iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Decart AI is a company developing models and infrastructure for AI that responds while video or interaction is happening. Its public product framing has three distinct layers: DOS, an inference and performance stack; Lucy, which transforms live video; and Oasis, which generates interactive worlds in response to user actions.
The key distinction is that Lucy edits an existing camera stream, while Oasis creates the next frames of a playable environment. Both are generative systems, but they solve different real-time problems and should not be evaluated with the same performance figures.
What Decart AI is
Decart positions itself as an AI company focused on live workloads: systems that must generate useful output continuously instead of waiting for a complete prompt, rendering a finished clip, or processing a recording after the fact. That requires both generative models and the infrastructure needed to run them quickly enough for an interactive experience.
Its public product structure separates the work into three parts:
#1 Best Overall
- EVOLUTION AMD RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
- DOS: the inference and performance layer intended to make demanding real-time models practical to run.
- Lucy: a live video-to-video system that changes an incoming stream as it plays.
- Oasis: an interactive world model that generates a visual environment from user actions rather than transforming a camera feed.
This separation matters when reading specifications. A frame-rate claim for Lucy is not automatically a frame-rate claim for Oasis, and a historical Oasis demonstration should not be treated as the specification of the current Lucy product.
How Lucy performs real-time video editing
The input and output
Lucy receives a live video stream and returns a transformed stream. Instead of recording footage, rendering an effect offline, and then playing the result, it generates frames while the scene changes. Decart describes each frame as responding continuously to motion, lighting, and the evolving scene rather than relying on a pre-rendered clip.
In practical terms, a camera or other live source supplies the visual context. A prompt, reference image, or selected transformation tells the model what should change. Lucy then produces an output frame that attempts to preserve the underlying action while applying the requested edit.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What Lucy can change
Decart’s Lucy 2.5 description lists several categories of transformation:
- changing clothing or character appearance;
- adding visual effects;
- changing the style of a scene;
- removing objects; and
- transforming an environment.
These are video transformations, not merely single-image filters. The difficult part is maintaining a coherent edit as people move, objects become occluded, illumination changes, and the camera view evolves.
How temporal consistency is trained
For Lucy 2.5, Decart says it created synthetic training data by applying image edits and then using specialized models to propagate those edits across video sequences. This gives the training process examples of how an intended change should persist from one frame to the next.
Decart also describes a technique called self-anchoring. After a stream starts, Lucy takes a snapshot of its own generated output and uses that result as a reference. The stated goal is to keep the chosen transformation more stable over time, reducing the tendency for an identity, costume, or scene style to drift as new frames arrive.
Lucy 2.5 performance claims
In a July 2026 publication, Decart states that Lucy 2.5 can perform live transformations at 30 frames per second and 1080p resolution. The same publication says its low-precision techniques can provide up to 4x speedups on compute-bound operations when compatible hardware is used.
Those numbers are company-published capabilities, not independent benchmark results. “Up to 4x” applies to the operations and hardware conditions described by Decart; it is not a promise of a fourfold reduction in total camera-to-display delay for every setup.
Rank #2
- Built for Local AI Development: AMD Ryzen AI Halo is designed for local AI development and inference, featuring 128GB unified memory and support for up to 200B parameter models to build and run intensive AI workloads locally.
- 128GB Unified Memory: Features 128GB LPDDR5x unified memory at 8000 MT/s with 256 GB/s memory bandwidth, providing a shared memory pool across the CPU, GPU, and NPU to support larger AI models.
- AMD Ryzen AI Max+ 395 Processor: Features 16 cores, 32 threads, and Zen 5 architecture, paired with AMD Radeon 8060S integrated graphics featuring 40 RDNA 3.5 compute units and an AMD XDNA 2 NPU with up to 50 TOPS.
- Linux AI Developer Platform: Purpose-built for Linux-based AI development with full AMD ROCm software support and preloaded tools, models, and workflows optimized for local AI development.
- Compact, Connected Design: Includes a 2TB M.2 SSD, 10GbE LAN, Wi-Fi 7, Bluetooth 5.4, USB-C connectivity, and HDMI 2.1b.
How Oasis generates an interactive world
Actions are the control signal
Oasis addresses a different problem. Its input includes keyboard and mouse actions, and the model generates the next visual frames of an environment in response. Decart describes the result as a playable experience that internally simulates elements such as physics, rules, and graphics.
