ElevenLabs Scribe v2 Realtime достигает задержки 150 мс, поддерживает более 90 языков — точность 93,5%.

Higgsfield opened Transitions: creators can now feed both videos and stills to cut between shots using 17 built‑in effects, including Morph, Raven, and Smoke Launch post Transitions app. For the next 9 hours, like/retweet/comment earns 204 credits to test it Launch post.

This removes the old image‑only constraint and targets faster, cleaner pacing for edits and reels; an independent recap frames it as pro‑grade and X‑native for working video teams Recap thread. Early users are already calling out smooth transitions in live tests User reaction.

ElevenLabs launched Scribe v2 Realtime, a low‑latency speech‑to‑text model targeting voice agents, meetings, and live apps, with a stated ~150 ms median latency and coverage for 90+ languages, available today via API and inside ElevenLabs Agents Launch thread and Product page. Following up on Summit agenda, this is the headline release many teams were waiting for.

On accuracy, ElevenLabs cites 93.5% across 30 European and Asian languages, edging Gemini 2.5 Flash at 91.4% and GPT‑4o MiniTranscribe at 90.7%, with Deepgram Nova 3 at 88.4% (see chart) Benchmarks chart.

Operational features include streaming with voice activity detection, manual commit control for when to finalize segments, support for PCM (8–48 kHz) and µ‑law, and optional speaker diarization Builder guide. Enterprise controls cover SOC 2, ISO 27001, PCI DSS L1, HIPAA, GDPR, plus EU and India data residency and a zero‑retention mode Builder guide. For builders, the pitch is simple: faster turns without sacrificing accuracy or compliance.

Runware rolled in two Runway engines: Gen‑4 Turbo at ~$0.25 per 5‑second clip and Aleph at ~$0.75 per 5‑second clip, available by API and in the Playground launch thread, with the model collection live on its catalog Runware models. A follow‑up note frames Gen‑4 Turbo for quick tests and storyboards, while Aleph targets higher‑fidelity looks model note.

For small teams, the fixed per‑duration pricing makes budgeting sprints easier. Turbo is the obvious pick for boards and animatics; switch to Aleph when you lock look and need polish.

fal.ai is now hosting FlashVSR for fast video super‑resolution to 4K, touting sharper details, cleaner motion edges, and fewer artifacts release thread, with a runnable page for creators today Model page. This follows FlashVSR 17 FPS performance chatter, now packaged as a point‑and‑run service.

If your pipeline already outputs 720p/1080p drafts, route finals through FlashVSR to lift perceived quality without re‑generating the whole shot.

Runware posted a detailed Riverflow 1.1 Pro JSON brief for a photoreal arctic explorer shot—covering facial identity constraints, environment, attire, and camera (85mm, mid‑shot)—so teams can reproduce a look consistently across frames or retakes JSON preset. The example frame shows the intended grade and detail targets.

Use this pattern to standardize briefs for recurring characters; it saves time when you switch engines or need to match pickup shots.

Vidu demoed a single‑button "Extend" that pushes a clip from roughly 1–8s to 9–14s in one action, no re‑prompting feature demo. For editors and social teams, this is a quick way to hit platform length targets or test alt endings without rerendering the whole sequence.

Try it on B‑roll or ambient shots first; narrative clips may still need continuity fixes.

A logs screenshot shows models/gemini-3-pro-image-preview-11-2025-dark-launch returning 200, implying internal endpoints are lit and a public image preview could be close Logs screenshot. Following up on Rename Ketchup that flagged NB2 as “Ketchup” in Gemini code, this is the clearest on‑platform signal yet that Google’s next image stack is near. Expanded logs view

Creators surfaced a hidden model label “mandarin” on LM Arena and suggest it’s tied to a forthcoming Grok Imagine image model refresh—positioned as competition for Google’s Ketchup/NB2 and Flux 2 in creator workflows Arena sighting. If confirmed, it would extend Grok’s recent momentum from stylized shorts into higher‑fidelity stills.

