Studio Robert Seidel

🙂🪽 can’t believe this online exhibition i curated with Tina and Carlotta is already 9 years old still worth a visit a snapshot of the era, featuring key figures, some still active today 💻💻 (better viewed on desktop) 🦄🦄 mermaidsandunicorns.net ✨✨@kimasendorf, @laturboavedon, @arambartholl, @gregorychatons, @mrkdrf, @Katie_Torn, @brb_irl, @beafremderman, @carlagannis, @YesClaudiaHart, @asugarhigh, @GLillemon,Douglas Repetto & @lovidlovid, @_menkman, @RossnerManuel, @Nicolas_Sassoon, @robertseidelcom✨✨ (👇🏼 more)

@dnlklr Had one request for an interview last year… they spent more time on my fake questions than I’m expecting these days in general from art people... The mixture of flattering and time pressure for a “deadline of the next issue” is very effective… so yes, be safe!

came sorta close to getting phished by a relatively elaborate fake podcast interview scheme. There were enough red flags that I was prepared to bail if they asked me to download anything, which ofc they eventually did but...be safe out there

@peterkirn It has an easy implementation for AE: aescripts.com/tracery/

Okay, folks -- what is this effect and why am I suddenly seeing it everywhere? Is it like a comfyUI thing? Wait. I realize I just *literally* became an old man yelling at clouds. (it's fine, just curious)

To Fail the World with an Open Heart Night Lights Denver Daniels & Fisher Tower May 1 – 31, 2026 Lillian Schwartz, @nonperformant, Patricia Detmering, Zeitguised, Natalia Stuyk, Grischa Lichtenberger, @lm_netwebs Curated by @robertseidelcom Trailer Score by @Sallwitz

@cfryant @PurzBeats @ComfyUI The potential possibilities of 16GB VRAM would be good to know ✨

@PurzBeats @ComfyUI Could you go over options/setup for the outpainting workflow if you don't have much ram/vram? If it's even possible, of course.

I think we're going to explore LTX 2.3 IC LoRAs tomorrow on the @ComfyUI stream. Definitely gonna check out Video Outpainting and Anime2Real -- any other suggestions?

@GoodWillsmith For me Ovalcommers by @markuspopp_oval has changed a lot about my expectations in music ✨

obvious answer for me. the record whose sound i still try to chase in some aspect every time i sit down to make music. so many tracks that i’ve made are temporarily called something like “Icefall 4” before they have a title

@longlivemikey Sounds great, good luck with polishing the process ✨

@robertseidelcom Working on premiere next but AE first I’ll show off the process soon, but basically it similar to how Red Giant builds universe plugins, they didn’t want to let me use their system so I built my own.

I built 7 After Effects plugins today. Actual plugins, not presets, not scripts or extensions. C++ plugins.

@Merzmensch @ComfyUI 🔥🔥🔥

#StyleGAN2 with #ComfyUI. GAN Interpolation video By using Video Combine node, you will be able to generate a video from your latent space. Select the framerate, video format (mp4, gif etc.), and loops if desired. Read more: medium.com/merzazine/styl… @ComfyUI

@Austen 🟥🟢🔷 Was the video model mentioned somewhere?

I’ve never wanted to see a prompt more in my life

@opiumhum Ah, thanks for letting me know!

@robertseidelcom He cancelled unfortunately!

CTM Festival kicks off in 48 hours and while there's plenty to get excited about, nothing tops bringing over DJ Chad, son of my dear friend DJ Rashad, for his first gig abroad, representing the new generation of Chicago footwork. The night is fully sold out already!

@_nic_hamilton_ @Vimeo So in the end, there is only Youtube for all of us?

🔥MLHA🔥Premiering at @signalfestival October 16 – 19 Praha Artwork: Robert Seidel @robertseidelcom Sound: Nikolai von Sallwitz @Sallwitz Location: Clam-Gallas Palace

@_motherslug open.spotify.com/album/1HtDaat9…

Give me your nicest music that's still a little esoteric, experimental, or avant-garde in some way. I need something nice but not pure today.

@matdryhurst Thanks for clarifying… so it’s random scraped data instead of working with specific hq sources ✨

@robertseidelcom In the paper it suggests popular holiday songs would be reproduced countless times in family videos etc, just as that Eminem song is probably duplicated through many thousands of motivational ugc Points to training on YouTube etc

This paper is very cool, although I'd argue its more complicated re: memorization As the paper outlines, the likelihood is that memorization occurs due to overrepresentation in the training data. Memorization is strong with popular holiday songs, and cultural touch points like the Eminem tracks (particularly when coupled with transcripts) that likely appear in at least hundreds of thousands of videos. Of those few overrepresented songs, those with tighter lyric-rhythm coupling are easier to surface. We knew that already from the old Carlini paper that demonstrated the ability to reconstruct the Mona Lisa due to its cultural omnipresence. The Mona Lisa is interesting in that its one of a few paintings that classification systems like CLIP would classify as "Mona Lisa" rather than "generic painting with x features", and it no doubt appears millions of times in a poorly curated dataset. What this paper doesn't say is that training on any song a handful of times will lead to memorization, which is the case for 99.99% of things a model will be trained on. It would be interesting to understand what that threshold is, but of course difficult when no-one releases info on training data. One could also argue that for the extreme minority of media this applies to traditional copyright protections on outputs are sufficient, as in order for a song to be omnipresent enough to be memorized, it will be obvious when it is regurgitated. Good luck releasing a "Lose Yourself-ish" track commercially. Where we all agree is this points to needing much less sloppy pre-training data ingestion practices.

@OdinLovis @Adobe @FAL Interesting! Comfy support could be nice for local tweaking…

I dont know why Content Fill AI for video was not available in @Adobe suite. So I made it as a plugin and it’s open-source for After effects!! github.com/lovisdotio/Con… Super easy Installation in two click for Mac and PC, it’s using @FAL .ai api for the generation. Two mods are possible : - Content fill from mask using Wan 2.1 ( more precise ! ) - Content fill from prompt using Wan 2.2 ( less precise, better quality!) It’s still a prototype and it’s possible that you encounter some bug, that is why I also made it open for now. But could be cool to develop that with much more functionalities ! For automatic download of the final file it’s working good on Mac, but on Pc it will show a popup to download it manually.