LTX 2.3 Pro's Extend Mode is the first AI video tool to condition on multiple prior frames at the architecture level, producing seamless video continuations that hold lighting, motion physics, and scene geometry consistent across every extension. This article breaks down how it works and what sets it apart from every other approach currently available.
Most AI video tools treat extension as an afterthought. You generate a clip, it ends, and if you need more footage, you regenerate from scratch, accept the visual drift, or paste two clips together and hope the cut is not too obvious. LTX 2.3 Pro solves this at the architecture level, not the post-processing level, and that distinction changes everything about what you can actually produce with it.
Extend Mode is the headline capability in LTX 2.3 Pro, Lightricks' flagship latent video diffusion model available on PicassoIA. It lets you feed the model the last frames of an existing clip and prompt it to produce the next logical sequence, with the lighting, motion physics, and scene geometry held consistent across the join. No hard cut. No color grade mismatch. No temporal artifacts creeping in at the boundary. It is a categorically different approach to video continuation than anything else available right now.
The simplest way to frame it: Extend Mode does not start from noise. It starts from your clip.
Standard video generation models sample from a random latent noise vector and decode that toward a target prompt. Each new generation is statistically independent from the previous one. This is fine when you want a standalone clip, but catastrophic when you need continuity. The model has no structural memory of what your original scene looked like, how the shadows were falling, what trajectory the camera was on, or what direction objects were moving.
LTX 2.3 Pro's Extend Mode conditions the generation process on the encoded latent representation of your existing clip's final frames. Those frames are encoded into the model's latent space and become part of the diffusion process's starting constraints for the extension. The denoising is no longer operating in a vacuum. It is operating in direct relationship to what actually happened before it.
💡 In practical terms: if your original clip has a woman walking through a park with dappled afternoon light falling from the upper left, the extension will continue with that same woman, that same park, and that same light direction. The physics of the scene persist because the model was conditioned on them, not asked to guess at them.
This is not a small distinction. It is the whole game.
The phrase "temporal coherence" gets thrown around loosely in AI video discussions. What it actually means in the context of LTX 2.3 Pro is worth unpacking precisely, because the mechanism is what makes the output feel different from every I2V alternative.
Temporal coherence refers to the consistency of visual information across frames in a video sequence. At the pixel level, this means objects maintain their shape, color, and position relative to what the physics of motion would predict. At the semantic level, it means the scene reads as a continuous experience rather than a series of loosely related frames.
LTX 2.3 Pro achieves this through causal attention mechanisms in its transformer backbone. When generating the extension, the attention heads can look backward into the prior frames' latent representations. This is architecturally different from simply using the last frame as an image prompt. A single image prompt gives the model a snapshot, a reference without velocity or direction. Causal attention over multiple prior frames gives the model velocity vectors, lighting gradient direction across time, camera trajectory data, and object position trajectories simultaneously.
The result is motion prediction that feels physically grounded rather than hallucinated. Water that was flowing left continues flowing left. A person mid-stride continues the stride without resetting to a neutral pose. A panning shot continues panning at the same angular speed with the same horizon line. These details are what separate footage that holds up in a professional context from footage that is "good enough for a quick social post."
One frame tells the model where things are. Multiple frames tell it where things are going, and at what rate.
The comparison is direct. Take a model like Kling v3 Video or Veo 3. Both are excellent text-to-video generators in their own right. Neither was designed around seamless video continuation as a first-class architectural feature.
When you use image-to-video (I2V) mode on these models to "extend" a clip, you are feeding the last frame as a static image. The model then generates motion from that frame based on your text prompt. Here is where the seam shows:
The visual seam where the original clip ends and the I2V extension begins is visible in most models if you look carefully. At 24fps played on a large screen, it is often visible without looking carefully at all.
💡 This is why professional video creators often describe I2V extension as "fine for social media" but unsuitable for anything that will be screened in a professional context, presented to a client, or included in a production reel. The seam is the tell.
Sora 2 Pro and Wan 2.7 T2V also produce strong standalone footage but similarly lack native multi-frame prior conditioning for extension workflows. They are generation tools, not continuation tools. The distinction matters more and more as creators try to build longer, more coherent video narratives with AI.
