REAL-TIME. LOCAL. YOURS.

Your camera.
Another
reality.

Turn a live camera into a creative canvas. Transform your character with AI that runs on your own machine.

Discover Dash See what’s possible

BUILT FOR CREATIVE EXPERIMENTSPOWERED BY YOUR GPU

THE DASH EFFECT01 / CHARACTER STUDIO

CAMERA INPUTCHROME CHARACTER

ONE CAMERA.
A NEW CHARACTER.AI-GENERATED
STYLE CONCEPT

TRY A DIRECTION

“Give me a polished chrome character.”

Drag the divider to explore.ILLUSTRATIVE PREVIEW

FOR THE ONES WHO MAKE THINGS.CreatorsLive streamersDigital artistsCreative studios

01 / THE POSSIBILITIES

A different kind
of camera presence.

Explore a new look. Build a character. Give your next creative experiment a world of its own.

CHARACTER CONCEPT / 01Still you.
A whole new look.

Change the character.

Use a prompt to explore character styles while your camera supplies the pose and expression.

Your cameradash.

inference.locationyour_machineYour GPU does the creative work.

LOCAL BY DESIGN

Keep it on your machine.

Run core AI inference on your own GPU. Open the studio in your browser and keep your workflow close.

Live cameraIN THE MOMENT

Source videoYOUR FOOTAGE

Batch tasksONE MODEL LOAD

ONE STUDIO. THREE WAYS IN.

Work at your own pace.

Start with your camera, edit existing footage, or queue multiple tasks in a single session.

THE TRANSFORMATION / IN MOTION

A closer look at what’s possible.

Watch appearance edits follow the source motion, in a presentation redesigned for Dash.

2 minutes · 720p. Supplied research demonstration with Dash branding and a new visual layout. The original footage, timing, and audio are preserved; this is not a separate Dash performance benchmark.

Open full video ↗

02 / YOUR CREATIVE WORKFLOW

You bring
the idea.
Dash brings
the possibility.

From a blank prompt to a character experiment, all in one local workspace.

Explore the studio

A compatible CUDA GPU and initial model setup are required.

Make it your studio.

Set up the models on your GPU machine, open the browser interface, and prepare your preferred configuration.

LOCAL WORKSPACE

Tell Dash what you imagine.

Write your character direction. Try a different material, an animated look, or a softer creative style.

PROMPT-BASED EDITING

Step into the experiment.

Enable your camera, frame the shot, and start editing. Reset, change the prompt, and explore again.

CREATE. RESET. REPEAT.

03 / UNDER THE HOOD

The technology behind
the transformation.

Appearance, motion, and temporal context come together in a streaming video diffusion pipeline.

BASE MODEL14B

Wan2.2-Animate

MOTION CONDITIONINGBody + face

Explicit pose and expression signals

TEMPORAL CONTEXTRolling KV

Chunk-by-chunk streaming cache

STAGE 01 / EDITING

Separate appearance
from motion.

Only the reference appearance is edited synthetically. A reversed prompt then trains reconstruction of the original video. Clean-reference conditioning, body pose, and facial features separate appearance from motion.

TrainableSelf- and cross-attention LoRAs
FrozenFace blocks

STAGE 02 / ADAPTATION

Make the context
causal.

Each target chunk sees its own chunk, the clean reference, and preceding clean context. Future context and other noisy chunks are masked. Flow-matching loss is applied only to targets.

TrainableSelf-attention + FFN LoRAs
FrozenCross-attention and face blocks

STAGE 03 / DISTILLATION

Align training
with the rollout.

Training completes the denoising rollout and backpropagates at the sampled step. A retained attention sink and rolling local cache support continued generation; fixed RoPE and FPSA preserve the intended temporal context.

CacheRolling keys and values
PositionFixed RoPE
AttentionFirst-frame preserved sparse attention (FPSA)

The figures and animation explain the published training design. The studio runtime uses pretrained adapters; training code is not currently included. Source paper: Figures 3 and 5, Appendix B.1 ↗

WATCH & EXPLORE

Three stages. One guided walkthrough.
1:26 · 1080p · Silent animation · Use fullscreen for the details.

A paper schematic with animated focus and connector markers. The original equations, masks, and training roles are preserved. Source: Figures 3 and 5, Appendix B.1 · Open animation ↗

The research, explained by an AI presenter.
2:01 · 1080p · English captions available with CC · Fullscreen for the details.

An original AI-generated presenter and voice explain the three training stages, cache behavior, and Dash’s use of pretrained components. The script was checked against the source paper · Open presenter video ↗

Read the full transcript

Welcome to Dash. Let’s unpack the research pipeline behind its local video-editing workflow. This diagram describes three training stages, rather than a live recording of a model running.

First, appearance and motion are separated. Only the reference image is edited synthetically, and a reverse instruction trains reconstruction of the original video. Body-pose signals and facial features guide movement, while the reference and editing instruction jointly control appearance.

