Why ByteDance Paused Seedance 2.0 (And Which AI Video Model Actually Wins in March 2026)

On March 14, 2026, ByteDance quietly pulled the global rollout of Seedance 2.0. No press release. No blog post. Just a brief notice on the developer portal: "Global availability postponed until further notice."

Within 48 hours, the AI video community had three competing theories: quality regression at scale, regulatory pressure from the EU AI Act, and unresolved deepfake safety concerns. The truth is likely all three.

But here is the thing most coverage missed: Seedance 2.0's pause did not happen in a vacuum. It happened during the most competitive 90 days in AI video history. Kling 2 launched its "Master" tier. Sora 2 dropped its API. Hailuo 02 started generating 60-second clips that rival Hollywood B-roll. And Runway Gen-3 Alpha quietly became the workhorse that nobody talks about.

This article breaks down the entire March 2026 landscape. Every major model, compared head-to-head, with real data on quality, pricing, speed, and features. If you generate AI video for any reason, this is the only comparison you need right now.

Breaking News: Why ByteDance Paused Seedance 2.0

Let's start with what we know, what we can reasonably infer, and what remains speculation.

What ByteDance Has Confirmed

ByteDance's official statement was minimal: Seedance 2.0 would remain available in the Chinese domestic market through Jimeng (the Chinese-facing product) but global API access and the international web app would be paused "to ensure quality and safety standards meet regional requirements."

The Quality Regression Problem

Early testers of Seedance 2.0's global beta reported a pattern: the model performed brilliantly on carefully chosen demo prompts but degraded on edge cases. Specifically:

Character consistency across scenes dropped sharply after the first 5 seconds of generation
Text rendering in video (a headline feature of 2.0) produced garbled or mirrored text roughly 40% of the time
Physics simulation for fluid dynamics and cloth worked well in isolation but broke down in multi-subject scenes
Temporal coherence in clips longer than 8 seconds showed visible "drift" -- subjects slowly morphing or environments shifting

These are not trivial bugs. They suggest Seedance 2.0 was trained on a distribution that works exceptionally well for short, single-subject clips (the demos ByteDance showcased) but struggles to generalize.

The Regulatory Factor

The EU AI Act's provisions on synthetic media came into full enforcement in February 2026. Any AI video tool available to EU users must now implement:

Mandatory C2PA metadata in all generated content
Real-time watermarking that survives compression and re-encoding
A documented model card describing training data provenance
Rate limiting on face-generation features

ByteDance had reportedly implemented watermarking but had not published the required model card or integrated C2PA metadata at the standards required. Launching without compliance would have exposed them to fines of up to 3% of global annual revenue.

The Safety Dimension

Seedance 2.0's most impressive feature -- photorealistic human generation with controllable expressions and lip sync -- is also its most dangerous. Internal red-teaming reportedly surfaced scenarios where the model could generate convincing deepfakes with minimal prompt engineering. ByteDance's safety filters caught most cases in testing, but the gap between "most" and "all" is where reputational and legal risk lives.

What This Means for Users

If you were planning to build workflows around Seedance 2.0, pause. The domestic Chinese version (via Jimeng) remains available but requires a Chinese phone number and has different usage terms. The global relaunch timeline is genuinely uncertain. ByteDance has not committed to a date.

In the meantime, every other model in this comparison is available today.

The March 2026 Video AI Landscape: Complete Reshuffling in 90 Days

The last quarter has been extraordinary. Here is a timeline of what changed:

January 2026: Kling 2.0 launches globally with "Standard" and "Master" quality tiers. Hailuo 02 enters open beta.
February 2026: Sora 2 opens its API to all developers. Google quietly updates Veo 3 with improved motion. Runway Gen-3 Alpha ships a major update to camera controls.
March 2026: ByteDance pauses Seedance 2.0. Minimax (Hailuo) raises $600M at a $6B valuation. Luma Dream Machine 2.0 launches with real-time preview.

The result: there are now six credible, production-quality AI video models. Choosing between them requires understanding their specific strengths.

