Loop packs · Pack 09
Minds: AI, Made Visible
What happens inside a neural network, drawn from the network itself: latent space, attention, diffusion, a language model thinking and the moment a model understands.
Founders: 30% off until October 19. Applied at checkout, no code needed.
Previews
Watch them loop.
Watermark-free previews at 540p. The pack is the same frames at 4K, graded for HDR.
- TrefoilA trefoil-knot surface of living points, one full turn of the cameraLoops natively
- Seashell MöbiusA seashell becomes a Möbius strip and turns backLoops natively
- Grand TourA grand tour of an eight-dimensional space, out and backLoops natively
- Grand Tour IIThe grand tour, through a second windowLoops natively
- Diffusion ButterflyNoise to butterfly to noise: one denoising cycle per loopLoops natively
- Diffusion WhaleA whale condenses out of noise and dissolves backLoops natively
- Attention LatticeAn attention lattice pulsing on every kickLoops natively
- Helix SphereA helix folds into a sphere and back, on the beatLoops natively
- KnotA knot of attention, kick waves running through itLoops natively
- Galaxy ThoughtA galaxy forming inside a thought, with shockwavesLoops natively
- Saturn ThoughtSaturn forming, one spin per loopLoops natively
- Evolution FishA swimming fish evolves into a lizard and backLoops natively
- Evolution HumanA walking human, six strides a loopLoops natively
- The ClickA tiny transformer groks addition: the tangle snaps onto four ringsLoops natively
- River of Thought4,624 tokens flow through all 29 layers of a real language modelLoops natively
- CondensationA diffusion model dreams a jellyfish into a nautilus, a coral and a mantaSeamless crossfade
- The Click IIThe same snap for multiplication, in logarithm orderLoops natively
What’s included
The science, and the files.
Driven by computations we ran ourselves: a transformer trained until it grokked modular arithmetic, real hidden states of Qwen3-1.7B reading real text, diffusion models we trained and sampled, and Everdust's latent-space and attention scenes rendered natively at 4K. The 10-second scenes pulse on a synthetic 120 BPM envelope (20 beats a loop), so they sit on a beat grid.
Every loop is our own computation and render. River of Thought shows hidden states of Qwen3-1.7B (Apache 2.0), used as data; no third-party footage.
Made in Everdust. Step inside these worlds: try the free beta on iPhone, iPad, Mac, Apple TV and Vision Pro on TestFlight.
- 4K UHD
- 3840×2160, 60 fps, H.264 High, SDR BT.709. Plays anywhere. · ~132 MB a loop
- 4K HDR10
- 3840×2160, 60 fps, HEVC Main 10, PQ / BT.2020, 1,000-nit grade · ~143 MB a loop
- Vertical
- 2160×3840 (9:16) for vertical video, reframed, not cropped blind · ~131 MB a loop
- Vertical HDR10
- 2160×3840 HEVC Main 10, PQ / BT.2020 · ~143 MB a loop
- ProRes 4444 / 422 HQ
- 4K60 per-loop downloads; 4444 with alpha where a loop has a solid surface, 422 HQ otherwise · ~2.1 GB a loop
- HAP
- 1080p60 for VJ and media-server software · ~495 MB a loop
- DXV3
- 1080p60 for VJ software that plays DXV · ~397 MB a loop
- Silent
- No audio, so there is no music license to clear
Everdust_Minds/ ├── README.txt ├── LICENSE.txt v1, personalized on your order page ├── 4K_H264/ 17 files ├── 4K_HDR10_HEVC/ 17 files ├── Vertical_9x16_H264/ 17 files ├── Vertical_9x16_HDR10_HEVC/ 17 files ├── ProRes/ per-loop downloads ├── HAP_1080p60/ VJ download └── DXV3_1080p60/ VJ download
Questions
- Do they really loop?
- Yes. Each loop is either periodic in the simulation itself, a camera orbit around a frozen moment, or a 2-second crossfade we check frame by frame at the seam. A loop that shows a seam or a mid-fade ghost doesn’t ship. The exception is marked “Full-event cycle”: an event that only runs one way, like two black holes merging, plays from start to finish and then dissolves back to its first frame over the last 1.5 seconds. The seam is checked the same way, but during that dissolve you see the end and the start at once.
- How are these made?
- Every frame is simulated, not generated video: stars under gravity, fluids under pressure, solved by our own code. Nothing is stock footage.
- Can I use them in a paid music video or a client ad?
- Yes, under Standard. TV, OTT, cinema, unlimited venues and products you sell need Extended.
- Which formats do I get?
- All of them, at every tier: 4K H.264, 4K HDR10, vertical 9:16 in both, ProRes per loop, and HAP and DXV3 for VJ software.
- How do downloads work?
- After checkout you get an email with a private link to your order page. It works for 7 days, and you can re-send a fresh link from your receipt page any time. Each file can be downloaded up to 10 times.
- Can I train an AI model on them?
- No. Training, fine-tuning or evaluating AI models on the loops, putting them in a dataset, or feeding them to generative AI services isn’t allowed on any tier. Ordinary editing features like upscaling or stabilization in your own software are fine.
- Can I play them unedited in a VJ set or on a venue screen?
- Yes. Showing the loops as they are is fine within your tier; what you can’t do is hand anyone the files.
- Refunds?
- Within 14 days if you haven’t downloaded anything, and always if a file is broken. Write to hello@everdust.ai.
- Do I need to credit you?
- Not us: “Visuals: Everdust (everdust.ai)” is always welcome, though. Two of the free sampler loops, Kerr Disk and Kerr Face On, show real star data from the HYG Database (CC BY-SA 4.0); no loop in a paid pack does. If you share those two, or a work that adapts them, that license asks for its credit, “Star sky: HYG Database v4.0, David Nash / astronexus, CC BY-SA 4.0, modified by Everdust” (license section 12).