A credible watcher says a Flux 2 image model beta is “incoming,” framed to compete directly with Google’s Ketchup/NB2 and xAI’s rumored “mandarin” Grok update Flux 2 tease. If timing holds, creatives should plan side‑by‑side tests across identity retention, style control, and artifact rates as these land within days.

Windsurf quietly introduced three stealth models—Aether Alpha, Beta, and Gamma—now selectable in Windsurf Next and for a small group of Stable users Stealth models note. Some in the thread speculate these map to a GPT‑5.1 line, but treat that as rumor until specs or evals land Speculation thread. For builders, it’s a heads‑up to isolate tests in a sandbox before routing production prompts.

A creator ran a single‑prompt, one‑gen basketball dribble shootout across Kling Turbo 2.5, Veo 3.1, and Grok Imagine 0.9, finding Kling best on ball‑handling physics, Veo stable but conservative, and Grok a credible dark horse comparison notes. Following up on camera moves, this puts concrete motion fidelity stakes in the ground for sports/action shots; Sora 2 was excluded due to new realistic‑face rules.

• Route dynamic ball‑handling to Kling 2.5; keep Veo 3.1 for safer, polished takes; try Grok 0.9 when you want motion ambition on a single pass.

Adobe is rolling out access to Firefly Custom Models (beta), letting creators fine‑tune on their own images, illustrations, and characters; invites are on a rolling basis via the waitlist Waitlist post, with Adobe emphasizing that training assets remain private to your model Adobe waitlist. For brand and character work, this promises faster on‑model consistency than prompt engineering alone.

Leonardo showcased two plug‑and‑play Blueprints: Restyle My Room for instant interior re‑renders and Outfit Inspo that returns three looks from a full‑body photo Room restyle demo, Outfit inspo demo. This follows Font Matcher adding typography recreation, rounding out a practical blueprint stack for brand boards, set design, and wardrobe pre‑viz.

Rory shows a node graph in Weavy that imports the Qwen Edit Multi‑Angle LoRA from Replicate, duplicates nodes with varied angle/tilt settings, and generates a grid of 18 alternate views from a single input image Workflow thread. The setup highlights batchable camera experimentation for look‑dev and coverage without re‑shoots; settings screenshots detail rotation and aspect controls Settings preview.

A creator says they finally completed their “best work yet” by blending two style frames per shot with Leonardo’s Blueprints, then animating the sequence, turning a previously too‑linear cut into a finished piece Case study thread. For teams, this shows how a Remix‑style blueprint can lock look and transitions while you swap start/end frames per beat.

A Munich regional court ruled OpenAI violated German copyright by using lyrics from nine songs to train ChatGPT and ordered damages; OpenAI plans to appeal Ruling summary. The decision treats both storing protected lyrics in model weights and reproducing them in outputs as infringement, a signal that EU licensing pressure on training data will intensify for creative AI Case details.

The Wikimedia Foundation asked AI companies to stop scraping and license content via Wikimedia Enterprise, saying bot traffic masquerading as humans coincided with an 8% YoY drop in human pageviews; it also asked for clear attribution to contributors Policy summary. For teams building RAG or pretraining pipelines, budget for API access and provenance—or risk rate limits and reputational pushback.

A creator’s side‑by‑side sports‑motion comparison excluded Sora 2 because its newer policy disallows realistic reference faces, preventing apples‑to‑apples tests against Kling and Veo Benchmark comparison. Expect tighter likeness guardrails to affect commercial workflows that rely on real‑person inputs or client‑provided casts.

Higgsfield released a Face Swap Browser Extension that lets you drop your face into any image you see online, right from the page. For the next 9 hours, follow + retweet + comment to receive 202 credits via DM, following up on face swap tips where creators shared practical workflow advice. See the announcement and link to the installer in the launch post Launch thread and the product details on the site Product page.

Early replies show playful adoption and brand tone (“best upgrade your face has ever met”) as users test swaps in the wild Playful reply. For creatives, this moves identity control into the browser, speeding comps, memes, and mockups without round‑tripping through separate apps.

