Google debuts Whisk, an image generator that takes other images as prompts to suggest the subject, scene, and style, and uses the new, “latest” Imagen 3 version

Google has announced a new AI tool called Whisk that lets you generate images using other images as prompts instead of requiring a long text prompt.

The Verge 2024-12-17 Jay Peters

Context & Ripple Effects

Whisk extends Google’s image-generation arc from Imagen 2’s quality and feature upgrade to ImageFX, where Google introduced suggested prompt terms to make text-led creation easier. Its shift is from helping users write prompts to letting reference images express creative intent.

The approach was subsequently validated by Whisk’s rollout to more than 100 countries, preserving the same subject, scene, and style remixing model. That makes this launch an interface decision around Imagen 3, not merely a model-version update.

First-order effects

Whisk gives users a visual input path for image generation: separate images can indicate the desired subject, setting, and aesthetic rather than requiring a detailed written prompt.
Google gets a new consumer-facing surface for Imagen 3 that emphasizes remixing and composition of references over standalone text prompting.

Second-order effects

Image-generation products are pushed to compete on controllability and ease of direction, not only output quality; reference-image workflows reduce the prompt-writing skill needed to get closer to an intended result.
Creators can treat existing visual material as a lightweight creative brief, increasing the value of tools that preserve distinct elements such as subject, scene, and style during generation.

Third-order effects

If image-reference interfaces become standard, generative-image competition will increasingly center on multimodal editing workflows that translate visual intent into outputs rather than on text-to-image generation alone.
The same simplification can make reuse of recognizable visual styles and subjects more commonplace, keeping likeness and provenance questions central as these tools broaden.

The trend: Generative-image products are moving from prompt-centric creation toward multimodal, workflow-native controls that let users direct models with visual references.

