Mickey Finkelson

1/7 New paper 🧵 ScenA generates a multi-speaker audio scene: overlapping speech, laughter, real room noise, from a text description and a few reference voices. When trained in the obvious way, it ignores the text and decides who speaks on its own. We found out why and fixed it.

SCENA by Lightricks. Reference-driven multi-speaker audio scene gen! Prompt "two friends arguing in a rainy cafe" + two ref voice clips -> full 20s audio scene in one pass. Overlapping speech, room echo, background sounds - all baked in. - based on LTX-2.3 - beats ZipVoice-Dialog/MOSS-TTSD - up to 3 ref speakers - perfect text alignment -Identity-aware positional encodings - SOTA performance finmickey.github.io/scena/

🔊 Audio on for this one! Code + checkpoint for ScenA, our multi-speaker audio scene generation model, are out! Reference voices + a prompt → a full conversational scene, prompt alone decides who speaks where. github.com/finmickey/scena arxiv.org/abs/2606.19325

Awesome work for consistent video editing! Comes with a recommended promotion video

🌟🚀 Excited to share our latest work: "Keep The Essentials: Efficient Reference Conditioned Generation via Token Dropping"! TL;DR: Stop wasting compute on redundant tokens! We introduce SparseContext that drops reference tokens for speeding up reference-based image generation⚡

Generating images with AI is fun 🤓 Finding the image? Not so much… The first result is rarely the one. Or the second. Or the third. 🥵 So we built a way to generate diverse possibilities along meaningful semantic axes 🧭 and organize them into a gallery you can effortlessly browse 🧚🏻 Introducing our #ECCV2026 paper: Semantic Browsing: Controllable Diversity for Image Generation ✨ 1/5

📄 arxiv.org/abs/2606.19325 🌐 Samples + code (soon): finmickey.github.io/scena Thanks @OPatashnik & @yoavhacohen and the team at @ltx_io! Happy to take questions! 🙋‍♂️🙋‍♂️🙋‍♂️

7/7 We don't think the Reference Shortcut is specific to audio. It should hit any reference-conditioned flow-matching model trained with a logit-normal schedule, subject-driven image and video generation included. The lever is the timestep distribution.

1/7 New paper 🧵 ScenA generates a multi-speaker audio scene: overlapping speech, laughter, real room noise, from a text description and a few reference voices. When trained in the obvious way, it ignores the text and decides who speaks on its own. We found out why and fixed it.

Excited to share our work accepted to #SIGGRAPH2026 ! Video generation models struggle with something few talk about: their transformations don't evolve smoothly. You get long boring stretches... then a sudden semantic jump where everything "catches up" at once.

1/7 When rewards conflict, what should RL post-training of diffusion models optimize? In visual generation, objectives are often in tension: Prompt adherence can conflict with source preservation. Photorealism can conflict with stylization. In our new paper, ParetoSlider, we introduce a multi-objective RL framework that trains a single diffusion model for continuous control over competing reward objectives 🧵