves prompt following and physical plausibility relative to the untuned base model while preserving camera-control performance. These results establish the PE as an operational component rather than a cosmetic prompt-rewriting layer. They also expose a dependency: the quality of the world model’s interaction behavior partly reflects the PE’s ability to infer and formalize user intent.

Evaluation and empirical findings

Hy-Nomas-Bench contains 280 test cases across four benchmarks: 60 multi-turn text-interaction cases, 100 camera-navigation cases, 60 interleaved text–camera cases, and 60 spatial-memory cases. Evaluators compare model outputs pairwise on prompt following, physical plausibility, camera control, and memory consistency. The benchmark includes sequences longer than 60 seconds and tests reversible camera trajectories, mixed control modes, and viewpoint revisitation.

The comparisons involve Hy-Nomas, Genie 3, LingBot-World v2, and Happy Oyster, although the models do not share identical interfaces. Genie 3 is evaluated only on camera navigation and spatial memory because its official interface lacks free-form text interaction. Happy Oyster uses separate Directing and Wandering modes and does not provide native interleaved control. LingBot-World v2 is adapted through its open-source implementation. These interface differences make the comparisons informative about supported capabilities but limit their interpretation as strictly controlled model comparisons.

Multi-turn text interaction

Hy-Nomas achieves a 65.0% win rate against HappyOyster-Directing and a 75.0% win rate against LingBot-World v2 on prompt following. For physical plausibility, it achieves win rates of 45.0% and 68.3%, respectively. The results support the claim that Hy-Nomas is particularly competitive on free-form instruction execution and action sequencing.

The asymmetry between prompt-following and physical-plausibility gains is notable. Against Happy Oyster, the 45.0% physical-plausibility win rate does not by itself establish dominance without the corresponding loss and tie rates, which are not provided in the textual discussion. The paper’s stronger and more clearly supported claim is that the advantage is larger against LingBot-World v2, whose text interface relies on VLM-generated proposals rather than direct free-form user commands.

Camera navigation

Hy-Nomas achieves a higher win rate than loss rate against Genie 3, Happy Oyster, and LingBot-World v2 on the shared camera trajectories. The paper emphasizes that its advantage is smaller than on text interaction, consistent with camera navigation being a more established capability in prior world models. Exact aggregate win, tie, and loss values are not included in the supplied text, so the numerical strength of this comparison cannot be assessed as precisely as the text-interaction results.

Interleaved text and camera control

On the 60 interleaved-control cases against LingBot-World v2, Hy-Nomas achieves a 75.0% win rate for prompt following and 73.3% for physical plausibility. Camera control is much closer: Hy-Nomas wins 26.7% of comparisons. This pattern is important. It indicates that the unified architecture and structured conditioning provide a strong advantage for semantic interaction under mode switching, but do not produce a comparable advantage in geometric camera execution.

The result also qualifies the paper’s broader claim of unified control. Hy-Nomas succeeds in supporting both modalities within one session, but its relative strength depends strongly on the evaluation dimension. Interleaving controls is therefore not equivalent to solving the underlying camera-control and spatial-consistency problems.

Spatial memory

Spatial memory remains the clearest limitation. Against Genie 3, Hy-Nomas wins 13.3% of comparisons, ties 56.7%, and loses 30.0%. Against Happy Oyster, it wins 30.0%, ties 46.7%, and loses 23.3%. Against LingBot-World v2, it performs substantially better, with a 58.3% win rate and only a 1.7% loss rate.

These results contradict any interpretation that strong text interaction automatically implies persistent world-state modeling. Hy-Nomas can preserve many local consequences across segments, yet it remains inferior to Genie 3 when previously viewed environments must be reconstructed after camera departure and return. The paper attributes Genie 3’s advantage partly to grounding in Street View data, but this remains a hypothesis rather than an experimentally established causal explanation.

