# AI News, 23 October: editable artwork, shot planning and faster models

Ideogram Vector, CozyClay and Colourlab update creative workflows. EdenLive and Mercury 3 tackle live rendering and response speed.

Published: 2026-10-23. Updated: 2026-10-23. Author: XANTRANS editorial.

Today’s additions cover editable artwork, controllable shot planning and Colourlab’s public access update. EdenLive and Mercury 3 target rendering and response speed. Release dates differ; none has been independently tested by XANTRANS.

## EdenLive renders the scene that another system controls

VAST AI Research released EdenLive on 18 October with 5B and 14B model weights, streaming inference code and a playable example. The renderer takes a reference video, a target first frame and a text prompt. The reference controls where subjects are, how the camera moves and when actions happen. The first frame and prompt determine the new appearance. A rough 3D game render can therefore drive a more detailed image while the game continues to own its rules and object positions. This is a different approach from asking one video model to invent both the action and every frame.

The useful separation is state from appearance. Imagine an interactive gallery scene in which a visitor opens a door, leaves the room and returns. A game engine can keep the door state and camera path exact; EdenLive can render those incoming frames in a chosen style. VAST also describes recorded footage, simple animations and normal maps as possible references. Its real-time mode works in streaming chunks using only the reference frames received so far, while an offline mode can look across a complete sequence. The release includes an EdenGolem skill that creates a rough playable world from a gameplay description, though that is a separate upstream step rather than evidence the renderer understands arbitrary world rules.

VAST reports continuous 720p output at 24 frames per second or higher for its 5B model on a single RTX 5090. That is a maker demonstration, not a benchmark of an ordinary creator laptop or a production stream. The code documentation lists a verified Linux and NVIDIA setup, and the complete weight tree is about 210 GB. The model card lists different licenses by component: some weights are Apache 2.0, the vision encoder follows DINOv3 terms, and the optional first-frame workbench uses a FLUX component restricted to noncommercial, nonproduction use. Anyone considering a commercial product needs to check the exact component path and rights before deployment.

For a live interactive video project, EdenLive deserves attention if a reliable reference stream can be supplied. It does not replace a text-directed stream by changing a model name: an upstream scene or camera feed, first-frame preparation, GPU hosting and distribution path are additional work. VAST has published code and demonstrations, but we have not independently run the model, measured first-frame delay, checked visual continuity or calculated effective cost.

## Mercury 3 puts diffusion text generation behind an API

Inception launched Mercury 3 on 19 October as a diffusion language model available on its own API platform and through partners including OpenRouter, Baseten and Vercel. Its calls are described as OpenAI API compatible. Instead of producing an answer strictly one token after another, a diffusion language model refines a draft across multiple tokens. For an interactive assistant that has to answer repeatedly during a live session, the practical question is whether that model can shorten the pause before a useful response without making the directions less reliable.

The company lists a one-million-token context window and regular API prices of $0.20 per million input tokens and $0.70 per million output tokens. It also advertises a launch discount of 50%, but the announcement does not state a cutoff date, so buyers should confirm the current price before relying on it. Inception reports 1,068 output tokens per second on NVIDIA hardware, along with comparisons against other fast models. Those are vendor figures obtained under its serving conditions. Tokens per second alone omit network time, prompt processing, tool calls and the quality of the final answer.

A concrete evaluation would send the same bounded set of audience directions or agent tasks to Mercury 3 and the existing model, then score whether each answer follows constraints, how long it takes to become usable and what the entire task costs. A fast wrong instruction can make an interactive story worse. Inception cites early use in voice support, tutoring, search and browser agents; these show intended workload types, not independent proof that it suits every creative assistant. API availability makes a small comparison possible, but we have not made a paid call or measured an application-level gain.

## Ideogram Vector turns a prompt or image into editable artwork

Ideogram announced its first vector model on 22 October. Vector 1.0 generates SVG artwork from a prompt, optionally guided by reference images, or converts an existing image into vector shapes. Unlike a flattened picture, an SVG can be scaled and its individual elements recolored or repositioned in a vector editor. A designer preparing a label and a large shop sign could start from one editable illustration, then adapt its colors and composition for each format.

