A brand new paper out this week at Arxiv addresses a problem which anybody who has adopted the Hunyuan Video or Wan 2.1 AI video turbines may have come throughout by now: temporal aberrations, the place the generative course of tends to abruptly pace up, conflate, omit, or in any other case mess up essential moments in a generated video:
Click on to play. A number of the temporal glitches which can be changing into acquainted to customers of the brand new wave of generative video methods, highlighted within the new paper. To the precise, the ameliorating impact of the brand new FluxFlow method. Supply: https://haroldchen19.github.io/FluxFlow/
The video above options excerpts from instance take a look at movies on the (be warned: quite chaotic) venture web site for the paper. We are able to see a number of more and more acquainted points being remediated by the authors’ technique (pictured on the precise within the video), which is successfully a dataset preprocessing approach relevant to any generative video structure.
Within the first instance, that includes ‘two kids taking part in with a ball’, generated by CogVideoX, we see (on the left within the compilation video above and within the particular instance beneath) that the native era quickly jumps via a number of important micro-movements, dashing the kids’s exercise as much as a ‘cartoon’ pitch. Against this, the identical dataset and technique yield higher outcomes with the brand new preprocessing approach, dubbed FluxFlow (to the precise of the picture in video beneath):
Click on to play.
Within the second instance (utilizing NOVA-0.6B) we see {that a} central movement involving a cat has not directly been corrupted or considerably under-sampled on the coaching stage, to the purpose that the generative system turns into ‘paralyzed’ and is unable to make the topic transfer:
Click on to play.
This syndrome, the place the movement or topic will get ‘caught’, is among the most frequently-reported bugbears of HV and Wan, within the numerous picture and video synthesis teams.
A few of these issues are associated to video captioning points within the supply dataset, which we took a take a look at this week; however the authors of the brand new work focus their efforts on the temporal qualities of the coaching information as a substitute, and make a convincing argument that addressing the challenges from that perspective can yield helpful outcomes.
As talked about within the earlier article about video captioning, sure sports activities are notably tough to distil into key moments, which means that crucial occasions (similar to a slam-dunk) don’t get the eye they want at coaching time:
Click on to play.
Within the above instance, the generative system doesn’t know get to the following stage of motion, and transits illogically from one pose to the following, altering the perspective and geometry of the participant within the course of.
These are massive actions that acquired misplaced in coaching – however equally susceptible are far smaller however pivotal actions, such because the flapping of a butterfly’s wings:
Click on to play.
Not like the slam-dunk, the flapping of the wings just isn’t a ‘uncommon’ however quite a persistent and monotonous occasion. Nonetheless, its consistency is misplaced within the sampling course of, because the motion is so fast that it is rather tough to determine temporally.
These will not be notably new points, however they’re receiving larger consideration now that highly effective generative video fashions can be found to fanatics for native set up and free era.
The communities at Reddit and Discord have initially handled these points as ‘user-related’. That is an comprehensible presumption, because the methods in query are very new and minimally documented. Subsequently numerous pundits have prompt various (and never at all times efficient) cures for a number of the glitches documented right here, similar to altering the settings in numerous parts of various varieties of ComfyUI workflows for Hunyuan Video (HV) and Wan 2.1.
In some instances, quite than producing fast movement, each HV and Wan will produce gradual movement. Ideas from Reddit and ChatGPT (which principally leverages Reddit) embrace altering the variety of frames within the requested era, or radically reducing the body price*.
That is all determined stuff; the rising fact is that we do not but know the precise trigger or the precise treatment for these points; clearly, tormenting the era settings to work round them (notably when this degrades output high quality, as an illustration with a too-low fps price) is just a short-stop, and it is good to see that the analysis scene is addressing rising points this rapidly.
So, apart from this week’s take a look at how captioning impacts coaching, let’s check out the brand new paper about temporal regularization, and what enhancements it would supply the present generative video scene.
The central thought is quite easy and slight, and none the more serious for that; nonetheless the paper is considerably padded with a view to attain the prescribed eight pages, and we are going to skip over this padding as obligatory.
The fish within the native era of the VideoCrafter framework is static, whereas the FluxFlow-altered model captures the requisite modifications. Supply: https://arxiv.org/pdf/2503.15417
The new work is titled Temporal Regularization Makes Your Video Generator Stronger, and comes from eight researchers throughout Everlyn AI, Hong Kong College of Science and Expertise (HKUST), the College of Central Florida (UCF), and The College of Hong Kong (HKU).
(on the time of writing, there are some points with the paper’s accompanying venture web site)
FluxFlow
The central thought behind FluxFlow, the authors’ new pre-training schema, is to beat the widespread issues flickering and temporal inconsistency by shuffling blocks and teams of blocks within the temporal body orders because the supply information is uncovered to the coaching course of:
The central thought behind FluxFlow is to maneuver blocks and teams of blocks into sudden and non-temporal positions, as a type of information augmentation.
The paper explains:
‘[Artifacts] stem from a elementary limitation: regardless of leveraging large-scale datasets, present fashions typically depend on simplified temporal patterns within the coaching information (e.g., mounted strolling instructions or repetitive body transitions) quite than studying various and believable temporal dynamics.
‘This difficulty is additional exacerbated by the shortage of express temporal augmentation throughout coaching, leaving fashions vulnerable to overfitting to spurious temporal correlations (e.g., “body #5 should comply with #4”) quite than generalizing throughout various movement eventualities.’
