acceptodds
Withdrawn from ICLR 2027

StreamCeption: Self-Correcting Latent Repair for Long Streaming Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Streaming video generation usually suffers from severe error accumulation when imperfectly generated latents are iteratively reused as fixed history. Inspired by variable-length generation paradigms like Flowception, we introduce StreamCeption, which adopts per-frame denoising and bidirectional attention but fundamentally shifts the generation paradigm. Instead of inserting latents into temporal gaps, StreamCeption consecutively appends noisy latents within a bounded active window, decoupling denoising completion from final frame output. While output frames become fixed history, the active window temporarily retains the oldest fully denoised latent. As subsequent trailing latents gain structural coherence during their denoising trajectories, a learned repair head adaptively determines whether to re-noise and refine this retained latent, correcting early errors before finalization. Trained alongside flow matching via self-sampled repair trials, our method uses reconstruction improvements to supervise repair selection and reconstruction loss to guide execution. Instantiated on the Wan2.2-TI2V-5B backbone, StreamCeption demonstrates the immense potential of this variable-length repair mechanism. Extensive experiments show that our approach scales effectively to this 5B model, achieving stronger long-horizon consistency than competitive streaming baselines while maintaining comparable visual quality, and seamlessly generalizing from an 81-frame training limit to highly coherent 162- and 320-frame videos with limited additional degradation over extended horizons.

Primary area:
generative models
Keywords:
Streaming Video generation, Flow-matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WVJROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes StreamCeption, a streaming video generation framework that consecutively appends and denoises noisy latents within a bounded active window. It learns to selectively repair fully denoised internal latents using subsequently developed temporal context to mitigate error accumulation during long rollouts, while leaving previously displayed video unchanged.

The authors evaluate the effectiveness of the proposed method on VBench, where StreamCeption achieves the highest score at the five-second duration on consistency metrics among several baselines, including Causal Forcing, Mask Forcing, and Rolling Forcing. Additional results in the appendix (Table 3) validate the performance gain at the ten-second duration achieved by the proposed repairing mechanism.

Critical strengths

This paper implements a sound and straightforward idea to mitigate temporal drift caused by error accumulation. It learns to repair the history that contains prediction errors and reuses it to generate new frames.

Critical weaknesses

The authors provide only weak experimental evidence, which is insufficient to support their claims.

  1. Table 1 does not report Dynamic Degree scores. The higher consistency scores may therefore reflect reduced motion in the generated videos.

  2. The paper lacks comparisons with baselines of comparable model size, as Rolling Forcing, Causal Forcing, and Mask Forcing are all implemented using 1.3B models. It also does not report results for Causal Forcing and Mask Forcing at longer durations, such as 10 and 20 seconds.

  3. The qualitative comparison in Figure 4 includes only five-second results. As a method for “long streaming video generation,” the paper should provide more qualitative comparisons at longer durations.

  4. The main text contains no ablation studies, and the appendix provides only one ablation examining the effect of enabling the repair mechanism.

Questions for authors

Novelty and Positioning.
Beyond the learned repair mechanism in Contribution 2, it would be helpful to clarify what independent methodological novelty Contribution 1 introduces. In particular, I would like to understand how separating video output from internal history modification differs from existing strategies. It appears to be a simple and common approach that refines or re-noises existing history contexts (e.g., Diffusion Forcing, Helios). I am not fully convinced that only outputting newly generated frames first and then reusing them via re-noising or refining is sufficiently novel. It would also be helpful to clarify the technical connection to Flowception, especially since StreamCeption removes Flowception's learned insertion mechanism and instead appends latents sequentially.

Effectiveness of the Repair Operation.
I have several concerns regarding the effectiveness of the repair operation.

  1. One concern is whether the error in a frame to be repaired may already have propagated to newer frames, leaving no reliable information from which to repair the target frame. In that case, choosing to repair could lead to worse results, while choosing not to repair may allow the error to continue accumulating and ultimately cause the method to fail.

  2. Another concern is that newer frames depend on earlier frames that need repair. It is unclear how one can guarantee that a newer frame predicted from a frame containing prediction errors can help recover a better frame to be generated.

  3. It would also be useful to study the effect of performing repair multiple times. As in point 2, if a newer frame predicted from an erroneous frame can effectively repair the frame to be generated, then re-predicting with the repaired frame should yield better results. Conversely, if better repair results can be obtained from a better newly predicted frame, it is worth examining whether multiple repairs lead to better generation results. If multiple repairs make the results worse, this may suggest that the method fails in such cases.

  4. During training, repair is always required to construct pseudo-labels. It would be helpful to further clarify the additional training overhead this introduces. In addition, it is unclear how one can ensure that the constructed labels are correct during training, especially in the early stage.

  5. It is also unclear why only one latent is repaired at a time. If the motion generated by the model collapses, it is not clear how it could be repaired. More ablation studies would be helpful to explain why only one latent is repaired, since this may make it impossible to repair the motion generated by the model.

Experiments.

  1. VBench results should report the Dynamic Degree score to demonstrate that the improvement in consistency is not achieved by generating more static videos.