Head-to-Head Comparison Table

Here is the complete picture as of March 23, 2026:

Feature	Seedance 2.0	Kling 2 Master	Sora 2	Runway Gen-3 Alpha	Hailuo 02	Veo 3
Max Resolution	1080p (4K claimed)	1080p	1080p	1080p	1080p	4K
Max Duration	16s	10s	20s	10s	60s	8s
Text-to-Video	Yes	Yes	Yes	Yes	Yes	Yes
Image-to-Video	Yes	Yes	Yes	Yes	Yes	Yes
Video-to-Video	Yes	Yes	Limited	Yes	No	No
Camera Controls	Advanced	Advanced	Basic	Advanced	Basic	Moderate
Character Consistency	Excellent (short clips)	Very Good	Good	Good	Very Good	Good
Physics Accuracy	Good	Very Good	Very Good	Good	Good	Excellent
Text in Video	Moderate	Poor	Moderate	Poor	Poor	Good
API Available	Paused (Global)	Yes	Yes	Yes	Yes	Yes (Waitlist)
Global Availability	Paused	Yes	Yes	Yes	Yes	Limited
Watermarking	Yes	Yes	Yes	Yes	Yes	Yes

Key takeaway: no single model dominates every category. The best choice depends entirely on your use case, which we will cover in detail below.

Quality Comparison: What Actually Matters

Quality in AI video is not one dimension. It is at least four, and they trade off against each other.

Resolution and Visual Fidelity

Winner: Veo 3

Google's Veo 3 is the only model delivering true 4K output with consistent detail. The catch: generation times are long (more on that below), and the maximum duration is only 8 seconds. For short, high-fidelity clips, nothing matches it.

Seedance 2.0 claimed 4K but delivered closer to upscaled 1080p in testing. Kling 2 Master, Sora 2, and Runway Gen-3 Alpha all produce genuine 1080p. Hailuo 02 outputs 1080p but with slightly more compression artifacts than competitors.

Practical ranking for visual fidelity:

Veo 3 (true 4K, limited duration)
Kling 2 Master (clean 1080p, excellent detail)
Sora 2 (clean 1080p, good detail)
Runway Gen-3 Alpha (clean 1080p, occasional softness)
Seedance 2.0 (upscaled 1080p, good but inconsistent)
Hailuo 02 (1080p with artifacts, but 60s duration)

Motion Coherence

Winner: Sora 2

Motion coherence means objects move in physically plausible ways without warping, disappearing, or suddenly changing shape. Sora 2 leads here. OpenAI's training pipeline produces motion that "feels right" even when the subject is complex -- a person walking through a crowded market, a bird landing on a branch, a car turning a corner.

Kling 2 Master is close behind. It handles complex multi-subject motion better than Sora 2 in some cases, particularly for scenes with many interacting elements.

Hailuo 02 deserves special mention. Its motion coherence degrades less over time than any other model. When generating 30-60 second clips, other models would show severe drift, while Hailuo 02 maintains reasonable consistency throughout.

Practical ranking for motion coherence:

Sora 2 (best short-form motion)
Kling 2 Master (best multi-subject motion)
Hailuo 02 (best long-form motion)
Veo 3 (good but limited duration)
Runway Gen-3 Alpha (good, occasional stiffness)
Seedance 2.0 (good under 5s, degrades after)

Physics Accuracy

Winner: Veo 3

Physics accuracy covers gravity, fluid dynamics, cloth simulation, reflections, shadows, and light transport. Veo 3 leads, likely because Google's training data and simulation pipelines are more extensive.

Notable physics behaviors by model:

Veo 3: Best water, reflections, and shadows. Cloth drapes realistically. Light bounces correctly.
Sora 2: Good gravity and momentum. Water is decent. Shadows sometimes lag behind movement.
Kling 2 Master: Strong on rigid body physics. Handles collisions and object interactions well. Fluid dynamics are average.
Runway Gen-3 Alpha: Good ambient physics. Struggles with fast motion and impacts.
Hailuo 02: Acceptable physics for most scenes. Noticeable errors in complex interactions.
Seedance 2.0: Good single-subject physics. Breaks down with multiple interacting objects.

Character Consistency

Winner: Kling 2 Master (with character reference images)

Character consistency -- maintaining a character's appearance across cuts, angles, and movements -- is the hardest problem in AI video. Kling 2 Master handles this best, particularly when you provide reference images.

Seedance 2.0 was designed to excel here, and its demos showed remarkable consistency. In practice, the consistency held for clips under 5 seconds but degraded noticeably in longer generations, which contributed to the global pause.