That makes Oasis an interactive world model, not a video editor. There is no requirement for a pre-existing camera scene to be preserved. The model has to decide what the next part of the generated world should look like after an action such as moving, turning, or interacting.
Free tools Windows power users keep installed
One-click scans. No signup required.
The published Oasis architecture
Decart’s October 2024 Oasis article gives the clearest public explanation of its architecture. It describes a transformer that predicts the diffusion process in a compressed latent representation. Temporal-attention layers provide context from earlier generated frames, while the system produces new frames autoregressively.
“Autoregressive” means that the next output is conditioned on what has already been generated. In Oasis, user actions are part of that conditioning. A movement input can therefore influence the visual state that follows, rather than merely selecting a pre-rendered branch.
This is a specific explanation of the published Oasis system from 2024. It should not be read as a disclosure that every current Decart model uses exactly the same architecture or training recipe.
Historical Oasis figures
Decart reported 47 milliseconds of inference per frame for the 2024 Oasis system. The same publication described a demonstration running at 360p and 20 frames per second on NVIDIA H100 GPUs.
These are dated, model-specific figures from an Oasis demonstration. They are not current Lucy 2.5 specifications, and they do not establish the end-to-end latency a user will experience on a different model, browser, network, or hardware configuration.
Limitations disclosed for the published demo
The Oasis publication acknowledged several weaknesses in that historical system:
- the output could become hazy in some situations;
- details from many frames earlier could be difficult to recall; and
- an initial image outside the model’s training distribution could produce unclear results.
Those caveats describe the published Oasis demo. They should not automatically be assigned to Lucy 2.5 or to later Oasis releases whose behavior may differ.
Rank #3
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
What makes a real-time generative system feel real time
Inference is part of the delay
A camera-to-screen experience has more stages than a network ping. The camera must capture a frame, the browser must send media, the service must run inference, and the resulting frame must travel back and be decoded and displayed.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Decart’s developer SDK documentation warns that network round-trip time alone can understate perceived lag because model inference contributes to the total. Where browser support permits, its documented timing approach measures glass-to-glass latency: the interval from the source frame entering the system to the transformed frame appearing on the display.
Performance methods described for Lucy
Decart identifies several inference optimizations for Lucy 2.5:
- Low-precision quantization, including MXFP8 and NVFP4, to reduce the computational cost of model operations on compatible hardware.
- Dynamic sparse attention, which limits attention work to the most useful relationships instead of treating every possible interaction equally.
- Deeper kernel fusion, which combines operations so the processor spends less time moving intermediate data between separate steps.
These are vendor-described implementation techniques. The published material does not provide an independent, full-system benchmark showing how much each technique improves camera-to-display latency in a particular deployment.
AdaOr and controllable edit strength
In a February 2026 research publication, Decart described AdaOr, an editing-control method that introduces an identity token meaning “no change.” The model’s guidance origin can then be interpolated between faithful reconstruction and a fully prompt-driven edit.
Recommended Free Tools
The intended effect is a smooth edit-strength control. A small change to the wording of a prompt should not, by itself, cause an unpredictable jump from nearly unchanged footage to a radically different result. AdaOr is a published research method; the reviewed material does not establish which production model versions expose it as a user-facing control.
How a camera integration is structured
Decart’s SDK documentation illustrates a browser-based service pattern rather than a downloadable offline editor. Exact API names and supported models can change, but the documented flow has these stages:
- Prepare a supported browser and camera source. The browser obtains a camera
MediaStream, subject to the user’s camera permission and the browser’s media capabilities. - Select a real-time model. The integration chooses the available model intended for the desired transformation.
- Send the live stream. The browser transmits the camera stream to the service while the session is active.
- Receive the transformed stream. The service returns a remote media stream containing generated video.
- Measure both startup and steady-state behavior. The SDK documentation distinguishes the time needed to begin displaying output from the continuing glass-to-glass delay once the stream is running.
This arrangement means an application must account for camera permissions, browser support, network conditions, inference capacity, and display buffering. A low network RTT can still feel slow if the selected model takes substantial time to generate each frame.
