harley turan

1.4K posts

harley turan banner
harley turan

harley turan

@hturan

exploring dimensionality 🎛️ ex principal eng @ cloudflare

London, UK Katılım Temmuz 2007
627 Takip Edilen13.1K Takipçiler
Sabitlenmiş Tweet
harley turan
harley turan@hturan·
gradually filling in the back catalogue on the personal site — hturan.com
English
3
4
55
24.6K
harley turan
harley turan@hturan·
generated multi-modal embeddings of my bookmarks using @GoogleDeepMind’s new Gemini Embedding 2, allowing low-dim spatialisation using UMAP and high-dim similarity querying via different media types.
English
6
37
326
24.2K
harley turan
harley turan@hturan·
@patcon_ @poetengineer__ @DefenderOfBasic totally! there’s space here for a component library with inputs outputting a normalised 0 → 1 mapping which can then be adapted as needed. general feeling is that interactions like this benefit exploratory gestures rather than precision input, but i haven’t gone too deep.
English
0
0
1
23
jason liu
jason liu@jxnlco·
Steal his look @dylan522p Dylan looks so stylist here man.
jason liu tweet mediajason liu tweet media
English
10
1
173
30.5K
harley turan retweetledi
grant
grant@GrantCuster·
A nano-banana canvas experiment. Generating based on a source image and the overlapping prompts, caching the results. Like adding semantic layers to an image.
English
3
6
68
3.4K
Lukas Gjetting
Lukas Gjetting@lukasgjetting·
@hturan are the interpolated images really embedded in the full latent space, though? if so, shouldn't the flower<>sunset images at 0:28 be turning more green as they approach the park pic? otherwise it feels more like just a series of 1d interpolations
English
2
0
1
294
harley turan
harley turan@hturan·
exploring paths through latent space by interpolating between known embeddings
English
51
121
1.6K
96K
harley turan
harley turan@hturan·
interpolating image embeddings across arbitrary numbers of vertices. using polygons as lenses through n-dimensional latent space.
English
17
23
326
22.8K
harley turan
harley turan@hturan·
@bee_human_ yeah, i’d love this! video too. time-based media is hard from an interaction perspective (would you place images at certain points on a timeline to steer the mood of a song?).
harley turan@hturan

@DanielleFong i’ve got it working across text and image modalities! end goal is really just high dimensional media as a way to pierce the curse of dimensionality. give me some solid ground to stand on in the latent space.

English
1
0
5
1.7K
Bee 🐝
Bee 🐝@bee_human_·
@hturan so sick! you think we could do the same with music?
English
1
0
4
1.3K
harley turan
harley turan@hturan·
@DanielleFong i’ve got it working across text and image modalities! end goal is really just high dimensional media as a way to pierce the curse of dimensionality. give me some solid ground to stand on in the latent space.
English
1
0
1
1.8K
harley turan
harley turan@hturan·
@MatthewWSiu yeah! i’ve got it working multiple times-dimensionally (interpolating across n points) and multimodally too. i haven’t tried it, but part of me thinks mixing too many sources would be like mixing too many pigments and you’d end up with nothing. curious to see!
English
1
0
9
1.6K
Matthew Siu
Matthew Siu@MatthewWSiu·
@hturan this would be a fun way to explore your photo library as well, it'd prob be less continuous but still curious how it would feel
English
2
0
14
1.9K
harley turan
harley turan@hturan·
@prathyvsh i've been toying with multimodal interpolation as a method of constraining a path within image diffusion models, and it's feeling like it's striking the balance well. blending between (semi) known points in visual possibility space.
harley turan tweet media
English
1
0
3
115
Prathyush
Prathyush@prathyvsh·
Thinking of making a catalog of possibility space explorations soon. Are there any such examples of seeking through the adjacent possible in a latent space like this that has come across your attention?
tetsuo@tetsuoai

English
6
3
28
2.7K