Hailuo 02 is surprisingly strong on character consistency in long-form clips. Because it can generate up to 60 seconds, maintaining character appearance across that duration is genuinely impressive.

Practical ranking for character consistency:

Kling 2 Master (best with reference images)
Hailuo 02 (best over long durations)
Seedance 2.0 (excellent under 5s, degrades after)
Sora 2 (good, occasional subtle changes)
Veo 3 (good within 8s limit)
Runway Gen-3 Alpha (acceptable, frequent hair/clothing changes)

Pricing Comparison: Cost Per Second of Generated Video

Pricing in AI video is confusing. Some models charge per generation, some per second, some via subscription tiers. Here is a normalized comparison based on API pricing as of March 2026, calculated as the cost for one second of generated video at each model's highest quality setting.

Model	Cost Per Second (Highest Quality)	Cost Per Second (Standard Quality)	Subscription Option
Seedance 2.0	$0.12 (domestic pricing)	$0.06	Jimeng subscription (China only)
Kling 2 Master	$0.10	$0.04	$30/mo for 600 credits
Sora 2	$0.15	$0.08	ChatGPT Plus ($20/mo, limited)
Runway Gen-3 Alpha	$0.08	$0.05	$12/mo for 625 credits
Hailuo 02	$0.05	$0.03	Free tier available
Veo 3	$0.20	$0.10	Vertex AI pricing

Key pricing insights:

Cheapest per second: Hailuo 02 at $0.03-0.05/second. For bulk content generation, nothing comes close.
Best value for quality: Kling 2 Master at $0.10/second delivers premium quality at a reasonable price.
Most expensive: Veo 3 at $0.20/second. You pay a premium for 4K and superior physics.
Best subscription deal: Runway Gen-3 Alpha's $12/month plan offers the lowest barrier to entry for occasional users.
Hidden costs: Sora 2's ChatGPT Plus subscription includes limited video generation, but the limits are tight. Heavy users will need API access at $0.15/second.

Cost for Common Use Cases

To make this more practical, here is what it costs to generate common content types:

Use Case	Duration Needed	Cheapest Option	Cost	Best Quality Option	Cost
Social media clip	5s	Hailuo 02	$0.15	Kling 2 Master	$0.50
Product demo	10s	Hailuo 02	$0.30	Sora 2	$1.50
Ad creative (3 variants)	3 x 5s	Hailuo 02	$0.45	Kling 2 Master	$1.50
Music video scene	15s	Hailuo 02	$0.75	Sora 2	$2.25
Long-form B-roll	60s	Hailuo 02	$1.80	Hailuo 02	$3.00

Speed Comparison: Generation Time

Speed matters when you are iterating on prompts or generating content at scale. Generation times vary significantly:

Model	5-Second Clip	10-Second Clip	Max Duration Clip
Seedance 2.0	~45s	~90s	~180s (16s clip)
Kling 2 Master	~60s	~120s	~120s (10s clip)
Kling 2 Standard	~20s	~40s	~40s (10s clip)
Sora 2	~90s	~180s	~360s (20s clip)
Runway Gen-3 Alpha	~30s	~60s	~60s (10s clip)
Hailuo 02	~40s	~80s	~480s (60s clip)
Veo 3	~120s	N/A	~200s (8s clip)

Speed takeaways:

Fastest for iteration: Kling 2 Standard mode generates a 5-second clip in about 20 seconds. Use this for prompt testing, then switch to Master for final output.
Fastest high-quality: Runway Gen-3 Alpha delivers solid quality at ~30 seconds for a 5-second clip.
Slowest: Veo 3. The 4K quality comes at a real time cost. Plan for 2+ minutes per clip.
Sora 2 is slow: OpenAI's model consistently takes longer than competitors. The quality is high, but the iteration cycle is painful.
Hailuo 02 scales linearly: Its 60-second clips take about 8 minutes, which is reasonable given the output length.

Feature Comparison: Beyond Text-to-Video

Raw text-to-video is table stakes. The differentiation now happens in advanced features.

