Here’s a little holiday-appropriate experiment featuring a shot of my dad & me (in Lego form, naturally) at my grandmother’s family farm in County Mayo. Sláinte!
Speaking of reskinning imagery (see last several posts), check out what’s now possible via Google’s Gemini model, below. I’ve been putting it to the test & will share results shortly.
Alright, Google really killed it here.
You can easily swap your garment just by uploading the pieces to Gemini Flash 2.0 and telling it what to do. pic.twitter.com/pNPBkIdRqy
This temporally coherent inpainting is utterly bonkers. It’s just the latest—and perhaps the most promising—in myriad virtual try-on techniques I’ve seen & written about over the years.
I love seeing the Magnific team’s continued rapid march in delivering identity-preserving reskinning
IT’S FINALLY HERE!
Mystic Structure Reference!
Generate any image controlling structural integrity Infinite use cases! Films, 3D, video games, art, interiors, architecture… From cartoon to real, the opposite, or ANYTHING in between!
This example makes me wish my boys were, just for a moment, 10 years younger and still up for this kind of father/son play. 🙂
Storyboarding? No clue! But with some toy blocks, my daughter’s wild imagination, and a little help from Magnific Structure Reference, we built a castle attacked by dragons. Her idea coming to life powered up with AI magic. Just a normal Saturday Morning. Behold, my daughter’s… pic.twitter.com/52tDZokmIT
“Rather than removing them from the process, it actually allowed [the artists] to do a lot more—so a small team can dream a lot bigger.”
Paul Trillo’s been killing it for years (see innumerable previous posts), and now he’s given a peek into how his team has been pushing 2D & 3D forward with the help of custom-trained generative AI:”
Traditional 2d animation meets the bleeding edge of experimental techniques. This is a behind the scenes look at how we at Asteria brought the old and the new together in this throwback animation “A Love Letter to Los Angeles” and collaboration with music artist Cuco and visual… pic.twitter.com/3eWSdgckXn
A passing YouTube vid made me wonder about the relative strengths of World War II-era bombers, and ChatGPT quickly obliged by making me a great little summary, including a useful table. I figured, however, that it would totally fail at making me a useful infographic from the data—and that it did!
Just for the lulz, I then ran the prompt (“An infographic comparing the Avro Lancaster, Boeing B-17, and Consolidated B-24 Liberator bombers”) through a variety of apps (Ideogram, Flux, Midjourney, and even ol’ Firefly), creating a rogue’s gallery of gibberish & Franken-planes. Check ’em out.
Currently amusing myself with how charmingly bad every AI image generator is at making infographics—each uniquely bizarre! pic.twitter.com/U3cs8ySoVa
By combining @pika_labs Pikaframes and @freepik, I now have the magical ability to jump through space and time and in this example, music becomes a transformative element teleporting this woman to a new location. This is how it’s done. 1/6
Generate image (in this example, using Google Imagen).
Apply background segmentation.
Synthesize a new background, and run what I think is a fine-tuned version of IC-Light (using Stable Diffusion) to relight the entire image, harmonizing foreground/background. Note that identity preservation (face shape, hair color, dress pattern, etc.) is very good but not perfect; see changes in the woman’s hair color, expression, and dress pattern.
Put the original & modified images into Pika, then describe the desired transformation (smooth transition, flowers growing, clouds moving, etc.).
Another day, another ~infinite canvas for ideation & synthesis. This time, somewhat to my surprise, the surface comes from VSCO—a company whose users I’d have expected to be precious & doctrinaire in their opposition to any kind of AI-powered image generation. But who knows, “you can just do things.” ¯\_(ツ)_/¯
The capturing work was led by Harry Nelder and Amity Studio. Nelder used his 16-camera rig to capture the recent winners. The reconstruction software was a combination of a cloud-based platform created by Nelder, which is expected to be released later this year, along with Postshot. Nelder further utilized the Radiance Field method known as Gaussian Splatting for the reconstruction. A compilation video of all the captures, recently posted by BAFTA, was edited by Amity Studio
Check it out (probably easier to grok by watching vs. reading a description):
From the static camera feed, EditIQ initially generates multiple virtual feeds, emulating a team of cameramen. These virtual camera shots termed rushes are subsequently assembled using an automated editing algorithm, whose objective is to present the viewer with the most vivid scene content.
Tired: Random “slot machine”-style video generation Inspired: Placing & moving simple guidance objects to control results: Check out VideoNoiseWarp:
Every now and then something comes along that feels like it could change everything… NoiseWarp + CogVideoX lets you animate live action scenes with rough mockups!
Here’s the tutorial! This video combines AI-generated elements (balloon, kite, surfboard, and backgrounds) with my own real-world practical effects and stop motion.
The YouTube mobile app can now tap into Google’s Veo model to generate video, as shown below. Hmm—this feels pretty niche at the moment, but it may suggest the shape of things to come (ubiquitous media synthesis, anywhere & anytime it’s wanted).