ElevenLabs is formalizing consented celebrity voice use with its new Iconic Voice Marketplace, where projects route through talent approvals instead of scraping or unauthorized cloning. Variety reports Matthew McConaughey and Michael Caine are onboard to replicate their voices; McConaughey also invested and is using the tech for a Spanish newsletter audio track, while Caine is cleared for narration and other approved uses Variety report, Marketplace details.

For creative teams, this turns high‑profile voices into licensable assets with workflow guardrails (approvals, rights management) rather than ad‑hoc deals.

On Veterans Day, ElevenLabs highlighted Lt Col Thomas Brittingham, a pilot living with ALS, who recovered a personalized speaking voice through its Impact Program—showing a restorative, non‑synthetic use case that matters to filmmakers and storytellers working with real people and sensitive narratives Impact story, Impact blog post. The company’s summit schedule also features dedicated segments on veterans and social impact, signaling continued investment in this lane Summit schedule.

This is a concrete pattern: licensed voice tech isn’t only for famous IP; it can enable documentary work, accessibility, and character‑driven projects with consent at the center.

Adobe quietly rolled out Generative Soundtrack (beta) inside Firefly: upload a clip, tap Suggest Prompt, tweak vibe/genre or write your own, and it composes music to the exact duration of your edit Feature demo. This targets editors who need fast, on‑brand beds for promos, shorts, and social without hunting stock or hand‑timing cuts.

The flow looks production‑friendly: worded prompts, preset vibes, and automatic length matching reduce back‑and‑forth on timing. Missing details to watch: licensing scope for commercial use, stem/loop exports, and hand‑off into Premiere/After Effects. If you cut short‑form, this could shave minutes per edit and keep pace with rapid versioning.

This adaptable prompt frames wind‑eroded ruins, bones, and heat‑blurred horizons; plug your [SUBJECT] and color accents [COLOR1]/[COLOR2] to keep the haunting vibe consistent across shots Prompt post.

Today’s V7 preset lands a glossy editorial collage look: --chaos 33 --ar 3:4 --raw --sref 3394984291 --sw 500 --stylize 500, following up on MJ V7 collage where an earlier sref was shared Param recipe.

A creator shared a reusable Midjourney style reference that nails late‑80s/90s cinematic anime (Ghost in the Shell/Akira energy) with a modern polish: use --sref 1055666590 to lock the art direction Anime sref post.

Higgsfield is running platform‑native prompts: reply anywhere on X with @higgsfield_ai + #nanobanana + your prompt to get an auto‑generated image; they’re offering 250 credits for follows for 12 hours How-to thread.

For saturated, sun‑kissed editorial scenes, stack --sref 639275414 2854356852 with --profile jls63zz --stylize 1000 --exp 33 --raw to keep motion blur streaks and color pop Preset recipe.

Midjourney tends to pretty; this moodboard pushes glitchy, feral looks. Use --profile jls63zz with --stylize 1000, then iterate on composition to retain the grit Moodboard post.

A distinctive character recipe to reuse in sets and series: a rough burlap cube head, square eye cuts, and a dry‑leaf headdress against soft fall bokeh—great for surreal portraits and fashion stories Look card.

A tight, reproducible character brief: “1980s, model in a red one‑piece swimsuit, she tips down her sunglasses and peers over them looking intensely into the camera.” Grok matched the vibe on a single generation One-shot prompt.

Rainisto shared a compact, observational interior scene—muted palette, slow lensing, and a held beat that feels lived‑in Scene post. Good reference for domestic lighting and pacing.

A creator cut a fast‑paced reel branded “Sabotage,” flashing code, distorted type, and red‑lit frames; Kling is credited in the share path Short film post. If you’re testing type and UI as characters, this shows a workable motion language.

James Yeung’s particle‑driven butterfly loop was shared with a clean upscale pass via Topaz Bloom, showing how light FX hold up after enhancement Loop preview. The follow‑up clarifies the source image lineage before upscale Source note.

James Yeung’s particle‑driven butterfly loop was shared with a clean upscale pass via Topaz Bloom, showing how light FX hold up after enhancement Loop preview. The follow‑up clarifies the source image lineage before upscale Source note.

Billywoodward tested Grok on a hand‑drawn frame from a prior film project, nudging style while preserving layout and motion intent Look test. If you’re blending original art with AI motion, this is a practical reference.

