David Dinkevich

CS MSc student at @HebrewU. Generative AI Researcher.

[David Dinkevich @DavidDinkevich]

[1/6] 🎬 New paper: Story2Board
We guide diffusion models to generate consistent, expressive storyboards--no training needed.
By mixing attention-aligned tokens across panels, we reinforce character identity without hurting layout diversity.
🌐 https://t.co/aRG81nu5qK

[Mickey Finkelson @mikikiFin]

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.

[Noam Issachar @IssacharNoam]

1/6 Diffusion models are scaling up, but deploying a massive, monolithic network uniformly across the entire generative timeline is inherently inefficient.
Introducing Complexity-Balanced Splitting (CBS): a principled framework that allocates capacity exactly where needed!👇🧵

[David Dinkevich @DavidDinkevich]

[6/6] Huge thanks to my collaborators for making this happen. Excited to see how the community builds on training-free consistency. More examples + code: 👉 https://t.co/2Mawweuifj thanks to my collaborators @MatanLvy, @OmriAvr, @dvir_samuel, and @DaniLischinski

[David Dinkevich @DavidDinkevich]

[1/6] 🎬 New paper: Story2Board
We guide diffusion models to generate consistent, expressive storyboards--no training needed.
By mixing attention-aligned tokens across panels, we reinforce character identity without hurting layout diversity.
🌐 https://t.co/aRG81nu5qK

[David Dinkevich @DavidDinkevich]

[5/6] 📚 To evaluate properly, we introduce the Rich Storyboard Benchmark
100 open-ended fantasy stories. 700 total panels.
Each one with detailed character actions and evolving backgrounds, designed to push generative models beyond static templates.

[David Dinkevich @DavidDinkevich]

[4/6] 🎥 How do we measure expressive storytelling?
We propose Scene Diversity, a new metric that captures variation in pose, scale, and layout across panels.
It rewards models that treat characters like real cinematic subjects, not static cutouts.

[David Dinkevich @DavidDinkevich]

[3/6] 🧩 Step 2: Reciprocal Attention Value Mixing
Some tokens across panels naturally attend to each other, like matching parts of the character.
We softly blend their visual features during inference, reinforcing consistency while keeping each scene dynamic.

[David Dinkevich @DavidDinkevich]

[2/6] 🧩 Step 1: Latent Panel Anchoring
We generate two-panel images: top = reference, bottom = scene.
During generation, we keep the top half fixed across panels, giving the model a stable anchor to base each scene on.

[David Dinkevich @DavidDinkevich]

@MahDryBread Can you play some more HOI4?

[David Dinkevich @DavidDinkevich]

@MKBHD @ijustine Will you ever start covering laptops and not just smartphones?