The model is available through Ideogram and its API. The two workflows use separate generation and vectorization endpoints. The image converter accepts JPEG, PNG or WebP files up to 50 MB, with a simple mode for fewer shapes and a default mode that retains more detail. One important limitation is explicit in the API documentation: grouping converted shapes into layers is not yet available. Editable paths therefore do not guarantee a neatly organized design file. We have not tested typography accuracy, path quality or print output; check the current credits and API quote before generating.

## CozyClay updates the blocking stage before AI video

CozyClay is an open-source browser studio for staging characters, timing motion and planning cameras before generating a finished video elsewhere. Its 23 October changelog lists versions 3.1.0 through 3.1.2: better hand contact with props, throw-and-catch actions, new sculpt shapes and materials, and improvements to the agent repair loop. The agent writes a motion specification; a browser compiler calculates the take and checks it. These are documented changes, not our assessment of animation quality.

For a product scene, a creator could block a character picking up a box, choose the camera and export reference frames or a graybox clip before paying for video generation. Local installation requires Node.js 22.19 or newer and a Chromium browser, without a model download or GPU requirement for the studio. The code uses AGPL-3.0-or-later; connected AI services have their own access and costs. Existing projects need care: version 3.1.0 replaces the old Chair, Car and Plane set pieces with cubes. We have not installed or exercised this release.

## Colourlab opens public access through Houston

Colourlab’s official 22 October presentation announced public access and Houston, a hub for installation, updates, shared looks and subscriptions. The presenter said the grading application had changed little: this is an access and workflow update to the earlier beta. A colourist can manage the software and discover looks in one place before grading a Resolve sequence.

The current offer lists a 14-day trial, $49 monthly, $399 annually or $799 perpetual. Houston requires Apple Silicon and macOS 14 or later, with Apple or Google sign-in. Windows and RAW support remain forthcoming. The documented Resolve Studio MCP workflow fetches, balances and pushes clips, but pushing overwrites affected grades. We have not tested it.

## Sources

- [VAST AI Research: EdenLive release, 18 October 2026](https://www.tripo3d.ai/research/edenlive) (checked 2026-10-23)
- [VAST AI Research: EdenLive source code](https://github.com/VAST-AI-Research/EdenLive) (checked 2026-10-23)
- [VAST AI Research: EdenLive model card and component licenses](https://huggingface.co/VAST-AI/EdenLive) (checked 2026-10-23)
- [Inception: Introducing Mercury 3, 19 October 2026](https://www.inceptionlabs.ai/blog/introducing-mercury-3) (checked 2026-10-23)
- [Ideogram: Vector announcement, 22 October 2026](https://x.com/ideogram_ai/status/2113300317584720259) (checked 2026-10-23)
- [Ideogram: Vector model](https://ideogram.ai/models/vector/) (checked 2026-10-23)
- [Ideogram: Generate SVG API](https://developer.ideogram.ai/v2/api-reference/vectors/generate-svg) (checked 2026-10-23)
- [Ideogram: Vectorizer API and grouping limitation](https://developer.ideogram.ai/v2/api-reference/vectors/image-to-svg) (checked 2026-10-23)
- [CozyClay: 23 October changelog](https://github.com/NomaDamas/CozyClay/blob/main/CHANGELOG.md) (checked 2026-10-23)
- [CozyClay: installation, workflow and license](https://github.com/NomaDamas/CozyClay) (checked 2026-10-23)
- [Colourlab: official public launch presentation, 22 October](https://www.youtube.com/watch?v=GAqSYwSH3uw) (checked 2026-10-23)
- [Colourlab: Houston access and system requirements](https://colourlab.ai/download/) (checked 2026-10-23)
- [Colourlab: current trial and pricing](https://colourlab.ai/pricing/) (checked 2026-10-23)
- [Colourlab: current MCP workflow and grade overwrite warning](https://colourlab.ai/technology/claude-mcp/) (checked 2026-10-23)
- [Colourlab: product access and platform limitations](https://colourlab.ai/products/colourlab-ai-4/) (checked 2026-10-23)

Prepared with AI assistance from the linked primary sources. Maker demonstrations and proposed comparisons are not independent XANTRANS tests.


Original page: https://xantrans.com/blog/ai-news-2026-10-23-edenlive-mercury/