Most video era fashions, the authors clarify, nonetheless borrow too closely from picture synthesis, specializing in spatial constancy whereas largely ignoring the temporal axis. Although strategies similar to cropping, flipping, and shade jittering have helped enhance static picture high quality, they aren’t satisfactory options when utilized to movies, the place the phantasm of movement relies on constant transitions throughout frames.
The ensuing issues embrace flickering textures, jarring cuts between frames, and repetitive or overly simplistic movement patterns.
Click on to play.
The paper argues that although some fashions – together with Steady Video Diffusion and LlamaGen – compensate with more and more advanced architectures or engineered constraints, these come at a value when it comes to compute and suppleness.
Since temporal information augmentation has already confirmed helpful in video understanding duties (in frameworks similar to FineCliper, SeFAR and SVFormer) it’s stunning, the authors assert, that this tactic is never utilized in a generative context.
Disruptive Conduct
The researchers contend that straightforward, structured disruptions in temporal order throughout coaching assist fashions generalize higher to life like, various movement:
‘By coaching on disordered sequences, the generator learns to get well believable trajectories, successfully regularizing temporal entropy. FLUXFLOW bridges the hole between discriminative and generative temporal augmentation, providing a plug-and-play enhancement answer for temporally believable video era whereas bettering general [quality].
‘Not like present strategies that introduce architectural modifications or depend on post-processing, FLUXFLOW operates instantly on the information degree, introducing managed temporal perturbations throughout coaching.’
Click on to play.
Body-level perturbations, the authors state, introduce fine-grained disruptions inside a sequence. This type of disruption just isn’t dissimilar to masking augmentation, the place sections of knowledge are randomly blocked out, to stop the system overfitting on information factors, and inspiring higher generalization.
Checks
Although the central thought right here does not run to a full-length paper, on account of its simplicity, nonetheless there’s a take a look at part that we will check out.
The authors examined for 4 queries referring to improved temporal high quality whereas sustaining spatial constancy; skill to study movement/optical movement dynamics; sustaining temporal high quality in extraterm era; and sensitivity to key hyperparameters.
The researchers utilized FluxFlow to a few generative architectures: U-Internet-based, within the type of VideoCrafter2; DiT-based, within the type of CogVideoX-2B; and AR-based, within the type of NOVA-0.6B.
For truthful comparability, they fine-tuned the architectures’ base fashions with FluxFlow as an extra coaching section, for one epoch, on the OpenVidHD-0.4M dataset.
The fashions have been evaluated towards two fashionable benchmarks: UCF-101; and VBench.
For UCF, the Fréchet Video Distance (FVD) and Inception Rating (IS) metrics have been used. For VBench, the researchers focused on temporal high quality, frame-wise high quality, and general high quality.
Quantitative preliminary Analysis of FluxFlow-Body. “+ Authentic” signifies coaching with out FLUXFLOW, whereas “+ Num × 1” exhibits totally different FluxFlow-Body configurations. Finest outcomes are shaded; second-best are underlined for every mannequin.
Commenting on these outcomes, the authors state:
‘Each FLUXFLOW-FRAME and FLUXFLOW-BLOCK considerably enhance temporal high quality, as evidenced by the metrics in Tabs. 1, 2 (i.e., FVD, Topic, Flicker, Movement, and Dynamic) and qualitative leads to [image below].
‘For example, the movement of the drifting automotive in VC2, the cat chasing its tail in NOVA, and the surfer driving a wave in CVX change into noticeably extra fluid with FLUXFLOW. Importantly, these temporal enhancements are achieved with out sacrificing spatial constancy, as evidenced by the sharp particulars of water splashes, smoke trails, and wave textures, together with spatial and general constancy metrics.’
Under we see picks from the qualitative outcomes the authors discuss with (please see the unique paper for full outcomes and higher decision):
Alternatives from the qualitative outcomes.
The paper means that whereas each frame-level and block-level perturbations improve temporal high quality, frame-level strategies are likely to carry out higher. That is attributed to their finer granularity, which allows extra exact temporal changes. Block-level perturbations, against this, might introduce noise on account of tightly coupled spatial and temporal patterns inside blocks, decreasing their effectiveness.
Conclusion
This paper, together with the Bytedance-Tsinghua captioning collaboration launched this week, has made it clear to me that the obvious shortcomings within the new era of generative video fashions might not end result from person error, institutional missteps, or funding limitations, however quite from a analysis focus that has understandably prioritized extra pressing challenges, similar to temporal coherence and consistency, over these lesser issues.
Till just lately, the outcomes from freely-available and downloadable generative video methods have been so compromised that no nice locus of effort emerged from the fanatic neighborhood to redress the problems (not least as a result of the problems have been elementary and never trivially solvable).
Now that we’re a lot nearer to the long-predicted age of purely AI-generated photorealistic video output, it is clear that each the analysis and informal communities are taking a deeper and extra productive curiosity in resolving remaining points; hopefully, these will not be intractable obstacles.
* Wan’s native body price is a paltry 16fps, and in response to my very own points, I notice that boards have prompt reducing the body price as little as 12fps, after which utilizing FlowFrames or different AI-based re-flowing methods to interpolate the gaps between such a sparse variety of frames.
First printed Friday, March 21, 2025
