New models and features in AI image, video, voice and music generation, from Midjourney, Sora, Veo, Suno and more.
OpenAI just unveiled DALL·E 3, its next-generation image generator built natively inside ChatGPT. It follows complex text descriptions with nuanced accuracy without requiring prompt engineering tricks, while blocking requests for public figures and living artists' styles. Generated images belong to users and can be commercialized without permission.
Stability AI just launched Stable Audio, its first generative AI product for music and sound effects. Built on latent diffusion models trained on licensed AudioSparx tracks, it renders 44.1 kHz audio from descriptive text prompts and duration limits. A free tier generates up to 45 seconds of audio, while the Pro tier offers 90-second commercial tracks.
The creators of Würstchen just released a lightweight text-to-image diffusion model that drastically reduces training and inference costs. By pushing latent spatial compression to an unprecedented 42x, it generates images faster than SDXL while slashing training compute down to 9,000 GPU hours compared to SD 1.4's 150,000. It makes training and running custom image models accessible without a farm of A100s.
Adobe is rolling out an annual bonus for Adobe Stock contributors whose assets were used to train its Firefly generative AI models. Payouts factor in total approved submissions and license volume over the past year, and users can now submit compliant Firefly-generated art directly back into Stock. It marks a rare effort to share generative AI upsides with dataset creators.
Ghostwriter, the anonymous creator behind the viral AI Drake track, just dropped a new song mimicking Travis Scott and 21 Savage while submitting his previous AI hit for Grammy consideration. Recording Academy leadership confirmed the track is creatively eligible because a human wrote it, though commercial distribution hurdles remain. The music industry can no longer dodge the AI voice controversy.
Ideogram just opened its foundation image model, Ideogram 0.1, to everyone for free. The model tackles one of generative AI's biggest flaws by rendering clean, coherent text directly inside images alongside various style presets. Why it matters: It turns messy AI typography into a reliable tool for designing logos, posters, and graphic messaging.
Stability AI just partnered with NVIDIA to optimize SDXL using TensorRT, releasing open ONNX weights globally. The integration doubles performance on NVIDIA H100 GPUs, spitting out HD images in 1.47 seconds with a 41% latency drop and 70% throughput jump. Why it matters: high-end open-source image generation just got significantly faster and cheaper to scale.
Snapchat is working on a new generative AI feature called "Dreams" that places you and your friends into fantastical AI-generated scenarios. Discovered by app researcher Steve Moser, the tool asks for clear multi-angle selfies, supports shared portraits via "Dreams with Friends," and hints at paid "Dream Packs." Snap declined to comment.
Runway just launched "Watch" on its website and iOS app, giving creators a dedicated hub to share and stream longer-form AI-generated videos. Amid ongoing Hollywood strikes over generative AI, CEO Cristóbal Valenzuela argued that the debate needs more nuance, positioning the feed as evidence that AI augments filmmakers rather than replacing them.
Shutterstock is tapping NVIDIA Picasso to let creators generate and light 3D scene backgrounds using text or image prompts. The models, trained on fully licensed data, spit out 8K 360-degree environment maps on demand. Why it matters: It offloads the tedious chore of background mapping and scene lighting, letting 3D artists focus their energy on core hero assets.
AI creators just revived digital artist Greg Rutkowski's style using a custom LoRA model on Civitai, bypassing Stability AI's official removal. Rutkowski had campaigned to purge his work from Stable Diffusion 2.0, but open-source creators stepped in to recreate his aesthetic anyway. Why it matters: It shows that in an open-source world, opt-out requests are practically impossible to enforce.
Anonymous creators just launched an infinite Twitch stream pitting AI versions of Donald Trump and Joe Biden against each other in a never-ending debate. The bots mimic voices, facial gestures, and banter while trading unhinged insults and roasting chat users. Over 100,000 viewers tuned in, showing how entertaining—and chaotic—real-time AI impersonation has become.
Midjourney just rolled out version 5.2, introducing a camera-like "zoom out" feature that builds out wider scenes while keeping the original subject intact. The update also brings a revamped aesthetic system and a handy "/shorten" command to trim bloated prompts. The catch: you can't upload custom images to expand yet—it only works on Midjourney's own generations.
Stability AI just launched SDXL 0.9, bringing a significant upgrade to its flagship text-to-image pipeline over the April beta. It uses a 6.6B parameter two-stage ensemble to generate high-res 1024x1024 visuals while still fitting on a consumer GPU with 8GB VRAM. Why it matters: High-fidelity open models are closing the gap with closed competitors, with a full open release slated for mid-July.
Hidden Door just opened limited playtests for its Wizard of Oz role-playing platform, which uses generative AI to generate tailored story adventures for each player. An AI narrator adapts plots, characters, and artwork in real time based on player inputs, with the full game arriving in Winter 2023. Why it matters: It shows how generative tech can transform public domain lore into collaborative, endless narrative sandboxes.
Meta just introduced Voicebox, a generative AI speech model that matches a voice from an audio clip as short as two seconds. It generates natural text-to-speech across six languages, enables cross-lingual style transfer, and lets creators cleanly erase background noise like dog barks and re-synthesize speech.
Google just launched a new virtual try-on feature that generates photorealistic images of clothing on diverse real models. Moving away from awkward image warping, the new diffusion-based model generates garments from scratch to accurately capture wrinkles, drape, and stretch. It is rolling out first for women's tops from brands like H&M and Everlane.
Paul McCartney just revealed that AI helped bring John Lennon's voice back for what he calls the final Beatles record. Using audio separation tech from Peter Jackson's documentary, engineers extracted Lennon's vocal cleanly from an old cassette demo. Why it matters: AI is turning damaged, unusable archives into release-ready masters.
Meta just unveiled MusicGen, a generative AI model that turns text prompts and melody inputs into original songs. Trained on 20,000 hours of music and instrument samples, it can theoretically expand a simple hum into a full track. It opens music creation to anyone, though copyright fights with the record industry are likely looming.
Adobe just rolled out a major revamp of Adobe Express, integrating its Firefly generative AI directly into the browser workflow for text-to-image and text effects. The update brings an all-in-one editor covering drag-and-drop video creation, PDF enhancements, and real-time syncing with Photoshop and Illustrator assets.