Next, the model adapts to streaming. Each chunk can attend throughout itself, to the clean reference, and to preceding clean context. Future context and other noisy chunks are masked. Training applies the flow-matching loss to target chunks only.

The third stage is aligned self-rollout distillation. The model completes both denoising steps, while gradients pass through the sampled step. An extra clean cache pass builds the retained attention sink from the first generated chunk. Subsequent chunks pass their final denoising-step keys and values to following chunks. Frozen real-score and trainable fake-score diffusion models provide the distillation signal.

A retained attention sink and rolling local cache manage context. Fixed rotary positions and first-frame-preserved sparse attention help maintain appearance and temporal consistency. The numbers here identify latent context positions, not ordinary video-frame numbers.

Dash brings pretrained components from this pipeline into a local studio workflow. Actual output quality and speed depend on the input, settings, and hardware.

THE LOCAL RUNTIME

A configurable creative engine.

Choose the balance of image size, memory use, and throughput that fits your GPU.

01 / MODEL STACK

Base model: Wan2.2-Animate-14B
Adapters: Editing LoRA, streaming LoRA, LightX2V
Decoder: Camera default: Flash-VAED; Wan VAE available
Precision: Camera default: FP8; BF16 available

02 / CAPTURE & INPUT

Camera output: 672 × 384 or 832 × 480, landscape
Capture target: Adjustable 8–25 fps
Video input: Raw video, preprocessed folder, JSON batch
Reference: Camera frame; video inference also accepts a separate image

03 / EXECUTION

Acceleration: FlashAttention + FastVideo kernels
Compilation: torch.compile with warmup and a persistent cache option
GPU layout: One GPU; optional second GPU for pose and VAE work
Build environment: Python 3.10, CUDA toolkit 12.3+, GCC 10+ for kernels

FILE INFERENCE / H100 REFERENCE

Find your memory balance.

For file inference, move weights or the rolling KV cache to CPU memory to reduce GPU memory use, with a speed tradeoff.

FILE INFERENCE · H100 · FP8 · FAST DECODE · NO COMPILE

Reported reference memory and slowdown by output resolution
Memory mode	672 × 384	832 × 480
GPU resident	31.4 GiBBaseline	38.4 GiBBaseline
CPU weights	17.5 GiB1.02× slower	24.5 GiB1.02× slower
CPU weights + KV	14.4 GiB1.71× slower	15.0 GiB1.57× slower

These are reported source measurements, rather than Dash-specific validation. Capture FPS is a camera upload target, not guaranteed AI output speed. Low-memory modes and torch.compile are mutually exclusive.

SYSTEM REQUIREMENTS

Know your setup before you start.

Live camera and file editing have different memory paths. Choose hardware for the mode you intend to use.

LIVE CAMERA / PLANNING TARGET

48 GB+ GPU VRAM

The current camera interface uses compilation and GPU-resident weights and cache. CPU offload is not exposed here. Allow extra headroom for pose, decoding, and compilation.

A live-camera minimum has not been published. 48 GB is a provisional workstation target, not a validated minimum.

FILE EDITING / EXPERIMENTAL FLOOR

16 GB GPU VRAM

Start at 672 × 384 with FP8, Flash-VAED, CPU weights + KV offload, and compilation disabled. The lowest reported footprint for this configuration is 14.4 GiB on H100.

16 GB is inferred from that footprint. A 16 GB consumer GPU has not been validated here; 24 GB provides more headroom.

Hardware and software prerequisites, with planning estimates distinguished from documented requirements
Component	Documented requirement	Practical planning target
GPU architecture	NVIDIA CUDA. FP8 requires SM 8.9 or newer.	Ada or Hopper for the pinned FP8 workflow. RTX 4090 is an Ada candidate; H100 is the reported benchmark GPU.
System RAM	No validated RAM minimum published. CPU offload needs additional host memory.	64 GB RAM to plan model loading and offload; 128 GB for more build and multitasking headroom. Estimates require validation.
CPU	No validated core-count minimum published.	Modern 8-core x86-64 CPU or better. Build tools and preprocessing use the CPU; generation relies on the GPU.
Storage	Space for base weights, adapters, dependencies, caches, and outputs.	150 GB free SSD space as a setup allowance. Reserve additional space for downloaded caches and generated videos; not a measured minimum.
Operating system	Linux is the primary build path for the pinned attention library. Native Windows builds need additional testing.	Ubuntu 22.04 LTS, 64-bit is a CUDA 12.4-compatible starting point, not a validated Dash installation. Windows/WSL2 and webcam forwarding need separate testing; macOS, Apple GPUs, AMD, and CPU-only use are not advertised as supported.
Python & PyTorch	Audited stack: Python 3.10, PyTorch 2.6.0, torchvision 0.21.0.	Use the pinned CUDA-enabled packages in an isolated environment.
CUDA & compiler	CUDA toolkit 12.3+ and GCC 10+ for FastVideo kernels. The dependency stack was audited with CUDA 12.4.	CUDA 12.4 with a compatible NVIDIA driver. Match the driver, toolkit, PyTorch build, and GPU architecture.
GPU driver	The NVIDIA driver must support the installed CUDA build.	For the CUDA 12.4 GA stack, use the corresponding Linux driver level 550.54.14 or newer. CUDA minor-version compatibility has separate rules; confirm the complete setup.
Kernel packages	FlashAttention 2.7.2.post1, FastVideo kernel 0.3.0, and Ninja.	Limit compilation jobs when RAM is constrained. Kernel builds are part of initial setup.
Browser & camera	Browser camera access on localhost or HTTPS. Camera permission is required for live mode.	A working webcam and current desktop browser. Node.js and npm are needed to build the interface; FFmpeg is needed for video preprocessing.
Internet	Needed initially to download models and install dependencies.	Once installed, core inference runs locally. Using a remote GPU requires a connection to that machine.

