MiniMax H3 unifies video and audio; Seedance 2.5 moves toward every major creation platform
MiniMax H3 ships as an open, omni-reference video modelLuma adds Seedance 2.5 to its incoming model shelf
AI visual creation, as it happens.
The releases, tools, work, and techniques changing how ideas become visual—curated from the creators and labs we follow, with our read on what is actually worth attention.
Daily briefs follow immediate changes; weekly digests connect the larger picture. Every note separates the fact, why it matters, and what it changes for visual work, with a link to the original source.
Daily briefs
Luma turns flat images into editable layers; Lucy 2.5 edits live video in real time
Luma Agents extracts editable layers from any imageLucy 2.5 makes prompt-controlled video editing real time
A two-person team ships a finished AI ad; Scenario turns one mesh into editable semantic parts
Two creators finish an authored AI ad, not a model reelRodin Hyper3D Bang splits a finished mesh into semantic parts
Seedance 2.5 reaches the platform layer; World Labs builds generated worlds for robot learning
Runway previews Seedance 2.5 on its multi-model platformWorld Labs turns generated worlds into robot training environments
Artlist puts generation inside Claude; PixVerse uses depth maps to hold motion and space
Artlist MCP lets Claude generate into a persistent asset accountPixVerse Depth Map Control transfers motion with spatial structure
Video agents get a 16-metric benchmark; world models and FPV prompts test continuous control
An independent benchmark scores video agents across 16 production metricsOdyssey keeps a generated world alive beyond one revealA PixVerse FPV recipe threads one camera move through sailboats
AI 3D moves from demos to studio conversations; spatial intelligence becomes an infrastructure category
Tripo closes SIGGRAPH around production-ready AI 3DSpatial intelligence enters the AI infrastructure conversation
Runway Agent edits node workflows by language; Midjourney V8.2 makes aesthetics the release target
Runway Agent builds and edits node workflows through natural languageMidjourney V8.2 targets aesthetics, personalization and image qualityLuma packages a full character sheet as a reusable Skill
FLUX 3 unifies media and action; a 300-shot AI film shows the real continuity bill
FLUX 3 spans image, video, audio and action predictionLuma demonstrates a cast that feels persistent beyond one frameOVERGROWN reaches 13 minutes and more than 300 shots
Kling teaches complete MCP filmmaking; DaVinci Resolve adds finer retime curves
Kling publishes an end-to-end MCP filmmaking workflowCurious Refuge packages its YouTube AI workflow as a courseDaVinci Resolve 21.0.3 adds finer retime and curve controls
Seedance 2.0 holds 15+ shots on three prompts; Gemini 3.6 Flash trims token cost
A three-prompt Seedance 2.0 workflow keeps 15+ shots consistentGemini 3.6 Flash delivers higher quality at the same cost
A named director ships an AI film; Nano Banana Pro edits by instruction
Neill Blomkamp releases an AI-made filmNano Banana Pro edits an image by written instructionOne-shot aerial sequences from a single Seedance 2 prompt
Seedance 2.0 turns a football match into broadcast-style highlights
A Seedance 2.0 sports-highlight montage, with the full workflowPrompting Seedance 2.0 for authentic handheld home-video texture
Video agents face a shared test; local recasting gets a faster path
Runway Agent leads a 30-prompt human evaluationMaestro adds local automatic character recastingHiggsfield exposes the making of its OriginalsOmni turns live video into a manipulable surface
AI post moves into the timeline; an agent holds continuity across 14 shots
Magnific lands inside four editing suitesAgent One keeps a 2,000-word brief coherent for 14 shotsHiggsfield packages a four-stage cartoon studio through MCPHiggsfield opens a 100,000-credit filmmaker grant
Higgsfield chains four models into one MCP pipeline and opens a free academy
An MCP pipeline chains story, art and animation into one runA free AI-filmmaking academy with certification
Higgsfield funds filmmakers; Pika edits footage by prompt; DeepMind turns video into 4D
Higgsfield opens a 100,000-credit Filmmaker GrantPika turns footage into 'anything' by prompt via Gemini OmniDeepMind's GenCeption turns video into searchable 4D
An AI feature trailer drops; GPT Image 2 beats Meta's new model on consistency
Fountain 0 trailers a second AI feature, ODYSSEUS: The FallGPT Image 2 beats Meta's new image model on character consistencyRunway Agent gets a short-film how-toClaude drives After Effects through Higgsfield's Supercomputer
AI ads win at Cannes on Kling; Adobe makes Firefly conversational; precise camera-and-action control lands in Seedance
