Skip to content
Bud's Wiki

Context Length / Quality Trade-off in Video Generation

TL;DR

Section titled “TL;DR”

A fundamental trade-off in autoregressive video DiTs: longer context window → better identity / scene / motion consistency across the rollout, but quadratic attention cost and noisier conditioning. VAE channel bottlenecks, learned patchifier compression, sparse / FIFO / surprisal-driven KV caches, learned memory encoders, bounded positional encoding, and training-time degradation perturbations all sit on different points of this Pareto. The filed papers explore four operating regimes — explicit compression (PFP), functional cache partitioning (Context Forcing), curriculum-based consistency injection (LongVie 2), and partitioned-cache-with-learned-mid-compression (PackForcing).

Key claims

Section titled “Key claims”

Recent contributions

Section titled “Recent contributions”
  • [2026-11-05] GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling: GameVerse (He et al.) reframes the long-horizon-coherence problem as a training-data availability bottleneck rather than an architectural one: 1,358 h of continuous two-minute gameplay sequences with event-boundary-aligned captions and temporally stable per-clip camera trajectories. Direct counter-datapoint to the “short clips + architectural workaround” position of PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference PackForcing (5s training → 2-min extrapolation via learned mid-buffer compression) and Context Forcing: Consistent Autoregressive Video Generation with Long Context Context Forcing (short clips + partitioned KV cache): claims that training on genuinely minute-scale data substantially extends the temporal horizon on its own. Leaves open whether the data-side gain replaces or stacks with the cache-side workarounds — no head-to-head reported.
  • [2026-11-05] ERNIE 5.5 Flash Base: ERNIE-5.5-Flash-Base (Baidu, 280B/14B MoE) is the most aggressive context-compression ratio filed: 37 of 43 attention layers compress the history 128×-per-layer via HCA (Heavily Compressed Attention), while only 6 layers retain token-level access via MLA + DSA sparse attention. Trained with context gradually extended to 128K during pretraining; supports 1M fine-tuning and inference context. Despite the aggressive compression, long-context benchmarks hold up: HELMET-RAG 71.5 (leads vs V4-Flash 68.9 / Hy3-preview 67.2 / MiMo-V2.5 69.0), HELMET-ICL 85.9 (leads), RepoQA 91.7 (leads vs MiMo-V2.5 56.5), LongBench-v2 48.7 (leads). Explicitly demonstrates that keeping a small number of full-context layers is enough to cover the long-context recall the compressed layers cannot serve — a per-layer-specialization answer to the pure-sparse-attention position.
  • [2026-10-22] Interactive World Simulation with Cosmos for Physical AI: Cosmos Simulation’s streaming inference §4 names history eviction + K/V capacity management as a first-class operating lever — when adding new entries would exceed cache capacity, older history tokens are evicted to keep retained history within budget. Per-example configurations give different history-retention policies for action-conditioned (latent-time-order refresh within completed chunk) vs. multi-view AV (clean-history visibility rules per view).
  • [2026-10-16] Introducing Muse Spark Open — The Most Capable US Open-Weight Model to Date: Muse Spark Open advertises a 1M-token context window at the 816B-total / 42B-active scale, positioning it alongside DeepSeek-V4.1-Flash (1M context) and Kimi K3 (1M native) in the “1M context × frontier-scale open MoE” cohort. No architectural detail on how the context budget is realized (attention variant, sliding window, routing scheme) at the blog level.
  • [2026-10-09] SemanTok: Predictable Semantic Tokens for Efficient Autoregressive Video Generation: SemanTok extends VideoFlexTok’s “compress at the tokenizer” operating point: by closing the REPA leakage path (encoder-input DINO + prefix-only DINO reconstruction heads), the first few tokens carry genuine semantics and short prefixes are both cheaper to AR-predict and give better generation fidelity. Reinforces that the tokenizer-side approach is still improving along axes orthogonal to the cache-side (Context Forcing / PackForcing / PFP) approaches.