For the longest time, Firefly users’ #1 request was to use images to guide composition of new images. Now that Firefly Video has arrived, you can use a reference image to guide the creation of video. Here’s a slick little demo from Paul Trani:
These changes, reported by Forbes, sound like reasonable steps in the right direction:
Starting now, Google will be adding invisible watermarks to images that have been edited on a Pixel using Magic Editor’s Reimagine feature that lets users change any element in an image by issuing text prompts.
The new information will show up in the AI Info section that appears when swiping up on an image in Google Photos.
The feature should make it easier for users to distinguish real photos from AI-powered manipulations, which will be especially useful as Reimagined photos continue to become more realistic.
But how do we go from ironic laughs to actual usefulness? Krea is taking a swing by integrating (I think) the Flux imaging model with the DeepSeek LLM:
Krea Chat is here.
a brand new way of creating images and videos with AI.
It doesn’t yet offer the kind of localized refinements people want (e.g. “show me a dog on the beach,” then “put a hat on the dog” and don’t change anything outside the hat area). Even so, it’s great to be able to create an image, add a photo reference to refine it, and then create a video. Here’s my cute, if not exactly accurate, first attempt. 🙂
Wow—check out this genuinely amazing demo from my old friend (and former Illustrator PM) Mordy:
In this video, I show how you can use Gemini in the free Google AI Studio as your own personal tutor to help you get your work done. After you watch me using it to learn how to take a sketch I made on paper to recreating a logo in Illustrator, I promise you’ll be running to do the same.
We propose MatAnyone, a robust framework tailored for target-assigned video matting. Specifically, building on a memory-based paradigm, we introduce a consistent memory propagation module via region-adaptive memory fusion, which adaptively integrates memory from the previous frame. This ensures semantic stability in core regions while preserving fine-grained details along object boundaries.
Users can enter search terms like “a person skating with a lens flare” to find corresponding clips within their media library. Adobe says the media intelligence AI can automatically recognize “objects, locations, camera angles, and more,” alongside spoken words — providing there’s a transcript attached to the video. The feature doesn’t detect audio or identify specific people, but it can scrub through any metadata attached to video files, which allows it to fetch clips based on shoot dates, locations, and camera types. The media analysis runs on-device, so doesn’t require an internet connection, and Adobe reiterates that users’ video content isn’t used to train any AI models.
Goodbye, endless scrolling. Hello, AI-powered search panel. With the all-new Media Intelligence in #PremierePro (beta), the content of your clips is automatically recognized, including objects, locations, camera angles & more. Just input your search to find exactly what you need. pic.twitter.com/cOYXDKKaFI
Putting the proverbial chocolate in the peanut butter, those fast-moving kids at Krea have combined custom model training with 3D-guided image generation. Generation is amazingly fast, and the results are some combo of delightful & grotesque (aka “…The JNack Story”). Check it out:
God help you, though, if you import your photo & convert it to 3D for use with the realtime mode. (Who knew I was Cletus the Slack-Jawed Yokel?) pic.twitter.com/nuesUOZ1Db
Here’s another interesting snapshot of progress in our collective speedrun towards generative storytelling. It’s easy to pick on the shortcomings, but can you imagine what you’d say upon seeing this in, say, the olden times of 2023?
The creator writes,
Introducing The Heist – Directed by Jason Zada. Every shot of this film was done via text-to video with Google Veo 2. It took thousands of generations to get the final film, but I am absolutely blown away by the quality, the consistency, and adherence to the original prompt. When I described “gritty NYC in the 80s” it delivered in spades – CONSISTENTLY. While this is still not perfect, it is, hands down, the best video generation model out there, by a long shot. Additionally, it’s important to add that no VFX, no clean up, no color correction has been added. Everything is straight out of Veo 2. Google DeepMind
Here’s a nice write-up covering this paper. It’ll be interesting to dig into the details of how it compares to previous work (see category). [Update: The work comes in part from Adobe Research—I knew those names looked familiar :-)—so here’s hoping we see it in Photoshop & other tools soon.]
this is wild..
this new AI relighting tool can detect the light source in the 3D environment of your image and relight your character, the shadows look so realistic..
Part 9,201 of me never getting over the fact we were working on stuff like this 2 years ago at Adobe (modulo the realtime aspect, which is rad) & couldn’t manage to ship it. It’ll be interesting to see whether the Krea guys (and/or others) pair this kind of interactive-quality rendering with a really high-quality pass, as NVIDIA demonstrated last week using Flux.
3D arrived to Krea.
this new feature lets you turn images into 3D objects and use them in our Real-time tool.
Powered by advanced AI, TRELLIS enables users to create high-quality, customizable 3D objects effortlessly using simple text or image prompts. This innovation promises to improve 3D design workflows, making it accessible to professionals and beginner alike. Here are some examples:
Alpha channels are crucial for visual effects (VFX), allowing transparent elements like smoke and reflections to blend seamlessly into scenes. We introduce TransPixar, a method to extend pretrained video models for RGBA generation while retaining the original RGB capabilities. […] Our approach effectively generates diverse and consistent RGBA videos, advancing the possibilities for VFX and interactive content creation.