Image-to-Video

All six models support image-to-video, but the quality varies:

Kling 2 Master: Best image-to-video implementation. Maintains source image fidelity while adding natural motion. Supports pose-guided animation.
Runway Gen-3 Alpha: Strong image-to-video with the best camera control integration. You can specify exactly how the camera should move relative to the source image.
Hailuo 02: Good image-to-video that excels at extending the motion for long durations. The initial frames closely match the source.
Sora 2: Solid but sometimes "reinterprets" the source image, making subtle changes to composition or color.
Veo 3: High fidelity to source image but limited motion range in the 8-second window.
Seedance 2.0: Reportedly excellent but currently unavailable globally.

Video-to-Video

Video-to-video -- transforming existing footage into a new style or enhancing it -- is less universally supported:

Runway Gen-3 Alpha: The strongest video-to-video capabilities. Style transfer, motion retargeting, and video enhancement all work well.
Kling 2: Supports video-to-video with style transfer. Quality is good but options are more limited than Runway.
Seedance 2.0: Strong video-to-video in demos, including style transfer and motion editing. Unavailable globally.
Sora 2: Limited video-to-video. Can extend existing clips and apply basic style changes, but not full transformation.
Hailuo 02 and Veo 3: No video-to-video support as of March 2026.

Camera Controls

Camera controls let you specify movement: pan, tilt, zoom, dolly, orbit, and more.

Camera Feature	Kling 2	Runway Gen-3 Alpha	Sora 2	Hailuo 02	Veo 3
Pan	Yes	Yes	Yes	Yes	Yes
Tilt	Yes	Yes	Yes	Yes	Yes
Zoom	Yes	Yes	Yes	Yes	Yes
Dolly	Yes	Yes	No	No	Yes
Orbit	Yes	Yes	No	No	No
Crane	Yes	Yes	No	No	No
Custom Path	Yes	Yes	No	No	No
Speed Control	Yes	Yes	Limited	No	No

Camera control winner: Tied between Kling 2 and Runway Gen-3 Alpha. Both offer full camera path specification. If camera work is central to your workflow, these are your only real options.

Lip Sync and Audio Integration

A newer category that is becoming critical for marketing and social content:

Kling 2: Supports audio-driven lip sync. Provide an audio track and the model syncs character mouth movements. Quality is good for front-facing subjects, weaker at angles.
Seedance 2.0: Reportedly the best lip sync of any model, a key feature of the 2.0 release. Unavailable globally.
Sora 2: No native lip sync. Requires external tools.
Runway Gen-3 Alpha: Experimental lip sync in beta. Not production-ready.
Hailuo 02 and Veo 3: No lip sync support.

Best Model for Each Use Case

Based on our testing, here are the specific recommendations by use case.

Marketing and Advertising

Primary recommendation: Kling 2 Master

Marketing needs consistency, brand control, and iteration speed. Kling 2 Master delivers all three. Its character consistency with reference images means your brand spokesperson looks the same across every ad. Camera controls let you create professional-looking product shots. And Kling 2 Standard mode lets you iterate quickly on concepts before committing to the expensive Master renders.

Runner-up: Runway Gen-3 Alpha -- slightly cheaper and faster, with strong camera controls. Good for teams on tighter budgets.

Independent Film and Short Films

Primary recommendation: Sora 2

For narrative content where motion quality and physical plausibility matter most, Sora 2 produces the most "cinematic" output. Its motion has a naturalism that other models lack. The longer maximum duration (20 seconds) gives you more usable footage per generation.

Runner-up: Veo 3 -- for shots where visual fidelity matters more than duration. Its 4K output looks genuinely cinematic.

Social Media Content

Primary recommendation: Hailuo 02

Social media content is a volume game. You need lots of clips, fast, cheap. Hailuo 02's combination of low cost ($0.03-0.05/second), decent quality, and 60-second maximum duration makes it ideal. Generate 10 variants for the price of one Sora 2 clip.

Runner-up: Kling 2 Standard -- when you need higher quality for important posts but still want reasonable speed and cost.

Product Demonstrations

Primary recommendation: Kling 2 Master (image-to-video mode)

Start with a product photo. Use Kling 2's image-to-video with camera controls to create a rotating product shot, a zoom-in on details, or an unboxing animation. The fidelity to the source image ensures your product looks exactly right.

Runner-up: Runway Gen-3 Alpha -- its video-to-video mode can transform existing product footage into stylized demos.

Music Videos

Primary recommendation: Runway Gen-3 Alpha

Music videos benefit from stylization, and Runway's video-to-video capabilities excel here. Film basic footage and transform it through AI style transfer. Combine with text-to-video for surreal interludes. The camera controls enable dynamic visual sequences that match musical beats.