  • [2026-10-08] Long-WAM: Scaling the Context of World-Action Models: Long-WAM (NVIDIA + MIT + HKU + UCSD) adds a conditional law to the robotics side of this Pareto: longer visual context helps WAMs only when the video foundation was pretrained autoregressively. On RoboCasa GR-1, 0.0s → 19.2s context raises success 63.3% → 78.7% for an AR init while a bidirectional init shows no net gain over the same sweep. Builds on LongLive-2.0’s AR checkpoint with ~10k hours of causal robot+egocentric continued pretraining, then WAM-adapts while preserving the history-to-future structure. Also demonstrates that long context is compatible with real-time control via streaming observation encoding + async execution: 107.4 ms per action chunk (incl. future-video latent prediction) on RTX 5090. Sharper statement than RoboTTT — 8k-timestep robot policy via test-time training (Jim Fan / NVIDIA GEAR) / WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time: for WAMs, pretraining-time causality — not just raw history — determines whether a model can use context.
  • [2026-10-02] Veda: Scalable Video Diffusion via Distilled Sparse Attention: Veda characterizes the specific failure mode of overly-sparse attention in video DiTs: at >90% sparsity prior methods produce water-ripple warping, local geometric distortion, and frame-to-frame temporal flicker — not ordinary generator hallucination, but a signature of broken spatiotemporal attention connectivity. Also demonstrates 95% sparsity is sustainable at 720P / 241 frames when the sparsity mask is distilled to preserve the oracle (max-pooled full-attention) tile distribution rather than cherry-picked per-key.
  • [2026-10-01] Improving Video Sparse Attention with Fine-grained Router and Sparse Rebasing: VSA2 inverts the standard sparsity-costs-quality framing: models trained with Hard-to-Easy curriculum (higher sparsity in training than inference) reportedly outperform full-attention counterparts on motion quality. If the claim holds, the sparse-attention/quality frontier for video DiTs may be bounded by the training recipe rather than the attention pattern itself — adding a new datapoint on the open question of whether aggressive context pruning can be quality-free given the right training discipline.
  • [2026-09-30] Google DeepMind announces Gemini 4 Argon — new frontier model with 1M-token output for coding, enterprise knowledge work, and cybersecurity defense: Gemini 4 Argon claims a 1M-token output limit — a 16× jump over the Gemini 3.7 Flash 64K output cap (Gemini 3.7 Flash Model Card) — with no accompanying quality-vs-length data. The tension between the “in one go” framing and the tradeoff this concept tracks is unresolved until benchmarks land.
  • [2026-09-22] WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory: WorldCrafter adds a new operating point: a learned memory encoder + pose-conditioned readout module that compresses the entire history into a fixed set of target-view-specific tokens before denoising — view-dependent compression rather than the time-only compression used by MALT, Context Forcing, PackForcing, and PFP. Sibling to Mirage’s explicit 3D-indexed latent cache but skips the depth-based back-projection entirely; 3D awareness is implicit in the learned encoder. Combined with recent temporal context and few-step distillation for streaming minute-scale exploration.
  • [2026-09-16] Closed-Loop Policies: Extends this concept from the video-DiT side (KV cache, long-context video generation) to the manipulation-policy side: an independent replication of the history × execution-horizon findings for behavior-cloning policies, sibling to Revisiting Open-Loop Execution in Robotics: Toward Reactive, Higher-Performing Policies, Simulation Pre-training for Dexterity (SPD) SPD, and Behavioral Cloning Mystery. Load-bearing new claim (per author summary): long-context policies require longer training to learn to use the additional history — a training-budget cost, not just an architectural change.