A prompt‑driven, overhead desert convoy with explosions shows WAN 2.2 tracking dust, missiles, and vehicle drift with readable geography in a short runtime Prompt and output. Useful for testing top‑down blocking.

A quick PolloAI clip pushes into a bathroom mirror, using a tight zoom and eye contact to sell intimacy without dialog Mirror study. Strong example of micro‑story via camera distance alone.

ProperPrompter posted a compact choreography study—close cuts, city backdrop, and a clean motion arc that reads even without dialogue Choreo clip. It’s a good template for action beats under 10 seconds.

A short comedic bit cuts from a mundane elevator entrance to an alien reveal beat—simple shot list, clear twist, and it works Comedy clip. This is the kind of structure AI shorts can repeat well.

A playful sloth “back to work” sting shows how a single animated beat and one gag line can carry a social‑length post Short bumper. Consider this structure for episodic channel idents.

A new ComfyUI node, PainterI2V, gives WAN 2.2 far stronger camera prompt control and claims a 15–50% boost in overall motion responsiveness for shots like push‑ins and moves node announcement. For creators, this means fewer re‑blocks and more reliable camera behavior from text prompts.

Use it to tighten drone‑style push‑ins or gentle dolly moves when WAN ignores subtle camera cues node announcement.

A community Motion LoRA trained over ~40 iterations adds a clean, repeatable “Push‑in camera” effect to WAN 2.1 via a simple trigger word, useful for story beats and product shots lora release. It’s a fast way to layer controlled forward motion onto otherwise static, high‑quality frames without re‑framing the scene.

ElevenLabs’ San Francisco summit kicks off today with a full-day agenda and on-stage product updates. Sessions include a Salesforce AI talk, a startup grant showcase, a veterans segment, and a closing conversation with Mati Staniszewski and Jack Dorsey Agenda image. This matters if you build with voice and real-time agents. Expect demos, customer case studies, and roadmap hints.

Following up on agenda, the schedule now lists exact slots: keynote at 10:00, will.i.am and Larry Jackson at 1:45, and a Jack Dorsey chat at 4:30 Agenda image. Plan your watchlist accordingly.

Dreamina opened a holiday creative brief: remix the iconic turkey using Dreamina across Instagram, TikTok, or X for a chance at credits, subscriptions, and custom certificates; 30 winners total, submissions close Nov 16 (PT) Challenge details. Top entries will be cut into a global highlight film called “Creative Turkeys We Love.”

Good prompt practice for stylization and brand-ish composition. If you’re new to the tool, join the community first.

<u>Discord</u>: Discord

Hailuo launched a horror short competition with tiered credit prizes and a simple brief: submit a high‑quality 10s–2min video primarily made in Hailuo, watermark on, by Nov 30; winners announced Dec 10 Contest post. Gold winners earn 20,000 credits; Silver 8,500; Bronze 3,500; Spotlight 500 Contest page.

If you’re testing i2v look dev or motion beats, this is a clean way to get feedback and credits back into your budget. The promo reel shows the target mood and pacing.

A creator asserts average consumers judge outcomes, not whether AI was used, urging artists to optimize for final quality and not the method debate. The point is sharpened by a follow‑up that the end result can be bad with or without AI—audiences still file it under "computer stuff" either way Consumer comment, Follow‑up view.

Two fresh posts from a prominent creator reignited the AI‑art culture war, pulling in 300+ likes combined within hours. Following up on creativity value, the new claims argue blanket dismissal of AI art is prejudice and that critics’ drawings don’t stack up. See the latest volleys in Prejudice claim and 99.99% claim.

Credit disputes and provenance confusion popped again: one creator hunted the original artist for a widely shared image, while others felt compelled to state pieces were AI‑made, not Blender, due to realism. It’s a reminder to label, link, and credit before the pile‑on starts Attribution ask, Provenance note, Blender spoof.

After being labeled a “pro‑AI bot,” a creator countered with follower counts and account age receipts to argue authenticity, highlighting how the debate keeps sliding into identity attacks instead of work critique Bot thread.