Runner-up: Hailuo 02 -- when you need long continuous sequences. Its 60-second generation can create extended visual passages that maintain consistency.

Architectural and Real Estate Visualization

Primary recommendation: Veo 3

Architectural visualization demands physics accuracy and visual fidelity. Veo 3's superior lighting, reflections, and 4K resolution make it the clear choice. The 8-second limit is less problematic here since you typically need short walkthrough clips of individual spaces.

Runner-up: Kling 2 Master -- for exterior shots and longer sequences where 4K is less critical than duration.

The Chinese AI Video Race: Global Ambitions, Local Realities

Three of the six models in this comparison come from Chinese companies: Seedance (ByteDance), Kling (Kuaishou), and Hailuo (Minimax). Understanding the dynamics between them explains a lot about where AI video is heading.

ByteDance and Seedance

ByteDance has the deepest pockets and the most compute. TikTok gives them unparalleled video data for training. Seedance 2.0 was supposed to be the model that proved Chinese AI could lead, not just compete, in generative video.

The global pause is a setback, but ByteDance is not going away. Their domestic version continues to improve, and internal roadmaps reportedly target a Q2 2026 global relaunch with full EU AI Act compliance. When Seedance 2.0 comes back, it will likely be the most capable model on the market.

Kuaishou and Kling

Kuaishou's strategy is different from ByteDance's. While ByteDance aims for the most impressive demos, Kuaishou focused on reliability. Kling 2 may not win benchmark comparisons on cherry-picked examples, but it consistently produces usable output. Their API documentation is excellent. Their pricing is transparent. They launched globally without major incidents.

This reliability-first approach has paid off. Kling 2 is currently the most-used AI video model by commercial users outside China, based on API call volume estimates from independent trackers.

Minimax and Hailuo

Minimax is the wildcard. Their $6B valuation, fueled by a $600M raise in March 2026, signals massive investor confidence. Hailuo 02's killer feature -- 60-second generation -- solves a real problem that no competitor has addressed.

The trade-off is quality. Hailuo 02's output is noticeably less polished than Kling 2 Master or Sora 2. But for many use cases, "good enough for 60 seconds" beats "excellent for 5 seconds."

Minimax's roadmap reportedly prioritizes "Hailuo 03" for Q3 2026, targeting Kling 2 Master quality at current Hailuo 02 prices. If they deliver, the economics of AI video generation change dramatically.

The Geopolitical Dimension

US-China tensions continue to affect the AI video landscape. Key developments:

Export controls on advanced GPUs have not stopped Chinese model development but have likely slowed training cycles and forced architectural innovations.
Data sovereignty requirements in multiple jurisdictions mean Chinese models must maintain separate infrastructure for global users. This adds cost and complexity.
Content moderation standards differ significantly between markets. All three Chinese models apply stricter content filters for domestic use and different filters for global markets.

For users, the practical impact is this: Chinese models may occasionally become unavailable in certain markets due to regulatory changes. Diversifying across providers is prudent.

How to Choose: Decision Framework

Rather than a flowchart (which does not render well in text), here is a systematic decision process:

Step 1: What is your budget?

Under $50/month: Hailuo 02 (free tier + cheap API) or Runway Gen-3 Alpha ($12/month subscription)
$50-200/month: Kling 2 Master with Standard for iteration
$200+/month or enterprise: Choose based on quality needs below

Step 2: What duration do you need?

Under 5 seconds: Any model works. Choose on quality and price.
5-10 seconds: Kling 2 Master, Sora 2, or Runway Gen-3 Alpha.
10-20 seconds: Sora 2 (only model with high quality at this length).
20-60 seconds: Hailuo 02 (only option for this duration).

Step 3: What quality dimension matters most?

Visual fidelity and resolution: Veo 3
Motion naturalness: Sora 2
Character consistency: Kling 2 Master
Physics accuracy: Veo 3
Camera control: Kling 2 or Runway Gen-3 Alpha

Step 4: What features do you need?

Video-to-video: Runway Gen-3 Alpha
Image-to-video: Kling 2 Master
Lip sync: Kling 2 (or wait for Seedance 2.0 global return)
API integration: Kling 2 or Hailuo 02 (best documentation and reliability)

Step 5: How risk-tolerant are you?