Before choosing a GPU

VRAM capacity alone does not establish compatibility or live speed. The 14.4–15.0 GiB numbers describe file inference with FP8 and CPU offload, not the webcam interface. Ampere cards such as RTX 3090 do not meet this FP8 path’s SM 8.9 requirement. Newer architectures still need compatible builds of the pinned kernels.

Check NVIDIA compute capability Attention library platform requirements CUDA 12.4 Linux environment CUDA driver compatibility

A few technical terms, explained

VAE & latents

The VAE encodes frames into compact representations called latents and decodes generated latents into images.

LoRA adapters

Additional model weights that adapt the base model for editing, acceleration, and streaming.

KV cache

Stored attention keys and values that carry context from previous video chunks into the current chunk.

Causal attention

An attention mask that controls which reference and temporal context each video chunk can access.

04 / MEET YOUR NEXT STUDIO

Big creative energy.
One simple price.

A dedicated local AI workspace for creators who want to explore beyond the ordinary camera feed.

Your setup. Your creative space.The AI runs on your hardware. GPU hardware is separate.

A few things to know before you start

DASH STUDIOPRE-ORDER

$599.90USD

One-time pre-order payment.

Real-time local camera workspace
Prompt-based character transformations
Source video and batch task workflows
Browser-based studio interface

05 / THE DETAILS

Good questions.
Clear answers.

Does Dash run locally?

Core AI inference runs on your GPU machine. Models and software dependencies are downloaded during setup, and you use a browser to open the local interface. A remote GPU machine can also be accessed through a secure tunnel.

What hardware do I need?

For planning, allow a Linux workstation with a modern 8-core CPU, 64 GB RAM, and 150 GB free SSD space. These are setup targets; validated CPU, RAM, and disk minimums have not been published.

Live camera: 48 GB+ NVIDIA GPU VRAM is a provisional planning target. The current interface does not expose CPU offload, and a live-camera minimum is not yet established.
File editing: 16 GB VRAM is an experimental target with FP8, Flash-VAED, CPU weights + KV offload, and compilation off at 672 × 384. The reported 14.4 GiB footprint was measured on H100, not a 16 GB consumer card.
FP8: NVIDIA compute capability 8.9+ is required. The pinned stack uses Python 3.10, PyTorch 2.6.0, CUDA 12.4, FlashAttention 2.7.2.post1, and FastVideo kernel 0.3.0.

View the complete hardware and software requirements

Will an 8 GB or 12 GB GPU work?

The documented configuration has no published 8 GB or 12 GB path. Its lowest reported FP8 file-inference footprint is 14.4 GiB. Lower-resolution or different-model experiments would need their own validation; they are not promised by this site.

Can I use a Mac or a normal Windows laptop?

The documented path is a CUDA-based Linux GPU environment. Apple GPUs, CPU-only machines, and laptops with integrated graphics are not advertised as supported. Native Windows and WSL2 require additional kernel and camera validation. A lighter computer can access a separate compatible GPU machine through a tunnel.

Can I edit an existing video?

Yes. The documented workflow accepts source videos, preprocessed video folders, and JSON task files for batch processing. A separate reference image can be used with a source video.

Are the previews actual Dash footage?

The portraits are AI-generated concept artwork, and the interactive comparison is a style explorer. The transformation demo is the supplied research demonstration with a redesigned Dash presentation. The animated walkthrough and AI presenter guide explain the published training design using adapted diagrams. These educational presentations do not establish new Dash performance measurements. Actual results and performance depend on the model, prompt, input, and hardware.

When can I buy Dash?

Pre-order Dash Studio for $599.90 USD through Exnode. The desktop app is in development; a delivery date has not been announced, and a pre-order does not provide an immediate download. GPU hardware and cloud rental are separate. The source workflow is currently described as academic research only; the pre-order does not grant a commercial licence to underlying models.

YOUR NEXT CHARACTER STARTS HERE.

Make something
only you could.

REAL-TIME LOCAL AI. A NEW CREATIVE DIRECTION.

Your camera.
Another
reality.