Most I2V approaches condition on a single frame. LTX 2.3 Pro conditions on multiple prior frames encoded together into a unified latent representation. This gives the model enough temporal data to infer object velocity vectors, camera movement speed and direction, and the progression of lighting gradients across the scene over time.
The difference in output quality between single-frame and multi-frame conditioning is not marginal. It is structural. Single-frame conditioning produces clips that look like they continue a scene. Multi-frame conditioning produces clips that are a continuation of the scene at the physics level.
LTX 2.3 Pro generates at 4K resolution natively. This matters for extension specifically because high-resolution output makes the join between original and extended footage far easier to work with in post-production. At 4K, you have more pixel data to work with if you need to apply a cross-dissolve over the join, and the sheer detail of the frame means any remaining inconsistencies at the boundary are less perceptible to the viewer.
Compare this to LTX 2.3 Fast, which prioritizes speed over resolution and is a strong option when you need rapid iteration, or the original LTX Video, which established the underlying architecture but operates at significantly lower resolutions. For final delivery, LTX 2.3 Pro is the model you want.
Extend Mode does not just continue the scene mechanically. You can provide a text prompt that guides what happens next within the constraints established by the prior frames. This means you can prompt for a character to turn around, for the camera to dolly forward, or for the light to shift as a cloud passes overhead, and the model will attempt those changes while maintaining consistency with the prior footage's established visual language.
This is scene continuation with directorial control, which is a fundamentally different tool than pure mechanical extrapolation.
Here is a direct feature comparison for video extension use cases:
| Model | Extension Method | Resolution | Temporal Coherence | Prompt Steering |
|---|---|---|---|---|
| LTX 2.3 Pro | Multi-frame prior conditioning | 4K | High | Yes |
| LTX 2 Pro | Multi-frame prior | 1080p | High | Yes |
| Kling v3 Video | Single-frame I2V | 1080p | Moderate | Yes |
| Veo 3 | Single-frame I2V | 1080p | Moderate | Yes |
| Wan 2.7 T2V | Text regeneration | 1080p | Low for extension | Yes |
| Hailuo 2.3 | Single-frame I2V | 1080p | Moderate | Limited |
| Ray 3.2 | Single-frame I2V | HDR | Moderate | Yes |
| Seedance 2.5 | Text and image | 30-second clips | Low for extension | Yes |
The table makes the positioning clear: LTX 2.3 Pro is the only model in this group combining multi-frame prior conditioning with 4K resolution output. Its predecessor LTX 2 Pro shares the conditioning architecture but operates at 1080p, making LTX 2.3 Pro the current ceiling for high-fidelity extension work.
LTX 2.3 Pro is available directly on PicassoIA. Here is how the Extend Mode workflow runs in practice:
Step 1: Generate or upload your base clip. Start with any video clip you want to continue. This can be a clip you generated earlier with LTX 2.3 Pro, LTX 2.3 Fast, or any external footage that meets the model's input requirements.
Step 2: Select Extend Mode. In the model interface, choose the extension option rather than standard generation. This activates multi-frame conditioning rather than fresh noise sampling.
Step 3: Set the extension length. Specify how many seconds of new footage you want generated. Shorter extensions in the 3-5 second range tend to have the highest coherence. Longer extensions can work well but may show more prompt drift toward the far end of the clip.
Step 4: Write your continuation prompt. Be specific about motion and scene changes. Instead of "woman continues walking," try "woman slows to a stop and turns to look at the camera, sunlight still from the left, same park background, light wind moving her hair." Specificity gives the model clear directorial intent to follow within the established scene.
Step 5: Review and iterate. Watch the extension at full playback speed, then step through frame by frame at the join point. If the join is not seamless, regenerate with a slightly adjusted prompt or try using the last two frames of the extension as a new base clip for a fresh extension call.
💡 For long-form content, the most reliable method is chaining: generate 5 seconds, extend by 5, extend again. Each extension conditions on fresh prior frames, keeping coherence high even as total clip length grows toward 30, 60 seconds, or beyond.