Low risk tolerance (enterprise, agency): Kling 2 Master or Runway Gen-3 Alpha. Both have stable APIs, predictable pricing, and good support.
Medium risk tolerance (startup, creator): Sora 2 or Hailuo 02. Good products with some API instability or occasional quality variation.
High risk tolerance (experimenter, researcher): Veo 3 (limited availability) or wait for Seedance 2.0's return.

What Is Coming Next: The Race to Real-Time Video Generation

The current generation of models takes 20 seconds to 8 minutes to generate a clip. The next frontier is real-time generation -- video created as fast as it is consumed.

Luma Dream Machine 2.0

Luma's February 2026 launch introduced "real-time preview" -- a low-resolution preview that generates in under 2 seconds, showing you approximately what the final output will look like. The full render still takes 30-60 seconds, but the preview loop dramatically speeds up iteration.

This is not true real-time generation, but it is the closest any commercial product has come. Expect every competitor to ship something similar by mid-2026.

NVIDIA's Cosmos Platform

NVIDIA announced Cosmos at GTC 2026, a platform for real-time video generation targeting robotics and autonomous vehicles. The initial focus is simulation, not creative content, but the underlying technology could reshape AI video within 12-18 months.

Key claim: Cosmos can generate 720p video at 12 frames per second in real time on an H200 GPU cluster. If this scales to consumer hardware (even at lower quality), it changes everything.

The Architecture Shift: From Diffusion to Flow Matching

Most current models use diffusion-based architectures. The emerging shift toward flow matching (used in parts of Seedance 2.0 and reportedly in Kling 3 development) promises faster generation with fewer denoising steps. Early research suggests 3-5x speed improvements are achievable without quality loss.

This is the technical change most likely to enable "real enough" real-time generation by late 2026 or early 2027.

What Real-Time Means for Users

When real-time video generation arrives, it will change use cases fundamentally:

Live content creation: Generate video during live streams, presentations, or meetings
Interactive media: Games and experiences that generate scenes on the fly
Conversational video: AI agents that respond with generated video in real time
Infinite content: Social media feeds with fully personalized generated content

We are not there yet. But the gap between "minutes per clip" and "frames per second" is closing faster than most people expect.

The Bottom Line: March 2026 Rankings

After testing all available models extensively, here are our overall rankings for March 2026:

Best Overall: Kling 2 Master

Kling 2 Master wins on the combination of quality, reliability, features, and pricing. It does not lead any single category outright, but it is top-three in almost every dimension. For most users generating AI video for real-world applications, this is the model to start with.

Best Value: Hailuo 02

At $0.03-0.05 per second with a free tier, Hailuo 02 makes AI video generation accessible to everyone. The quality is not best-in-class, but it is good enough for social media, prototyping, and high-volume content. The 60-second duration is unmatched.

Best Quality (Short Form): Veo 3

If you need the absolute best-looking 5-8 second clips and are willing to pay for them, Veo 3 delivers. The 4K output with accurate physics and lighting is genuinely stunning. Limited by duration, speed, and availability.

Best for Creative Work: Sora 2

Sora 2's motion quality gives it an edge for narrative and artistic content. When you need video that "feels" cinematic, Sora 2's output has an organic quality that other models lack. The slow generation speed is the main drawback.

Best for Production Workflows: Runway Gen-3 Alpha

Runway has the most mature ecosystem for professional video production. Camera controls, video-to-video, style transfer, and integration with existing editing tools make it the practical choice for studios and agencies that need AI video as part of a larger pipeline.

Most Anticipated Return: Seedance 2.0

ByteDance's model showed genuine leaps in its demos. If the global relaunch addresses the quality, safety, and compliance issues, Seedance 2.0 could take the top spot. But until it is available and proven at scale, it is a promise rather than a product.

The AI video landscape will look different again in 90 days. Models that are leading now will face new challengers. Pricing will drop. Quality will improve. But as of March 2026, you have more viable options than ever, and the practical gap between "AI generated" and "professionally shot" footage continues to narrow.

Choose the model that matches your specific needs, budget, and risk tolerance. Test with your actual use cases, not generic prompts. And keep an eye on Seedance 2.0's return -- when ByteDance comes back, the entire competitive landscape will shift again.