Tool01
Runway Agent leads a 30-prompt human evaluation
An independent human evaluation tested six AI-video agents on 30 cinematic prompts and 16 metrics; Runway says its Agent ranked first across Narrative Coherence, Cinematic Language and Production Quality. A shared brief and multidimensional scoring frame are more useful than isolated demo reels because they begin to separate directorial reasoning from a polished model output. Studios can borrow the test design for procurement: run shortlisted agents on the same real client brief, score continuity, shot intent and delivery readiness separately, then keep the scorecard beside the cost and revision log.
Source Runway @runwayml · 2026.07.17 ↗
Tool02
Maestro adds local automatic character recasting
Maestro v1.3.0 introduced Recast, a local workflow that identifies a described person in a clip, masks them automatically and replaces them from a character image. Its bundled SCAIL-2 Fast distillation runs in six steps without CFG and is reported to be about 13 times faster than the stock model. That makes character replacement testable without sending footage to a hosted service; teams can reserve it for approved-source plates, compare identity stability through occlusion and motion, and price cleanup separately when the automatic mask fails.
Source Blaine Brown @blizaine · 2026.07.17 ↗
Technique03
Higgsfield exposes the making of its Originals
Higgsfield opened the prompts, references, assets and settings behind shots from its Originals, including timed multishot direction, camera constraints and continuity rules. Releasing the production recipe turns a polished output into something peers can inspect, adapt and challenge, while also revealing how much orchestration sits outside the underlying model. Directors can use the material as a shot-spec template—separate references, action beats, camera behavior and hard exclusions—then version every change instead of treating one enormous prompt as an opaque final answer.
Source Angry Tom @AngryTomtweets · 2026.07.17 ↗
Case04
Omni turns live video into a manipulable surface
fofr prototyped an Omni interface for interactive live video generation with direct manipulation rather than a prompt-only control loop. The experiment points toward a different creative surface: intent can be expressed by touching and moving the image while generation continues, narrowing the gap between art direction and model instruction. Product teams should test this interaction first on spatially explicit tasks—moving a subject, changing a region or redirecting motion—and log whether direct control reduces prompt retries without introducing temporal instability.
Source fofr @fofrAI · 2026.07.17 ↗