The honest assessment, based on what the architecture promises and what it delivers in practice:
Where it excels:
Where it takes more prompt work:
💡 The more information the prior frames encode about the scene's physics, the better the extension performs. Stable, well-lit footage extends more cleanly than shaky, high-contrast footage. This is not a limitation of the model. It is a physical property of what multi-frame conditioning can infer from the data it receives.
The scenes that produce the most satisfying extensions on a first attempt are also the scenes that look best in live-action production: good light, deliberate camera movement, and a clear subject with a clear spatial relationship to its environment.
It is worth spending a moment on what Lightricks built to make this work, because the architecture is what separates LTX 2.3 Pro from models that approximate extension through workarounds.
LTX Video's core is a latent diffusion transformer rather than the U-Net based architecture that dominated the first generation of video diffusion models. Transformer architectures have a natural advantage for temporal modeling: the self-attention mechanism can be configured to operate across both spatial and temporal dimensions simultaneously, a property called spatiotemporal attention.
In standard text-to-video transformers, the temporal attention is self-contained within a single generation window. In LTX 2.3 Pro's Extend Mode, the temporal attention window is extended backward into the encoded prior frames. The model literally attends to its own past during the diffusion process, not as a post-processing step, but as an integral part of denoising itself.
This is why the results differ so substantially from I2V approaches. I2V gives the model a photograph and says "make this move." LTX 2.3 Pro's Extend Mode gives the model a memory and says "keep going from here." The first produces movement. The second produces continuation.
The 4K output resolution compounds this advantage. At higher resolution, the model has more spatial detail to condition on within each prior frame, which means temporal attention can track finer details: specific light reflections on surfaces, individual strands of hair movement between frames, small environmental elements like leaves shifting or fabric folds responding to body movement.
One of the most practical applications of Extend Mode is extension chaining. Because each call to Extend Mode conditions on the most recent frames of the previous generation, you can chain multiple extensions together to produce clips well beyond what any single generation window would allow, while maintaining the scene's visual consistency throughout.
The workflow:
Each step maintains the scene's established visual language. The cumulative visual drift over 5-6 chained extensions is minimal compared to regenerating from scratch at each step or stitching together I2V clips at the seams. Creators working on product demos, architectural walkthroughs, nature documentaries, and short film content have reported clean 30-60 second sequences assembled this way, with seams that are not visible at standard playback speed.
💡 If you are building a product demo, a real estate walkthrough, or an atmospheric scene for a short film, extension chaining with LTX 2.3 Pro is the fastest path to polished long-form footage that would otherwise require a physical camera crew.
The technique also pairs well with PicassoIA's broader video toolset. You can generate the base clip with LTX 2.3 Pro, apply visual effects from the Effects library (500+ options), and use AI video restoration tools to stabilize or upscale the final assembled sequence. The chain can be as simple or as layered as your project requires.
LTX 2.3 Pro is live on PicassoIA alongside its faster sibling LTX 2.3 Fast for when you need rapid drafts and iteration. The full model library at picassoia.com/en/all-models includes over 87 video generation models spanning text-to-video, image-to-video, video effects, and restoration tools, so there is a full pipeline available regardless of where your project starts.
If you have never worked with video extension before, start with a simple scene: a person walking down a street at golden hour, a river flowing through a rocky landscape, a candle burning on a table in a dark room. These scene types give Extend Mode enough temporal data to work with and produce the most satisfying results on a first attempt. The lighting is consistent, the motion has clear direction, and the spatial environment is easy for the model to track across frames.
From there, as you build a feel for how prior-frame conditioning responds to different kinds of input footage, you can push into more complex scenarios: multi-character scenes, rapid camera movements, and narrative sequences that unfold over 30 seconds or more through chained extensions.
The Lightricks approach to temporal coherence is not just a feature addition to an existing pipeline. It represents a different philosophy about what AI video generation should do: not just produce isolated clips, but produce footage you can actually build a production with.
Try LTX 2.3 Pro on PicassoIA now and see the difference that architecture-level temporal conditioning makes on your first clip.