Discussion

@jaypeters.net Jay Peters on bluesky
Google has a new AI tool that lets you generate images by using images as prompts. Everything I've made with it has been pretty strange, like this bear that has six claws on each hand — www.theverge.com/2024/12/16/ 2...
@emollick Ethan Mollick on x
Pretty good prompt compliance from Imagen 3. [image]
@jasonbaldridge Jason Baldridge on x
Whisk is a really fun and compelling new way to prompt our Imagen 3 model. You upload images that have subject, scene or style of interest, and then it composes these together to generate a new image based on these elements. Less wordsmithing, and fun results!
@joshtwoodward Josh Woodward on x
It's time to Whisk! What makes it unique? No more long, detailed text prompts. Whisk lets you prompt with images and easily blend them together. It's so fast and fun. Just drag in your images and start creating. We created Whisk based on conversations with filmmakers working on
@samhouston Sam Houston on x
Google's new AI Whisk tool is pretty neat [image]
@samhouston Sam Houston on x
it's nice to see Google shipping products across multiple Product Areas that all support the overarching AI mission. nice to see the company all moving in the same direction. seems like they're starting to unlock some innovation (ex: NotebookLLM) and some cool use-cases
@googledeepmind @googledeepmind on x
Today, we're announcing Veo 2: our state-of-the-art video generation model which produces realistic, high-quality clips from text or image prompts. 🎥 We're also releasing an improved version of our text-to-image model, Imagen 3 - available to use in ImageFX through [video]
@simonw Simon Willison on x
Veo 2 did pretty well at “A pelican riding a bicycle along a coastal path overlooking a harbor” - two of these videos have the pelican actually cycling! [video]
@jason @jason on x
Google/Gemini and Grok have reached parity, and in some cases are much better, than ChatGPT I use all three, every day, at least 50-100x ... ChatGPT went from 90% of my queries last month to < 50% It's now a data and UI game, and I think Reddit, X/twitter and google's dataset wil…
@andrewcurran_ Andrew Curran on x
Veo 2 looks great, they're not going to let Sora hog the spotlight. Interestingly, according to Google's test results, their strongest competition in generative video is not META, or even Sora, but Kling! [image]
@alexanderchen Alexander Chen on x
We launched a new, interactive video analyzer in @googleaistudio that you can use to test Gemini 2.0 for video understanding! 🎉 It parses the model's timecode results to make it way easier to explore results. Try it here: https://aistudio.google.com/ ... And here's a demo video o…
@officiallogank Logan Kilpatrick on x
Introducing the Gemini for Academic Research program, created to help researchers using Gemini models accelerate their research agenda. 🧪🥼 We are supporting researches exploring evaluations, benchmarks, embodiment, science, and more! Apply today: https://ai.google.dev/...
@trondw Trond Wuellner on x
I've had early access to Veo 2 and am BLOWN AWAY by how far video generation has come. I can't wait to see what you all do with it!
@thexeophon @thexeophon on x
Google's 12 days of shipping are more exciting and they skip days
@karpathy Andrej Karpathy on x
AI video generation today. When I was back in school, the story of the field of computer graphics (and physically based rendering etc.) was that we will carefully study and model all the object/scene geometry, physics, rendering etc., and after 1000 PhDs and 50 SIGGRAPHs get res…
@tkipf Thomas Kipf on x
Capybara gymnastics ✅ Generated with #Veo2 [video]
@bilawalsidhu Bilawal Sidhu on x
BREAKING: Google just dropped Veo 2 and Imagen 3 — their next gen video and image generation models. Turns out Google's been closing the gap quietly — not just on LLMs, but on visual creation too. Here's everything you need to know w/o the hype 🧵 [video]
@doomie Dumitru Erhan on x
Actually one of my favorite things to do with Veo 2 is spell things. It works like magic!
@rpnickson Roberto Nickson on x
Insane week for Google: - Willow - Gemini 2.0 - Veo 2 [image]
@matthewberman @matthewberman on x
Veo 2 is a cinematographer's dream It understands lens choices (18mm wide angle anyone?) and camera effects like depth of field. Plus, fewer AI “hallucinations” like those weird extra fingers we're used to seeing [video]
@rubenevillegas Ruben Villegas on x
A broccoli wearing a leather jacket and carrot wearing a tank top having a steak dinner #veo2 [video]
@_tim_brooks Tim Brooks on x
Great example of accurate physics in Veo 2!
@mayorkingai @mayorkingai on x
This is wild! Google DeepMind just announced Veo 2, their state-of-the-art AI video model Veo 2 outperforms other leading video generation models! Here are 10 insane examples👇 [video]
@gbarthmaron Gabriel Barth-Maron on x
Extremely proud to have been a part of this. Veo 2 is a significant upgrade and has achieved SoTA results in head-to-head comparisons of outputs by human raters over top video generation models. Looking forward to seeing what everyone creates with it. https://deepmind.google/...