  • [2026-09-03] SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models: SolarWM claims real-time interactive rollout from minutes to hours from causal models trained only on 5-second sequences — 720×–4320× extrapolation ratio in the same regime that Context Forcing: Consistent Autoregressive Video Generation with Long Context and Self-Forcing++: Towards Minute-Scale High-Quality Video Generation target, with the bidirectional-to-causal-to-DMD stack as the load-bearing recipe rather than explicit KV-cache partitioning or memory compression.
  • [2026-09-01] Atlas: A World Model for Spatial Intelligence: Atlas turns the reconstruction-vs-imagination trade-off into a runtime dial via spatial context: it accepts 1 to 100+ reference images placed at explicit 3D positions and behavior shifts monotonically from imagination toward faithful reconstruction as inputs grow — “the more it sees, the less it imagines” — supporting 1-minute 1440p output at the long end.
  • [2026-08-28] OpenDM — DM0.5: An Open-World Foundation Model for General-Purpose Embodied Intelligence: DM0.5 (Dexmal) supplies a robotics-domain datapoint on the “just make the KV window bigger” side of this axis — up to ~60 seconds of historical visual context in a Gemma3-4B VLM + 680M flow-matching Action Expert VLA, used both for long-horizon coherence and for a specific in-episode learning-from-a-single-human-demonstration experiment where the human’s start-of-episode demonstration and the robot’s subsequent execution phase must both fit in the same context window. Contrasts with the memory-augmentation answer (MEM: Multi-Scale Embodied Memory for Vision Language Action Models MEM’s dual-timescale memory) and the memory-into-weights answer (RoboTTT — 8k-timestep robot policy via test-time training (Jim Fan / NVIDIA GEAR) RoboTTT / WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time WAM-TTT) — three different substrates for the same operating goal, never benchmarked head-to-head at matched horizons.
  • [2026-08-27] Gemini Omni 1.1 Flash lets you build with more control: Gemini Omni 1.1 Flash surfaces the tradeoff as a product-side chunked API: 10 seconds of prior clip context per extension call (up from 1s in the launch model), extendable in 10s increments to a 40s cumulative cap — a much shorter horizon than open long-AR-video recipes (LongVie 2 at 5 min, PackForcing at 2 min) at unspecified fidelity trade-off; no benchmark tables, so unclear whether the cap is a quality floor or a product decision.
  • [2026-08-21] 4DAnyone: Create Anyone in 4D from a Casual Monocular Video: Reference Context Packing (RCP) is a novel operating point: instead of evicting or sinking old tokens, it packs a growing reference set into a fixed-length mixed-resolution context to keep reference-attention O(1) while preserving cross-group appearance guidance — applied to camera-controlled video DiTs generating dense target views for 4DGS lifting.
  • [2026-08-14] Qwen3.8 open weights release — 27B dense multimodal + 2.4T-A95B MoE (Alibaba Qwen): Qwen3.8-27B lists 262K native context extendable to 1M via YaRN — Alibaba continues to publish YaRN as the extrapolation path in launch materials, without noting any independent long-context quality benchmarks in the announcement.
  • [2026-08-10] MEM: Multi-Scale Embodied Memory for Vision Language Action Models: MEM’s video encoder factorizes attention into spatial + causal temporal (every 4th layer), reducing joint T×N attention cost from O((TN)²) to O(TN² + T²N) while dropping past-timestep tokens after the encoder — the VLA backbone sees the same token count as a single-image VLA, so temporal information is compressed into the current-frame tokens rather than expanding the downstream context length.
  • [2026-08-08] MiniWorld: Democratizing the Training of Video World Models from Scratch: MiniWorld combines a rolling KV cache with pipelined asynchronous denoising to keep per-step compute bounded during streaming AR video generation from a from-scratch-trained block-causal DiT.
  • [2026-07-31] Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers: Chimera reports zero-shot temporal extrapolation from 5-second training clips to 30-second video generation — 6× duration extension with only 6.5% FID degradation in the last five seconds and no length-specific fine-tuning. Attributed to the KDA + MLA hybrid attention’s fixed-state long-context behavior operating on a shared parameter path with no positional embeddings; adds a new operating point where the length-extrapolation mechanism is architectural (no RoPE to break, linear-attention state carries history) rather than cache-partitioning or memory-encoder-based.
