Dreamina's Lucy 2.5 restyles live video at 1080p/30fps using self-anchoring to stay stable, with API pricing far below typical video model costs.
What is Lucy 2.5? #
Lucy 2.5 is a real-time video restyling model from Dreamina that processes live footage at 1080p and 30 frames per second with near-zero latency. It can swap out a character, change the visual style of a scene, add effects, or erase objects, all while the camera is still rolling, guided by text prompts and reference images. The API is priced at roughly 2 cents per second at 720p, a fraction of what comparable video generation typically costs.
TL;DR #
Lucy 2.5 processes video live at 1080p/30fps, applying edits like character swaps, restyling, and object removal while footage is still being captured.- The model uses a technique called self-anchoring, where each generated frame becomes the reference point for the next, which keeps output from drifting into distorted or incoherent results over time. - Pricing is unusually aggressive: the API runs about 2 cents per second at 720p, well below the cost trajectory most video models have been on. - There’s no standalone consumer app yet. Access currently runs through a browser playground and the API, meaning most people will experience it through third-party tools or custom builds. - A practical workflow use case is live previsualization: shoot rough footage on a phone, restyle it instantly with Lucy 2.5 to block a shot, then send it through a higher-fidelity video-to-video model for the final render. - The term “video editing” is doing double duty here. Lucy 2.5 isn’t cutting timelines, it’s doing live compositing and VFX work on top of existing footage.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How does Lucy 2.5 actually work? #
Lucy 2.5 sits in the category of video-to-video models, but the “real-time” part is the distinguishing feature. Instead of processing a clip after it’s recorded and returning a finished file minutes later, it operates on a live or near-live feed, applying transformations frame by frame with latency low enough to feel instantaneous.
The demonstrated capabilities include:
- Swapping a person in frame for a different character, using a reference image to define what they should look like.
- Restyling an entire scene’s aesthetic on the fly.
- Injecting visual effects (like an on-body fire effect) directly into a live shot.
- Removing objects from the frame as it plays.
All of this is driven by two inputs: a text prompt describing the desired change, and optionally a reference image (a character, an object, a style) that the model should pull from. In one demonstrated example, a live camera feed of a person was restyled in real time into an animated character, then swapped again to match a completely different reference character, with the transformation holding up shot to shot.
What is self-anchoring and why does it matter? #
Self-anchoring is the mechanism Dreamina points to as the reason Lucy 2.5 doesn’t degrade into visual noise over time. In a naive frame-by-frame generation setup, small errors or stylistic drift can compound. Each new frame is influenced by the last, and without a stabilizing reference, that chain can spiral into distorted, flickering, or increasingly incoherent output, the kind of psychedelic breakdown that’s plagued earlier real-time video experiments.
Self-anchoring works by having the model treat its own most recent generated output as the new reference point for the next frame, rather than only relying on the original prompt or first frame. That constant recalibration against its own recent work is what keeps a character’s face, a scene’s color palette, or an object’s shape from sliding into distortion across a long sequence. In practice, this is what makes something like a multi-second live restyle usable instead of a fun tech demo that falls apart after a few frames.
Is Lucy 2.5’s pricing actually a big deal? #
Yes, and it’s arguably the more surprising part of the release. Generation costs across the AI video space have been trending upward as models get more capable and compute-hungry. Lucy 2.5’s API pricing, at around 2 cents per second at 720p, breaks that pattern. For context, that kind of cost structure makes it financially viable to run continuous or near-continuous video processing rather than treating each generation as a precious, expensive one-off.
That pricing matters most for anyone thinking about building on top of the API rather than using the hosted playground. At 2 cents a second, extended experimentation, iterative testing, and even production use in latency-tolerant workflows become realistic in a way they aren’t with pricier video models.
Where can you actually use it right now? #
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
There’s currently no consumer app for Lucy 2.5, despite demo footage showing someone using a phone to restyle their surroundings on the fly. The two ways to access it today are:
A browser-based playground hosted by Dreamina, which includes preset effects (like an on-body fire effect) and supports custom prompts plus reference image uploads.A direct API, priced as described above, which developers can build against.
The gap between the flashy phone-based demo and the current no-app reality means that, for now, anyone wanting that “point your phone at something and watch it transform live” experience will need to build it themselves using the API, or wait for Dreamina to ship a dedicated app.
What’s the best way to actually use a tool like this? #
The most practical near-term use case isn’t necessarily shipping Lucy 2.5 output as a finished product. It’s previsualization. Because the model is fast and cheap, it’s well suited to being a live storyboarding tool: shoot rough blocking footage on a phone, run it through Lucy 2.5 to instantly visualize a restyled scene, character swap, or effect, and use that to make creative decisions before committing to a full, higher-fidelity render pass through a slower and more expensive video-to-video model.
This mirrors a broader pattern showing up across AI video tools right now: cheap, fast, rough-draft generation paired with a separate, higher-quality finishing step. Rather than one model doing everything, workflows are increasingly splitting into a fast iteration loop and a slower final render.
Frequently Asked Questions #
Is Lucy 2.5 actually editing video in the traditional sense?
No. It’s not doing timeline-based cuts or nonlinear editing. It’s closer to live VFX or compositing, applying visual transformations (restyling, character swaps, object removal) on top of existing or incoming footage, without traditional editing steps like trimming or sequencing clips.
How much does Lucy 2.5 cost to use?
Through the API, pricing runs at roughly 2 cents per second at 720p resolution. That’s notably cheaper than the direction most video generation pricing has been heading, making sustained or iterative use more financially practical.
Can I use Lucy 2.5 without coding?
Yes, currently through a browser-based playground that supports prompt input, reference images, and some preset effects. There’s no dedicated mobile or desktop app yet, so anyone wanting a packaged app experience will need to wait or build a custom interface on top of the API.
What makes self-anchoring different from standard video generation?
Standard frame-by-frame video generation can drift or degrade as small inconsistencies compound over a sequence. Self-anchoring has the model use its own most recently generated frame as an ongoing reference point, which helps keep characters, color, and scene structure stable across longer real-time sequences instead of devolving into distorted output.
Does Lucy 2.5 replace higher-end video-to-video models?
Not necessarily. Its speed and low cost make it well suited for real-time previsualization or rough drafts, but final production output for many creators still runs through slower, higher-fidelity video-to-video models as a separate finishing pass.