☕ Affogato Competitor & Trend Digest

At 1:37am IST @krea_ai posted 'about to drop. sharing some early access codes below' (221 likes) with five K3-EARLY codes and a redemption page at krea.ai/k3-access. Earlier it asked 'fast, precise, and controllable. who else wants early access?' (318 likes). As of this run, krea.ai/krea-3 still reads 'Krea 3: Frontier Image Editing Model, 4x faster than Nano Banana Pro, Coming soon' and the Krea changelog's latest entry is Oct 8. Third parties are ahead of the official launch: the varg.ai changelog (Oct 13) says Krea 3 is available through its API as krea_3, free during early access, with up to 8 reference images, fast/high/ultra thinking modes (fast returns a 1K image in about 5 seconds), 1K or 2K output and one image per request; fal lists Krea 3 in early access with an edit endpoint taking 1 to 4 references. Speed and quality claims are Krea's or early testers', not independent benchmarks.

BFL says serving optimizations let it make FLUX 3 Video up to 2x faster and cut the price 50% for native 2K and 4K generation (408 likes, about 19K views). A follow-up claims '4K quality that outperforms Kling o3, at comparable pricing and latency'. That comparison is BFL's own. BFL's public release notes still show Oct 1 (FLUX 3 Image) as the latest entry, so the cut is announced on X and the pricing page rather than in the changelog.

'We asked Claude how good it is at motion design. A few hours later, we got a finished launch video.' Higgsfield says Katana for Motion Design is available now in Claude through the Higgsfield MCP (895 likes, about 185K views, the most-viewed company post in this window). Katana launched Oct 8 as a Claude skill for Aura Farming, Motion and Clipping edits from up to five reference images; this extends it to motion-design launch videos.

Higgsfield's changelog gives Pro-and-above subscribers seven days of unlimited Wan 3.0 and Wan 3.0 Prime (clips up to 30 seconds at up to 1080p, shared queue) and 60% off the same model for all API users until Oct 31, 2026.

Runway partnered with Baseten so enterprises can deploy Gen-4.5 and Aleph 2.0, including fine-tuned versions trained on their own data, in dedicated single-tenant environments with reserved GPU capacity. Baseten's Oct 14 post says Aleph 2.0 edits existing or generated footage and supports multi-scene videos up to 30 seconds at 1080p.

MiniMax says Design Agent V3 is live with 30%+ lower creation costs, asset and context reuse across chats on the same creation page, and customizable agent instructions. Members can claim 1,000 free credits a day from Oct 13 to 16 (up to 4,000). MiniMax Design is a desktop app (macOS and Windows) that orchestrates text, image and video models on a node canvas from brief to final cut.

Vidu is running 30% off all Q4 Preview credit packages on its API platform and highlighting direct audio-video generation, with reference workflows that accept up to 3 audio references. Its 'Wuxia slide transition' template post reached about 55K views; Q4 Preview is still invite-code gated (VIDUQ4PV5).

Mirelo's video-to-sound model keeps dialogue by default, removes non-dialogue audio such as music nobody prompted, and generates synced foley, effects and ambience that stay coherent across cuts for videos up to 10 minutes. It is live in Mirelo's playground, editor, API and MCP, on fal, and Mirelo lists Runway as a partner platform. API price is 15 credits per second per variant, discounted to 10 until Dec 12, 2026. Mirelo's own blind test reports listeners preferring it over Sony Woosh (81%) and Kling Foley (63%). Justine Moore (a16z) amplified it (136 likes).

HeyGen launched its in-house HeyGen Voice on Oct 9 and published a research write-up on Oct 14. It speaks a script in the user's cloned voice and adapts to the acoustics of the scene on screen. HeyGen says it ranked #1 on Artificial Analysis' controlled-voice leaderboard (Elo 1201, Oct 9 snapshot). It is free inside HeyGen videos; a Professional Voice Clone add-on costs $99 a month. HeyGen's Oct 14 X push got 921 likes, and Heather Cooper (@HBCoop_) highlighted the room-matching.

invideo says Claude generates the assets, codes the motion graphics, and builds a fully editable timeline in the invideo editor, with no After Effects needed. A YouTube setup video and invideo.io/mcp accompany the post.

@emmanuel_2m, posting for Scenario, announced an open-source Blender plugin billed as 'an entire 3D team' (concept artist and other roles), usable inside Blender or via Claude Code through MCP (2,855 likes). A day earlier Scenario gave away a free Opus 5.5 + Seedance 2.5 auto-edit skill. Theoretically Media (@TheoMediaAI) quote-posted it with plans to test it.

Manus's blog post 'Introducing Manus for Business' launches Team and Enterprise plans aimed at making whole teams 'AI native', for example by triaging Slack feedback and checking a codebase before suggesting support fixes. It was promoted on X on Oct 14.

OpenArt named Divyeshu Sinha as Ad of the Year ($15,000), with $3,000 category prizes for brand commercial, fashion and beauty, product ad and UGC ad, $5,000 model awards from Seedance, Wan and MiniMax, and partner awards from Machine Cinema and Wonder Studios. The sponsor list shows Chinese model labs paying to win ad-creative mindshare on aggregator platforms.

Kris Kashtanova (@icreatelife) posted 'I can't stop playing with the new AI Relight feature in Photoshop (beta)'. Relight shipped in Photoshop beta build 27.12.0 on Sep 28: up to three non-destructive lights, free previews, and 30 generative credits for each final render. Fstoppers' Oct 14 review found the rendered result falls short of the original file's resolution. Adobe Firefly also posted a tutorial on using Generate in Firefly from Lightroom to remove a person and create a new portrait.

Recraft posted a V4.1 range showcase (frozen lakes, surreal landscapes, objects, wildlife) with prompts in the replies. It was low engagement (13 likes) and is not a new release.

Midjourney (latest is the Oct 10 MCP tester call and Oct 8 thinking-mode test), Ideogram (latest Oct 12), Kling (latest Oct 12; Kling 4.0 is 'launching soon' per its Oct 7 post), Luma (latest Oct 13), Pika (latest Oct 7), Magnific (X account inactive since April; API changelog latest Oct 8, Magnific One), Dreamina (latest Oct 12), Leonardo, Stability, Hedra and Noiz.

Pollo AI announced Pollo Recast, which recreates any video with custom characters, products and scenes. That is the same 'recast a reference clip with my product' job Higgsfield's Genjutsu targets. Medeo announced Creative OS, taking AI video from trend discovery to monetization in one workflow (it launched VideoClaw, a viral-trend recreation agent, on Oct 1). PearlMountain launched CutBay, a chat-based AI video agent. Only the PR Newswire headlines could be read; the full release text was blocked, so treat feature details as the companies' own claims.

Startup Fortune ran a piece saying Seedance 3.0 extends single clips to 30 seconds and accepts up to 50 references (30 images, 10 videos, 10 audio). A same-day analysis on For You found resellers giving conflicting specs and cites an API marketplace saying there's no confirmed official release or public API. Dreamina's X account (latest Oct 12) still promotes Seedance 2.5. Treat this as rumor until ByteDance Seed or Dreamina announces it.

HeyGen (@HeyGen) and Mirelo (@MireloAI) were added for AI-video voice and sound. Scenario (@Scenario_gg) was added for game and 3D asset pipelines plus MCP tools. Medeo (@Medeo_AI) and Pollo AI were added for agentic video recreation; Pollo's X handle isn't verified yet (pollo_ai and PolloAI_ returned nothing).