  2. It would be helpful to add more baselines with similar model sizes, such as LongLive-2.0, because Rolling Forcing, Causal Forcing, and Mask Forcing are all 1.3B models.

  3. Metrics for measuring drifting should also be added, for example, by computing the difference between the score of the last chunk and that of the first chunk.

  4. It would also be helpful to report metrics for selective repair, such as the accuracy and recall of the model's decision on whether repair is needed on controlled or annotated frames.

  5. More visualization results at 20 seconds would be helpful, since the results shown on VBench (Figure 4) are only 5 seconds long.

Reviewer_A9ULOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes StreamCeption, a streaming video generation method that mitigates error accumulation by keeping fully denoised latents revisable. Built on Flowception's variable-length per-frame denoising, StreamCeption consecutively appends noisy latents within a bounded window and decouples denoising completion from final video output: when the oldest active latent reaches zero noise, it is held while later latents develop temporal context, and a learned binary repair head decides whether to re-noise it at fixed strength and refine it via bidirectional attention before it enters the fixed history; user-visible output remains immutable throughout. Repair selection and execution are trained jointly with flow matching through self-sampled repair trials on the model's own reconstructions, where measured reconstruction improvement supervises the keep-or-repair decision and endpoint reconstruction loss guides the correction. Instantiated on the Wan2.2-TI2V-5B backbone with LoRA fine-tuning on roughly 300K clips of about 3.4 seconds, and evaluated with VBench-I2V at 5, 10, and 20 seconds, StreamCeption achieves the best subject consistency, background consistency, and motion smoothness among the compared streaming baselines. A controlled ablation shows consistency gains of 0.75-1.08 points from enabling repair, and qualitative examples extend to one minute.

Critical strengths
  1. Clean conceptual reframing. Separating denoising completion from final video output, and repairing the internal conditioning history while displayed video stays immutable, is a genuinely useful idea; the n-versus-r bookkeeping is precise and consistently maintained.
  2. Mechanism matches the diagnosed failure mode. Prior forcing methods freeze a latent once its trajectory ends; retaining the completed latent for later-context repair targets exactly this gap, and the distinction from SDEdit-style re-noising and AutoRefiner is clearly articulated.
  3. Elegant training scheme without an extra evaluator. Self-sampled repair trials turn "would a repair help?" into a measurable reward that supervises keep-or-repair, while the same trial's endpoint loss trains the correction; A.6 correctly derives the weighted-CE minimizer as the improvement-weighted preference A/(A+B), not Pr(R>0|C).
  4. Controlled, mechanism-aligned ablation. Table 3 isolates the repair branch at matched context-development budget, with gains concentrated where the mechanism predicts (I2V-S/B, SC, BC up 0.75-1.08; MS flat) rather than diffuse gains.
Critical weaknesses
  1. Headline long-horizon claim rests on weak quantitative evidence. The core evidence is 10s vs 20s scores differing by ≤0.02, equally consistent with VBench metric saturation at the noise floor; the 5s→10s drop (SC -1.48, BC -1.00) shows drift occurs early. No 40s/60s quantitative results; the "162-/320-frame" and one-minute claims are qualitative only (single sequence in Fig. 1).
  2. Comparison is not compute-matched. Repair-on adds Review + repair forwards per candidate (authors acknowledge unequal model evaluations). With ~1-point gains, a matched-compute control (larger D with repair off, equal total forwards) is needed to attribute gains to repair rather than extra computation. No latency/throughput numbers — a first-class concern for real-time interactive simulation.
  3. Single long-horizon baseline. Causal Forcing and Mask Forcing appear only at 5s; at 10s/20s only Rolling Forcing is compared.
  4. Baseline confounds unaddressed. 14B teacher priors (Rolling/Causal Forcing), undescribed I2V adaptation (Mask Forcing), and Wan2.2 initialization are flagged in captions but never discussed as confounds between method and training recipe.
Questions for authors
  1. Frame counts: please reconcile the abstract's "81-frame training limit" and "162- and 320-frame" generalization with the evaluation protocol (81 frames at 24 FPS is about 3.4s; Fig. 1 reports 24 FPS outputs, so 5/10/20s correspond to 120/240/480 frames). Are 162/320 latent frames, RGB frames, or a different sampling rate?
  2. Can you report VBench-I2V (even on a prompt subset) at 40s and 60s, and can Causal Forcing and Mask Forcing be evaluated at 10s and 20s? These two results would most directly test the paper's central claim and could move my score.
  3. What does a compute-matched ablation give (repair-off with larger D, or repair-on at equal total forwards), and what are the measured latency and throughput versus the baselines?
  4. What is the actual r-n offset in the streaming experiments, and can repair at n<r create a visible discontinuity between the already-displayed frame n and subsequent frames conditioned on the repaired internal copy? Is there any measurement of such seams?
  5. What fraction of internal candidates trigger repair at inference on VBench prompts, and how does trigger behavior differ between the training distribution (local reconstructions) and free-running generation?
Reviewer_qnhKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes StreamCeption, a streaming video generation framework that addresses error accumulation by allowing fully denoised latents to be revised using subsequently generated context. Building on Flowception's variable-length flow matching with per-frame denoising and bidirectional attention, StreamCeption replaces gap-insertion with consecutive appending inside a bounded active window, and crucially decouples denoising completion from final frame output: when the oldest active latent reaches σ=0, it is temporarily retained rather than frozen. A learned binary repair head then decides whether to re-noise it (fixed strength 0.10) and refine it under the now-richer context. Training uses self-sampled repair trials where the reconstruction improvement R = MSE(y, x*) − MSE(x_rep, x*) supervises the keep-or-repair decision (weighted by |R|) while the same trial trains the repaired endpoint toward the clean target, jointly with flow matching. Instantiated via LoRA on Wan2.2-TI2V-5B and trained on 81-frame clips, it is evaluated on VBench-I2V at 5s/10s/20s against Rolling Forcing, Causal Forcing, and Mask Forcing.