The world moves on, and now NVIDIA has teamed up with Black Forest Labs to enable 3D-conditioned image generation. Check out this demo (starting around 1:31:48):
For users interested in integrating the FLUX NIM microservice into their workflows, we have collaborated with NVIDIA to launch the NVIDIA AI Blueprint for 3D-guided generative AI. This packaged workflow allows users to guide image generation by laying out a scene in 3D applications like Blender, and using that composition with the FLUX NIM microservice to generate images that adhere to the scene. This integration simplifies image generation control and showcases what’s possible with FLUX models.
The Former Bird App™ is of course awash in mediocre AI-generated video creations, so it’s refreshing to see what a gifted filmmaker (in this case Ruairi Robinson) can do with emerging tools (in this case Google Veo)—even if that’s some slithering horror I’d frankly rather not behold!
I’ve long wanted—and advocated for building—this kind of flexible, spatial way to compose & blend among ideas. Here’s to new ideas for using new tools.
Supporting Non-Linear Exploration
Creative exploration rarely follows a straight line. The graph structure naturally affords exploration by allowing users to diverge at various points, creating new forks of possible alternatives. As more exploration occurs, the graph grows… pic.twitter.com/Yq18Caj94T
It’s a touch odd to me that Meta is investing here while also shutting down the Meta Spark AR lens platform, but I guess interest in lenses has broadly faded, and AI interpretation of images may prove to be more accessible & scalable. (I wonder what’ll be its Dancing Hot Dog moment.)
I love seeing exactly how Chad Nelson was able to construct a Little Big Planet-inspired game world through some creative prompting & tweening in Open AI’s new Sora video creation model. Check out his exploratory process:
Director Matan Cohen-Grumi shows off the radical acceleration in VFX-heavy storytelling that’s possible through emerging tools—including Pika’s new Scene Ingredients:
For 10 years, I directed TV commercials, where storytelling was intuitive—casting characters, choosing locations, and directing scenes effortlessly. When I shifted to AI over a year ago, the process felt clunky—hacking together solutions, spending hours generating images, and… pic.twitter.com/pJUamLFgWI
Instead of generating images with long, detailed text prompts, Whisk lets you prompt with images. Simply drag in images, and start creating.
Whisk lets you input images for the subject, one for the scene and another image for the style. Then, you can remix them to create something uniquely your own, from a digital plushie to an enamel pin or sticker.
The blog post gives a bit more of a peek behind the scenes & sets some expectations:
Since Whisk extracts only a few key characteristics from your image, it might generate images that differ from your expectations. For example, the generated subject might have a different height, weight, hairstyle or skin tone. We understand these features may be crucial for your project and Whisk may miss the mark, so we let you view and edit the underlying prompts at any time.
In our early testing with artists and creatives, people have been describing Whisk as a new type of creative tool — not a traditional image editor. We built it for rapid visual exploration, not pixel-perfect edits. It’s about exploring ideas in new and creative ways, allowing you to work through dozens of options and download the ones you love.
And yes, uploading a 19th-century dog illustration to generate a plushie dancing an Irish jig is definitely the most JNack way to squander precious work time do vital market research. 🙂
I’m a near-daily user of Ideogram to create all manner of images—mainly goofy dad jokes to (ostensibly) entertain my family. Now they’re enabling batch creation to facilitate creation of lots of variations (e.g. versions of a logo):
Check out this wild video-to-video demo from Nathan Shipley:
Sora Remix test: Scissors to crane
Prompt was “Close up of a curious crane bird looking around a beautiful nature scene by a pond. The birds head pops into the shot and then out.” pic.twitter.com/CvAkdkmFBQ
Just a taste of the torrent the blows past daily on The Former Bird App:
Rodin 3D: “Rodin 3D AI can create stunning, high-quality 3D models from just text or image inputs.”
Trellis 3D: “Iterative prompting/mesh editing. You can now prompt ‘remove X, add Y, Move Z, etc.’… Allows decoding to different output formats: Radiance Fields, 3D Gaussians, and meshes.”
Blender GPT: “Generating 3D assets has never been easier. Here’s me putting together an entire 3D scene in just over a minute.”
This might be the world’s lowest-key demo of what promises to be truly game-changing technology!
I’ve tried a number of other attempts at unlocking this capability (e.g. Meta.ai (see previous), Playground.com, and what Adobe sneak-peeked at the Firefly launch in early 2023), but so far I’ve found them all more unpredictable & frustrating than useful. Could Gemini now have turned the corner? Only hands-on testing (not yet broadly available) will tell!
Diffusion models are ushering in what feels like a golden(-hour) age in relighting (see previous). Among the latest offerings is LumiNet:
[6/7] Here are a few more random relighting!
How accurate are these results? That’s very hard to answer at the moment But our tests on the MIT dataset, our user study, plus qualitative results all point to us being on the right track.