  • [2026-07-31] Towards Video World Models: Xun Huang position essay frames the wiki’s forgetting-drifting problem as “persistence” — a soft constraint for video WMs that is architecturally tightened by the real-time constraint: unlike LLMs where users tolerate growing inference time with context, video WMs face a hard latency ceiling and cannot slow down over time. Names three complementary axes tracked on this page and elsewhere — aggressive compression of temporally-distant frames (FAR, FramePack; “needle in a haystack” failure mode named), integration with persistent 3D conditions (GEN3C, WVD, SPMem; brittle in dynamic scenes), and linear-RNN long-range recurrence (Po et al. 2025, ConvLSTM ancestry).
  • [2026-07-29] Wonder: Video World Model Done Better: Wonder adds a new operating point: sparse-attention selection over full-fidelity history KV caches (no partition, no compression, no latent warping) for constant-latency retrieval at arbitrary context length; reports 16 FPS at 0.5s constant latency up to one minute of rollout with return-to-previously-observed regions preserved.
  • [2026-07-27] Kimi K3 Technical Report — Open Frontier Intelligence: Kimi K3 tech report — 1M-token context deployed via the KDA/Gated-MLA 3:1 hybrid with NoPE on all MLA layers (position-sensitivity carried by KDA), removing RoPE-frequency retuning / YaRN when extending context.
  • [2026-07-24] SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation: Compiled DiT forward pass is 3.2× faster than a matched full-softmax baseline at 720p/60s, with the gap expanding at longer durations — a controlled hybrid-vs-softmax runtime scan at video length that matches the LLM-side “throughput advantage grows with context” pattern from Jet-Nemotron and Kimi Linear.
  • [2026-07-21] AlayaWorld: Interactive Long-Horizon World Modeling: AlayaWorld realises constant per-chunk compute at in-principle unbounded horizon via a bounded rolling prefix — 1 fixed sink frame + a 6-latent-frame compressed temporal memory + a capped 10-frame spatial-memory render + 1 I2V nearby frame — attended jointly with the noised target. The sink is pinned at RoPE position 0 and held fixed across all chunks as a global identity anchor (during training drawn as a remote frame ≥8 latents away from the target so the model cannot extrapolate the next chunk from it).
  • [2026-07-17] HDR: Hierarchical Denoising for Multi-Step Visual Reasoning: HDR’s Sparse Hierarchical Attention Pattern gives each token access to local + parent + first-frame context only, so effective context grows logarithmically via a tree hop rather than linearly with a long window — a distinct point on the trade-off curve from KV-partitioning (Context Forcing) or memory-compression (PFP) approaches. Enables coarse plans to reach any leaf in at most one vertical hop while attention cost stays sparse.
  • [2026-07-16] Kimi K3 — Open Frontier Intelligence (2.8T MoE with KDA + AttnRes): Kimi K3 offers 1M-token native context and reports BrowseComp = 90.4 at 1M tokens with no context management, vs the Claude-card 300K compaction baseline — an explicit test of whether frontier agentic quality survives full-length context on a hybrid-linear-attention backbone.
  • [2026-07-15] RoboTTT — 8k-timestep robot policy via test-time training (Jim Fan / NVIDIA GEAR): RoboTTT (Jim Fan / NVIDIA GEAR launch tweet) extends the concept’s Pareto to robotics: instead of a KV-cache partition or a recurrent memory vector cross-attended into a DiT, the entire history is compressed into a small TTT core’s weights via one SGD step per sensor reading, at constant per-step inference cost. The reported Context Scaling Curve is the strongest “no saturation” datapoint yet from the robotics side — 128 → 8K timesteps of pretraining hill-climbs monotonically, with 8K beating 1K by 62% closed-loop; Fan claims 1M-timestep robot context is reachable. Launch-tweet only; no benchmark numbers.
