Lucy 2.5 tests real-time live video filters at $0.04/sec listing

• Lucy 2.5 edits live video at 30 FPS and 1080p, according to Decart's launch post, and marckohlbrugge's Yay demo showed typed camera filters changing a live feed in real time.
• fal exposes a WebRTC endpoint and browser playground for Lucy 2.5, and the $0.04 per second price also appears in marckohlbrugge's price reply.
• Creator tests moved past simple style transfer: the first test by venturetwins showed avatar, clothing, and background changes on a webcam feed.
• Reference images are part of the live workflow: venturetwins' Nano Banana follow-up used an edited screenshot as the Lucy reference to preserve room lighting.
• The live identity use case landed first with creators: venturetwins' streamer reply said some streamers want to go live without showing their real face, and carolletta wanted to speak as a cyborg avatar.

fal already lists a WebRTC endpoint, a playground, the model slug decart/lucy-2-5/realtime, and a $0.04 per second price on its Lucy 2.5 page. Decart's technical post names Self-Anchoring as the trick for long streams, while its public Lucy page still carries mixed performance copy: sub-40ms and 100 FPS in one block, under 200ms at 22 FPS in the FAQ. Marc Kohlbrugge's Yay press kit says the current photo app takes about a minute per image, so the live video demo is a large UX jump for the same product.

Decart says Lucy 2.5 generates every frame live at 30 FPS and 1080p, responding to motion, lighting, and scene changes without using pre-rendered video in its launch post.

The shipped editing range is broad:

• Character swaps
• Clothing changes
• Physically aware VFX, including slime, sand, water, and fire
• Environment transformations
• Object removal
• Text prompts and reference images
• Global style changes that preserve structure and identity

Kohlbrugge framed the feature as real-time video filters coming soon to Yay, an iOS AI camera app whose press kit describes still-image filters, credits, and roughly one minute of server processing per photo.

The prompt bar is the interface. In the Yay demo, a live camera feed changes styles after typed prompts, and marckohlbrugge said new styles can be added because "it's just text prompts."

That is the creator-friendly part: a filter library can become a prompt field instead of a fixed preset tray.

The workflow in venturetwins' Nano Banana follow-up had three steps:

1. Grab a screenshot from the scene.
2. Ask Nano Banana to change the person into a different character.
3. Use that result as Lucy's reference image so the stream edit keeps the original lighting.

A second clip from venturetwins ran through restyling, outfit changes, hair changes, environment edits, a 3D animation look, and an alien character.

Decart's named consistency mechanism is Self-Anchoring. Shortly into a stream, Lucy adopts a snapshot of its own generated output as the new reference, then trains and runs under those same self-anchored conditions, according to the Lucy 2.5 post.

Decart says the result is three persistence gains:

• Edits made in the first seconds stay locked in minutes later.
• Subjects can leave frame, turn around, and return without the edit dissolving.
• Identity, geometry, and texture remain stable over longer streams.

Decart names three speed levers in the technical post:

• Low-precision quantization using MXFP8 and NVFP4, with up to 4x speedups on compute-bound operations.
• Dynamic sparse attention, using custom kernels to prune redundant attention work.
• Deeper kernel fusion and larger megakernels to cut launch overhead and HBM transactions.

The public performance numbers are messy. Decart's Lucy page lists sub-40ms latency, 100 FPS continuous generation, and 4x real-time inference speed, while the FAQ on the same page says "under 200 ms" with generation running at 22 FPS at 512×768×3.

The obvious first market is people who want a camera presence without a camera face. In venturetwins' reply, she said lots of people want to stream without using their real face, while another venturetwins reply said existing avatar upkeep had already tested her patience.

The reactions split into live formats:

• carolletta wanted to use it with a cyborg avatar and said cameras felt easier if her face could be changed.
• hasantoxr imagined audience members changing scenes, adding effects, or triggering moments during a stream.
• Uncanny_Harry said a theatre friend had been looking for a real-time AI solution for live performance.

fal's Lucy 2.5 listing says the model is available now through a real-time WebRTC endpoint, with live or recorded video input, text prompts, and reference images.

The implementation details listed by fal:

• Model slug: decart/lucy-2-5/realtime
• Endpoint type: real-time video to video over WebRTC
• Price: $0.04 per second, matching marckohlbrugge's reply
• Client path: @fal-ai/client, Python client, or REST API
• Auth pattern: short-lived JWT token fetched from a backend
• Previous model: Lucy 2.1 remains available on fal
• Commercial use: fal says outputs can be used in commercial projects, subject to its terms

The listing also compares Lucy 2.5 with Lucy 2.1: broader editing range, better reliability on small and non-human objects, improved prompt adherence, reference fidelity, realism, and claimed hours of stable output with no drift or identity collapse.