Two AI ads win at Cannes, both built on Kling — the maker breaks down whyAdobe makes Firefly conversational: one chat runs the whole creative workflowPrecise camera-and-action control, prompted into Seedance 2.0Luma leans on Skills — a reusable creator capability as the unit of workA full history documentary, cut by an agent (invideo agent one)
Kling's short-film circuit rewards authored work; Sol rebuilds San Francisco in 3D voxels
A Kling contest film treats perspective as the subject, not the effectSol Ultra reconstructs San Francisco as a 3D voxel cityMultiple frontier models coordinate inside one shared Slack channel
Prompt-based video editing lands in Pika; Meta enters video with Muse; the director-agent becomes a category
Pika wires in Gemini Omni for prompt-based video editingMeta enters pro AI video with Muse Video (not public yet)OpenArt Director makes the director-agent a category, not a productBenchmarking the director-brain: GPT-5.6 Sol vs Fable 5 driving SeedancePrompt length is not the lever — a built-in self-review loop is
One agent directs an entire AI sitcom; Runway Dev turns into a multi-vendor model shelf
An all-AI office sitcom, directed through a single agentRunway Dev stocks rival labs’ models on one enterprise shelfBorrowing a video model as a motion proxy to recover full 3D poseChaining an image model into video to fake 1960s documentary footageA creator vibe-coded a full video editor on top of Higgsfield
Pika opens its agent-run Director’s Suite; Runway ships a dev platform and folds in Seedream 5.0 Pro
Pika opens Director’s Suite, an agent that runs the whole video projectSeedream 5.0 Pro lands in Runway with legible in-image text in 14 languagesRunway launches Runway Dev, an AI media platform for developers and enterprisesPrompting Seedance 2.0 into documentary-grade historical footageKling unveils its NEXTGEN Awards winners, from 4K cinema to campus shorts
Meta enters the image race with Muse Image; an end-to-end 4K AI music video ships on Higgsfield
Meta Superintelligence Lab drops Muse Image, an agentic image modelAn end-to-end 4K AI music video: Fable 5 plans the choreography, Seedance 2.0 rendersA full SnorriCam sequence from a single image, no rigClaude Fable 5 access extended to July 12
Runway opens a Paris research hub; agent-built previs for Seedance takes off
Creators use Claude to build pose/depth previs rigs for SeedanceRunway opens its first office in France, a Paris research hubAnime football clips settle into a Magnific-to-Seedance recipeMidjourney 8.2 preview renders start circulatingGPT Image 2 plus Seedance makes tiny-world edits easyKling's NEXTGEN Awards ceremony lands in Seoul today
Weekly digests
Creative tools become agent-callable, while images and worlds become structurally editable
Week 31 shifted the unit of production from a generated clip to an editable system: MCPs moved creation into agents, FLUX 3 unified modalities, depth and layers exposed structure, and world models were judged by interaction rather than a single rendered view.
- IndustryMCP becomes a standard front door for creative platforms
- ModelFLUX 3 unifies image, video, audio and action prediction
- TechniqueEditing exposes structure: layers, depth and semantic parts
- IndustryWorld models move from scenery to persistent interaction
- IndustryVideo-agent benchmarks start measuring direction, not isolated clips
Seedance 2.0 becomes the workflow center; Higgsfield builds a whole stack around filmmakers
This week the craft news converged on one model and one platform strategy: Seedance 2.0 turned into the shared base for consistency and montage workflows, while Higgsfield stacked a grant, a free academy and MCP tooling around AI filmmakers. Alongside them, named directors and AI features kept normalizing authored AI work, and the model layer under all of it (image and Flash) kept shifting.
Meta joins the frontier, the director-agent becomes a category, and motion goes subject-agnostic
Week 29 turned last week’s experiments into shape: Meta entered pro AI video with Muse, the director-agent multiplied from one product into a category, platforms became aggregators while editing collapsed into a prompt, motion control went subject-agnostic with a video-model pose proxy, and craft consolidated around capture recipes and self-checking prompts.
Google ships Omni Flash to the API, Kling takes three Cannes Lions, and agents start directing the shots
A dense week: Google put video generation and editing into every developer's hands, AI-made films won at Cannes Lions for the first time in a dedicated AI Craft category, and the frontier moved from writing prompts to letting agents run the whole pipeline.