  • [2026-07-09] LingBot-World 2.0 / LingBot-World-Infinity — Infinite Worlds with Versatile Interactions: LingBot-World 2.0 claims unbounded interaction horizon at 720p / 60 fps via causal pretraining plus attention-sink + local-window inference (CLI: --sink_size 6 --local_attn_size 18) — a new open-source operating point on the horizon vs. quality Pareto that pushes past v1’s ~10-minute ceiling; no quantitative long-horizon benchmark numbers disclosed in the README at filing time.
  • [2026-06-09] Latent Spatial Memory for Video World Models: Mirage compresses long-horizon scene context via a 3D-indexed cache of diffusion latent tokens rather than via in-network KV partitioning, learned bottlenecks, or RGB-space point clouds. The depth-guided back-projection is content-adaptive geometric compression, but on the latent manifold the generator already uses — avoiding both the quadratic-attention cost of long context windows and the encode/decode information loss of SPMem-style RGB caches.
  • [2026-06-08] Scaling Video Training with Parallelism: The systems-side enabler for the long-video-DiT recipes filed in this concept. Argues that long-video training requires within-sample (sequence/context) parallelism, not just across-sample DP — and that for AR teacher forcing the SP unit must be a temporal chunk holding matched clean-history + noisy-target latents (Balanced SP), not a slice of the concatenated clean/noisy sequence. Pairs with LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation (whose training stack uses this), and with LongVILA’s MM-SP for the understanding side.
  • [2026-06-06] Direct Video-Action Models — Causal Video Models Are Data-Efficient Robot Policy Learners: Rhoda AI’s DVA makes “hundreds of frames of video context, natively” the load-bearing capability for robotics — explicitly contrasted with the few-frame context of typical VLAs. The Context Amortization training strategy is what makes the long-context regime trainable: every position in a long sequence supervises a noisy-future prediction off noise-free history, so the long-context conditioning the inference mask requires is also what’s seen densely at training. Validated downstream by a shell-game memory probe (object tracked across multiple swaps) and an end-to-end returns-processing deployment that resolves visually ambiguous current frames purely from video history.
  • [2026-06-02] Tilde Research — Wall Attention: length generalization via diagonal forget gates in softmax attention: Tilde’s Wall Attention announcement — claims 4k-training → 200k+-token generalization with no continued pretraining, by replacing RoPE with per-channel diagonal forget gates inside softmax attention. If reproducible, this is a much wider extrapolation ratio than the FoX / Gated-Attention baselines it’s pitched against.
  • [2026-05-27] Planning at Inference: MCTS Test-Time Scaling for Long Video Generation: Inference-time MCTS over candidate continuations as a way to push long-horizon consistency without enlarging the context window: the search tree externalizes the consistency burden onto a verifier-scored rollout rather than into more attention context.
  • [2026-05-26] MIGA: Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos: MIGA — first train-free operating point on the long-AR-video Pareto: a sampler wrapper over FIFO-Diffusion that uses a zigzag noise schedule (Two-stage Training-inference Alignment) to narrow the noise span per forward pass, plus Dual Consistency Enhancement (anomaly-triggered self-reflection at the queue tail + long-range frame guidance from the queue head). Sits parallel to PFP / Context Forcing / PackForcing / LongVie 2 / MALT, all of which require training; useful baseline for separating “training helps” from “scheduling + cross-time attention help”.
  • [2026-05-25] SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers: SEGA addresses the spatial analogue of RoPE long-context degradation: dynamically rescales RoPE components by FFT-derived per-band spectral energy to recover structure and detail at out-of-training-distribution image resolutions (4096² from FLUX / Qwen).