@ai_for_success AshutoshShrivastava on x
This is probably the best AI video of all time. At first, it looks simple , but then you realize how insanely good Veo-2 is at simulating the physical world. Damn Google how?? 🫡
@ada_rob Adam Roberts on x
Really proud of my teammates on the Video & Image portions of @GoogleDeepMind GenMedia for pulling this off. It's wonderful to be part of such an impressive (and fun) team! #veo2
@hhm Hernan Moraldo on x
Veo v2 generates a meeting of animals #Veo2 Prompt: A meeting of a lion, a bear and a giraffe, all of them wearing suits. Photorealistic, cinematic. [video]
@conorgrennan Conor Grennan on x
BREAKING: @Google just dropped Veo 2, their video model, and WHOA. I had a sneak peek at their Sora competitor and was blown away, both the quality and interface. I think 2025 is The Year of Google. Here's what you need to know about Veo 2, which is being rolled out to Google [vi…
@diesol Dave Clark on x
Google Veo 2. Prompt: A bartender making an old-fashioned cocktail. Text2Video. Two variations. #VideoFX #Veo2 @GoogleDeepMind [video]
@ai_for_success AshutoshShrivastava on x
Google is so back 🔥🔥 They just announced Veo 2, an AI video generation model, and it's incredible. They've also updated the Imagen 3 image-generation model. 10 examples below 👇 [video]
@poolio Ben Poole on x
the sweater frogs can moooove #veo2 [video]
@kimmonismus @kimmonismus on x
Google strikes again! VEO 2 AI Video released While OpenAI releases boring, small-scale seach updates, Google releases DeepMind Veo 2, their new SOTA AI video model. Goodbye Sora? “we're announcing Veo 2: our state-of-the-art video generation model which produces realistic, [vide…
@dylanneve10 Dylan Neve on x
Just got access to Veo 2 on AI Test Kitchen! Results look amazing #VideoFX #Veo2 [video]
@doomie Dumitru Erhan on x
Veo 2 is shipping! :) [video]
@hhm Hernan Moraldo on x
Soccer from the future, according to Veo 2 #veo2 [video]
@matthewberman @matthewberman on x
Safety Every Veo 2 video comes with invisible SynthID watermarks. You'll find it in VideoFX (Google Labs), with YouTube Shorts integration coming in 2025 [video]
@gbarthmaron Gabriel Barth-Maron on x
“An astronaut exploring an underwater alien shipwreck.” #veo2 [video]
@drjwrae Jack Rae on x
Really beautiful videos from Veo 2 and images from Imagen 3. The team are super talented and have delivered something pretty special! Try Imagen 3 out in labs fx: https://labs.google/fx
@web3willbefree Danny Ki on x
Veo 2 is the new SOTA in video generation! @Google has strategically timed its release after OpenAI's Sora. Their tests claim Veo 2 outpaces competitors, but take those with a grain of salt. #AI [video]
@doomie Dumitru Erhan on x
I love the synchronization of the flowers with the walk #veo2 [video]
@swyx @swyx on x
Veo 2 vs Sora comparison. apparently 58% win rate vs Sora turbo prompt: “the camera moves in a slow dolly shot, revealing the opulence of a Renaissance palace chamber adorned with gold-inlaid furniture, velvet drapes, and chandeliers casting soft, flickering light.
@tkipf Thomas Kipf on x
Check out Veo 2, our most capable video generation model. “Veo 2 outperforms other leading video generation models, based on human evaluations of its performance.” https://deepmind.google/... It was exhilarating to see the development of Veo 2 unfold, and to play a small part in
@rowancheung Rowan Cheung on x
Google just released Veo 2, a new state-of-the-art AI video model. In testing, Veo beat OpenAI Sora in BOTH quality and prompt adherence. The video compilation below is 100% created by AI (more details in thread): [video]
@doomie Dumitru Erhan on x
Veo 2 imagines neurips parties in the future ;) [video]
@rubenevillegas Ruben Villegas on x
A monkey and a potato riding a bike under water surrounded by colorful fish and sharks. #veo2 [video]
@angrypenguinpng @angrypenguinpng on x
The physics understanding with Veo 2 is insane. Prompt : “A man jogging on a treadmill” [video]
@_tim_brooks Tim Brooks on x
🔥🔥🔥 Veo 2 is DeepMind's newest state-of-the-art video gen model. Congrats to the Veo team on all the hard work that went into making it!! I'm especially impressed by the improved motion and physical accuracy. Learn more at https://deepmind.google/... https://youtube.com/...
@sundarpichai Sundar Pichai on x
Introducing Veo 2, our new, state-of-the-art video model (with better understanding of real-world physics & movement, up to 4K resolution). You can join the waitlist on VideoFX. Our new and improved Imagen 3 model also achieves SOTA results, and is coming today to 100+ countries …
@mattturck Matt Turck on x
Google buying DeepMind for $400M-$650M in 2014 may very well be the best acquisition of all times.
@krishnanrohit Rohit on x
True. And rather quaint how they said the acquisition was worth it because it helped reduce the power demands in data centers!

Chronicles