LoRA adaptation

The proposed LoRA workflow trains adapters on the base autoregressive model and merges them into the distilled model for deployment. The paper reports that adaptation can be performed with approximately 100 task-specific samples, although the specific examples use larger effective training schedules. A cinematic LoRA improves visual style and, relative to the distilled baseline, obtains 35% and 46% win rates for prompt following and physical plausibility, with losses of 24% and 13%. However, camera control regresses: the adapted model wins 19% and loses 37% of comparisons.

This trade-off is a useful empirical warning. Lightweight adaptation does not preserve all capabilities automatically. The likely issue is the narrow camera-motion distribution in the cinematic fine-tuning set, but the paper does not verify this with a data-diversity ablation. The robotics LoRA qualitatively improves dual-arm manipulation behavior on an out-of-distribution dataset, suggesting that the generator can absorb domain-specific visual and interaction priors without repeating full-model distillation.

Limitations and open questions

The main limitation is persistent spatial state. Hy-Nomas relies primarily on cached representations and learned visual continuation rather than an explicit scene graph, geometric map, or state transition model. The spatial-memory benchmark exposes the resulting weakness: performance is competitive with Happy Oyster but substantially below Genie 3. The paper proposes looped camera trajectories, retrieval of older observations, explicit state descriptions, and heterogeneous memory representations as possible remedies, but does not evaluate these mechanisms within Hy-Nomas.

The physical interpretation of the model is also limited. Hy-Nomas predicts visually plausible consequences of language and camera inputs, but the training data generally lack executable action annotations and explicit physical state variables. Consequently, physical plausibility in human evaluation should not be equated with reliable action-conditioned dynamics. The paper explicitly notes that robotics deployment would require videos paired with robot actions and evaluation based on task success, contact, grasp stability, and executable outcomes.

The benchmark has several methodological constraints. Human pairwise evaluation is appropriate for heterogeneous outputs and difficult world-model properties, but the reported comparisons are affected by unequal interfaces, model-specific initialization procedures, and the use of Hy-Nomas-generated initial frames for systems requiring images. The benchmark also focuses on four interaction dimensions and excludes conventional visual-quality metrics, making it unsuitable as a complete evaluation of generation quality. In addition, some claims are supported primarily by aggregate win rates without confidence intervals, inter-rater agreement, or statistical significance analysis.

The paper leaves open whether the reported gains arise primarily from structured captions, action-aware attention, the training curriculum, or the Prompt Enhancer. No systematic component ablation is reported in the supplied material. This matters because the system combines many interventions: data filtering, structured annotations, causal masks, camera cross-attention, self-generated histories, high-quality annealing, distillation, and PE rewriting. The empirical evidence establishes the performance of the complete system more strongly than it establishes the necessity or sufficiency of any individual component.

Finally, efficient inference introduces its own compromises. INT8 KV caches, sliding-window attention, sparse attention, CPU offloading, and a learned decoder reduce memory requirements, but the authors acknowledge that errors can become more visible over successive segments. The reported 20.53 GiB peak memory for a 100-second rollout is promising, yet latency, throughput, and quality degradation under longer horizons are not characterized comprehensively.

Conclusion

Hy-Nomas presents a coherent system for free-form interactive video world generation. Its most significant technical contribution is the alignment of structured temporal captions with action-aware causal attention: persistent descriptions remain available across segments, while action instructions are confined to their intended temporal scopes. This design is integrated with continuous camera-trajectory conditioning, self-generated-history training, four-step diffusion distillation, bounded-cache inference, and a VLM-based Prompt Enhancer.

The empirical results support strong performance in multi-turn text interaction and interleaved semantic control, including 75.0% prompt-following and 73.3% physical-plausibility win rates against LingBot-World v2 in interleaved tests. The results also establish a clear limitation: spatial memory remains weaker than Genie 3, with a 13.3% win rate and 30.0% loss rate in direct comparison. Thus, the paper demonstrates that free-form instruction execution and unified control can be improved through causal supervision and curriculum design, while also showing that these advances do not by themselves provide persistent geometric world state.