  • [2026-05-25] Decoupling the "What" and "Where" With Polar Coordinate Positional Embeddings: PoPE — Polar Coordinate Positional Embeddings. Re-expresses RoPE in polar form to localise the what-where confound to a single interaction term in the dot product, then removes it by giving each key/query element a softplus magnitude and a position-only phase p ωip\,\omega_i, with an optional per-head learnable phase bias. Solves an Indirect Indexing diagnostic at 95% vs RoPE’s 11%, lowers perplexity on MIDI / human genome / OpenWebText at 124M–774M, and extrapolates from 1024 to 10240 tokens on PG-19 without interpolation while RoPE diverges. Constructive counterpart to RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably‘s lower-bound result. No vision experiments — leaves open whether the axial RoPE schedules used in image/video DiTs admit the same drop-in re-parameterisation.
  • [2026-05-24] Rotary Positional Embeddings Explained | Transformer: Reference explainer covering the RoPE construction itself (2D rotations on Q/K feature pairs, frequencies set by base θ, relative-offset property of the post-rotation dot product, axial extension to 2D/3D inputs) and a from-scratch PyTorch implementation. Background for the RoPE-side claims and workarounds in this concept; not itself a new operating point.
  • [2026-05-24] RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably: Provides the missing theoretical foundation — proves that RoPE-based attention loses both locality bias and token-relevance consistency as context grows (failure probability → 0.5), and that the RoPE-base hyperparameter cannot simultaneously preserve position- and token-distinguishability. Recasts every RoPE-side patch already in this concept (Bounded Positional Encoding, incremental temporal-only RoPE adjustment, relative-PE-inside-memory-attention) as necessary workarounds for a content-independent lower bound. The author’s accompanying takeaway — that agentic / context-folding approaches may be more effective than scaling raw context — links this concept to the recursive-LM / context-folding cluster on the LM side.
  • [2026-05-24] MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation: MALT Diffusion — adds a fifth operating point: a single recurrent memory vector of fixed size, built by attention over prior segments’ hidden states, cross-attended at every diffusion-Transformer block, with relative positional encoding inside memory attention as the length-extrapolation mechanism and noise augmentation on the memory vector at training as the on-manifold regulariser. Training cost is independent of video length (stop-grad through prior segments). Conceptually upstream of PFP (explicit compression), Context Forcing (partitioned cache), and PackForcing (partitioned cache + mid compression) — all three can be read as evolutions of MALT’s memory-vector idea in different directions.
  • [2026-05-23] PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference: PackForcing — a three-partition KV cache (Sink + Compressed Mid + Recent) where the mid partition is compressed 16× by a learned dual-branch module (HR 3D-conv + LR pixel-space-re-encode), then accessed via dynamic top-k retrieval; an incremental temporal-only RoPE adjustment closes position gaps from mid-block eviction. Trained on 5s clips, extrapolates 24× to coherent 2-minute generation at 16 FPS on a single H200, with VBench Dynamic Degree 56.25 / Overall Consistency 26.07 at 60s and CLIP-T drop under 1.2 points over the full 60s. A fourth operating point on the Pareto: combines Context Forcing’s partitioning idea with explicit learned mid-buffer compression.
  • [2026-05-22] LongVie 2: Multimodal Controllable Ultra-Long Video World Model: Adds a third operating point — multi-minute (5-min) continuous generation via a multi-stage training curriculum (multimodal guidance → degradation-aware training → history-context guidance) rather than explicit cache or memory-encoder design.
  • [2026-05-22] Context Forcing: Consistent Autoregressive Video Generation with Long Context: Functional KV-cache partitioning (Attention Sink / Slow Memory / Fast Memory), surprisal-driven consolidation, and Bounded Positional Encoding. Names the “forgetting–drifting dilemma” as the underlying trade-off vocabulary.
  • [2026-05-22] Pretraining Frame Preservation in Autoregressive Video Memory Compression: First filed datapoint that isolates the compression sub-problem with a “preserve any random frame perfectly” pretraining objective; recommends 4×4×2 latent compression as the sweet spot.