Lucy and Oasis compared
| Dimension | Lucy | Oasis |
|---|---|---|
| Primary task | Real-time video editing and transformation | Interactive world generation |
| Typical input | Live video, with prompts or reference controls for the edit | Keyboard and mouse actions, with the model generating the environment |
| What the model must preserve | The motion and evolving content of an incoming scene while applying an edit | A coherent interactive state across newly generated frames |
| Public architecture detail | Training and inference techniques are described, but a complete architecture is not disclosed in the reviewed material | Transformer and diffusion components operating in compressed latent space, with temporal attention, were described for the October 2024 publication |
| Published performance figure | Decart states 30 FPS at 1080p for Lucy 2.5; no independent validation is established here | 47 ms inference per frame and 360p at 20 FPS on H100 GPUs for the 2024 demonstration |
| Main latency question | How quickly a camera frame becomes a stable transformed frame on screen | How quickly an action changes the next generated state |
| Known published caveat | Availability and supported model behavior can change by release | The 2024 demo could become hazy, forget distant details, or handle out-of-distribution starting images poorly |
How to interpret Decart’s numbers and claims
Real-time specifications are meaningful only with their conditions attached. Lucy’s 30 FPS and 1080p figures describe what Decart says its current product can do; they do not identify the exact hardware, scene complexity, prompt, stream duration, or network conditions for every session. The “up to 4x” figure concerns compute-bound operations and compatible hardware rather than guaranteed total latency.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOasis’s 47 ms figure is inference time per frame in a 2024 system. Even if that number remains reproducible, a user’s experience would also include input capture, transport, queuing, decoding, and display. Comparing it directly with a camera-to-screen measurement for Lucy would mix different workloads and different definitions of latency.
For a serious evaluation, record at least:
- the model and release date;
- input and output resolution;
- frame rate;
- hardware and precision mode;
- network path and round-trip time;
- startup time before the first usable frame; and
- steady-state glass-to-glass latency.
Without those details, a single FPS or milliseconds figure can give a false impression of how responsive an application will feel.
Where each Decart layer fits
DOS for the execution layer
DOS is best understood as the infrastructure layer. Its role is to make inference efficient enough for live workloads, where a model cannot simply spend several seconds rendering one result. It is relevant to deployment and performance, not a separate visual effect that a user applies to a camera feed.
Lucy for live visual transformation
Lucy fits applications in which the source scene already exists and the desired experience is a changing version of that scene. A live presenter, performer, game stream, or camera view can remain temporally connected while clothing, environment, style, or other visual properties are altered.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Oasis for generated interaction
Oasis fits applications in which the environment itself is generated and user actions determine what appears next. Its challenge is closer to maintaining a playable visual state than to preserving the identity of every object in a real camera recording.
Practical limitations and questions to resolve before deployment
Anyone evaluating a Decart-based experience should verify the current service documentation rather than relying on an old demonstration. Model names, browser support, output limits, access requirements, and latency can change as the system is updated.
- Check the actual model release. Do not use 2024 Oasis figures to size a 2026 Lucy deployment.
- Test the complete path. Measure from camera capture to displayed output, not just network RTT or server inference time.
- Test difficult motion and lighting. Temporal stability is most visible when subjects turn, objects overlap, or illumination changes quickly.
- Separate visual quality from responsiveness. A higher resolution or more elaborate transformation may require more inference time.
- Record the hardware and precision mode. Quantization benefits depend on compatible processors and implementation details.
- Plan for model-specific failure modes. The disclosed Oasis demo’s haze, long-range memory problems, and sensitivity to unusual starting images illustrate why a graceful fallback may be necessary.
Bottom line: what Decart AI actually represents
Decart AI is not one single “real-time AI” feature. It is a group of models and an inference stack aimed at making generative systems respond continuously. Lucy transforms an existing live video stream, while Oasis generates an interactive world conditioned on user actions. DOS addresses the execution problem underneath those experiences.
The most reliable way to assess a Decart system is to keep the product and date attached to every claim. Lucy 2.5’s 30 FPS and 1080p figures are current company statements, the 4x figure is an “up to” compute claim, and Oasis’s 47 ms and 360p/20 FPS numbers belong to an older 2024 demonstration. For an application that must feel immediate, measure end-to-end glass-to-glass latency on the exact model, hardware, browser, and network path you intend to use.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