Critical strengths
  1. A genuinely novel and well-motivated relaxation of a constraint every prior streaming method takes for granted. All existing samplers treat "denoising finished" as equivalent to "committed to history." StreamCeption's insight that these two events can be separated, with a retention period during which later context accumulates, is simple in hindsight and non-obvious. Importantly, the design respects the actual constraint of streaming (already-displayed frames stay fixed) while exploiting the freedom that remains (internal conditioning context may still be revised). This clean separation of "what must be immutable" from "what merely appears immutable" is the paper's real contribution and is conceptually transferable beyond this specific instantiation.

  2. An elegant training formulation where one trial supervises both selection and execution. The repair trial simultaneously yields (i) a label for the keep-or-repair head via sign(R), weighted by |R| so that large improvements and large degradations dominate, and (ii) a regression target for the repaired endpoint via MSE(x_rep, x*). Notably, harmful trials still contribute usefully — they supply a keep label and train recovery toward the clean target. Using the model's own detached reconstructions as repair candidates means the head is trained on the error distribution it will actually face, and the method requires no clean targets or error computation at inference. This is a well-designed objective.

Critical weaknesses
  1. The "selective" component is never validated against non-selective alternatives. The entire premise of the repair head is that blindly refining early latents "risks perturbing accurate content," motivating a learned decision. Yet the only ablation (Table 3) compares repair-on against repair-off. There is no always-repair baseline, no random-repair baseline at matched trigger rate, and no reported statistics on how often the head actually fires during generation. With Amax = 1, τ = 0.5, and τ_acc = 1.0, each candidate is repaired at most once and the decision is never revisited — so if the head fires on most candidates, the method is operationally close to always-repair and the learned classifier contributes little. Without the trigger rate and an always-repair comparison, the reader cannot determine whether the gains come from selectivity (the claimed contribution) or simply from having a repair step at all. Compounding this, the one ablation that does exist is not compute-matched, so part of its margin may be attributable to extra forwards rather than to repair itself.

  2. Dynamic Degree is omitted, and the reported metric profile is consistent with motion/detail suppression rather than genuine consistency gains. StreamCeption attains the best SC/BC/MS (97.91/97.40/99.36 at 5s) while scoring below every baseline on aesthetic quality and below three of four on imaging quality (61.89/71.27 vs. Wan2.2's 64.01 and Mask Forcing's 73.11). The paper correctly notes that Rolling Forcing's high IQ can coexist with declining BC/MS because IQ ignores temporal structure — but the symmetric caution applies to its own results: SC and BC measure inter-frame similarity and reward static or over-smoothed content, which is exactly what repeated re-noising-and-refinement under a consistency-oriented objective would tend to produce. Dynamic Degree is the standard VBench dimension for detecting this, it is reported by the baselines' own papers, and it is absent here. Similarly, the near-perfect 10s→20s stability (every score within 0.02 points) is presented as robustness, but such flatness is equally consistent with the generation having converged to a low-dynamics regime. Without Dynamic Degree, a motion-magnitude measurement, or any human preference study, the core claim of "stronger long-horizon consistency" cannot be separated from degenerate stability.

Questions for authors
  1. What fraction of repair candidates does the head actually select for repair, at 5s / 10s / 20s? And how does StreamCeption compare against an always-repair variant (τ = 0) and a random-repair variant matched to the learned trigger rate? This directly determines whether "selective" is doing any work.
  2. Can you report Dynamic Degree (and ideally a motion-magnitude statistic such as mean optical-flow norm) for all methods in Table 1? Given that SC/BC reward static content, this is necessary to establish that the consistency gains are not motion suppression.
  3. Table 1 evaluates Causal Forcing and Mask Forcing only at 5s, so all long-horizon comparisons rest on Rolling Forcing alone. Can the other baselines be run at 10s and 20s? Additionally, the baselines are marked as using a 14B teacher († ) while StreamCeption is LoRA-tuned on 5B — how should the reader account for this difference in training setup?
  4. The re-noising strength is fixed at a nominal 0.10 and Amax = 1. Was any sweep performed over these? A learned or adaptive strength seems like the natural extension of the paper's own argument that different candidates need different amounts of correction.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.