Open questions

Section titled “Open questions”
  • Can the five approaches be composed? Memory-encoder compression (PFP) + functional cache partitioning + surprisal slow-memory (Context Forcing) + learned mid-buffer compression with top-k retrieval (PackForcing) + curriculum-based consistency (LongVie 2) + recurrent memory vector (MALT) target different parts of the same Pareto and look stackable, but no filed paper has tested combinations. PackForcing and Context Forcing in particular share the partitioning idea but differ on whether to select or compress the long-term region — a combined approach (compress then select) is the obvious next experiment. MALT is the fully-recurrent extreme — does adding a partitioned/compressed cache on top of MALT’s memory vector help?
  • Where does the surprisal-promotion threshold τ come from? Context Forcing reports the choice is robust but doesn’t characterize its dependence on content type (dense-shot-shift vs. single-long-shot); PFP claims that dense-shot-shift hides error accumulation entirely.
  • PackForcing’s dual-branch HR+LR compressor, PFP’s bottleneck-bypassing residual encoder, and MALT’s recurrent-attention memory encoder are three different recipes for the same compression sub-problem — none has been compared head-to-head on the same backbone or benchmark.
  • Should compression rates be content-adaptive? PFP shows that combining 4×4×2 + 2×2×4 helps preserve text — a static per-region compression mix — but no filed paper has tried adaptive compression where the rate is chosen per-frame or per-region. PackForcing’s top-k retrieval is temporally adaptive but not spatially adaptive.
  • The Pareto evaluation protocols don’t agree: PFP uses user-study ELO + Cloth/Identity/Instance consistency on internal benchmarks; Context Forcing uses VBench + windowed DINO/CLIP-F/CLIP-T on MovieGenBench prompts; PackForcing uses VBench-Long with 128 MovieGen prompts + CLIP-T at 10s intervals; LongVie 2 introduces LongVGenBench; MALT reports Kinetics-600 long-video prediction. A shared benchmark for the long-AR-video Pareto is missing.
  • All four “Wan-era” filed papers are Wan-family. How much of the reported tradeoffs are an artifact of Wan2.1/Wan2.2’s specific VAE channel width, RoPE schedule, and patchifier rates? MALT’s results are on W.A.L.T (a different DiT-family) and would be the cleanest cross-backbone replication target.
  • Bounded Positional Encoding (Context Forcing), incremental RoPE adjustment (PackForcing), and relative-PE-inside-memory-attention (MALT) all address a RoPE-side problem in long-horizon AR rollouts but with different mechanisms. Are they substitutes or complements? RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably gives the lower-bound framing but doesn’t compare workarounds directly.
  • If RoPE has a content-independent failure floor (RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably), how much of the long-context win in the video-DiT setting is from the workarounds vs. from non-RoPE inductive biases (e.g. spatial attention windows, compressed memory)? The video setting may be more forgiving than the text setting because spatial structure provides redundant positional signal.
  • Does PoPE (Decoupling the "What" and "Where" With Polar Coordinate Positional Embeddings) work on axial / 2D / 3D positional schedules used in image and video DiTs? PoPE’s text-side extrapolation result (1024 → 10240 PG-19, zero-shot) is the strongest length-generalization datapoint among RoPE alternatives so far, but the paper runs no vision experiments and the standard image/video extension of RoPE applies separate frequency banks per spatial axis — it isn’t obvious that PoPE’s per-element softplus-magnitude / position-only-phase construction transfers cleanly.
  • Is PoPE composable with the cache-side workarounds in this concept (Bounded PE, incremental temporal RoPE adjustment, relative-PE-inside-memory-attention)? Both directions plausibly help — PoPE removes the underlying confound; the cache-side patches restrict the regime — but they touch overlapping parts of the model and no filed paper has tried the combination.

Papers

Section titled “Papers”

2026-05

Section titled “2026-05”

2026-04

Section titled “2026-04”

2026-03

Section titled “2026-03”

2026-02

Section titled “2026-02”

2026-01

Section titled “2026-01”

2025-12

Section titled “2025-12”

2025-11

Section titled “2025-11”

2025-10

Section titled “2025-10”

2025-09

Section titled “2025-09”

2025-08

Section titled “2025-08”

2025-06

Section titled “2025-06”

2025-05

Section titled “2025-05”

2025-02

Section titled “2025-02”

Unknown date

Section titled “Unknown date”