Nick | KLIYΞR.eth 👾🥨

364 posts

Nick | KLIYΞR.eth 👾🥨 banner
Nick | KLIYΞR.eth 👾🥨

Nick | KLIYΞR.eth 👾🥨

@nickstracke_

PhD student working on Stable Diffusion at LMU. Founder @PretzelDAO. Background in Data Science and AI. Crypto wizzard.

Munich Katılım Eylül 2016
1.1K Takip Edilen536 Takipçiler
Sabitlenmiş Tweet
Nick | KLIYΞR.eth 👾🥨
Nick | KLIYΞR.eth 👾🥨@nickstracke_·
Can we condition LoRAs to use their great performance for 0-shot adaptations? 🤔 So far, LoRAs have been used extensively but their adaptions remain static after training. We introduce LoRAdapter: efficient and flexible conditioning of LoRAs, which we validate on SD. 🧵👇
Nick | KLIYΞR.eth 👾🥨 tweet media
English
1
5
21
3.9K
Nick | KLIYΞR.eth 👾🥨 retweetledi
Nick Stracke
Nick Stracke@rmsnorm·
Video diffusion models learn motion indirectly through pixels. But motion itself is much lower-dimensional. We introduce 64× temporally compressed motion embeddings that directly capture scene dynamics. This enables efficient planning -> 10,000× faster than video models. 🧵👇
English
10
48
328
44.1K
Nick | KLIYΞR.eth 👾🥨 retweetledi
Nick Stracke
Nick Stracke@rmsnorm·
🤔 What if an LLM could edit the game world itself during gameplay? I built a sandbox engine where an LLM directly modifies the world state in real time. New biomes, new blocks, even new minerals added mid-game 🤯 No scripts, no reload. Video below ⬇️
English
2
1
9
329
Nick | KLIYΞR.eth 👾🥨 retweetledi
Nick Stracke
Nick Stracke@rmsnorm·
Actually had a similar idea last weekend and vibe coded a Minecraft-like game that gets extended with new content and mechanics as the user plays. Code a demo in the reply.
Simo Ryu@cloneofsimo

Trazillion dollar idea. Make a game engine, a decent one, that supports basic stuff, but that could be heavily extended and modified. But make it unconventionally console / test driven, make it easily verifiable. Then, make a lot of toy game with it. Get lot of instructions, Train (fine-tune) an LLM on it. Sell the LLM & game engine as a paired product. Train on the user-generated games. SFT. RLVR. I would do this if I didnt have a job

English
2
1
6
246
Nick | KLIYΞR.eth 👾🥨 retweetledi
jo.schb
jo.schb@jo_schb·
🤔 What if you could generate an entire image using just one continuous token? 💡 It works if we leverage a self-supervised representation! Meet RepTok🦎: A generative model that encodes an image into a single continuous latent while keeping realism and semantics. 🧵👇
jo.schb tweet media
English
8
23
112
20.2K
Nick | KLIYΞR.eth 👾🥨 retweetledi
Stefan Baumann
Stefan Baumann@StefanABaumann·
🤔 What happens when you poke a scene — and your model has to predict how the world moves in response? We built the Flow Poke Transformer (FPT) to model multi-modal scene dynamics from sparse interactions. It learns to predict the 𝘥𝘪𝘴𝘵𝘳𝘪𝘣𝘶𝘵𝘪𝘰𝘯 of motion itself 🧵👇
Stefan Baumann tweet media
English
5
15
39
6.4K
opinions.fun
opinions.fun@opinionsdotfun·
OPINIONS FUN V2 IS NOW LIVE. Turn your opinions into assets. Turn every viral X debate into tradable markets. Opinions matter. Keep reading 🧵
English
147
77
546
122.5K
Nick | KLIYΞR.eth 👾🥨 retweetledi
opinions.fun
opinions.fun@opinionsdotfun·
As we prepare for our upcoming launch, we're excited to announce our acceptance into @heliuslabs Startup Launchpad Opinions > Memes Accelerate
opinions.fun tweet media
English
43
19
87
18.3K
@levelsio
@levelsio@levelsio·
I'll pay you $10,000 If you can make a Replicate model (not ComfyUI, not anything else, ONLY a Replicate model) That takes an image Detects the hands area Fixes the hands consistently: - regular hands - crossed hands - hands facing camera And is fast like 10-20 seconds
@levelsio tweet media@levelsio tweet media
English
179
45
897
934.9K
Nick | KLIYΞR.eth 👾🥨
Nick | KLIYΞR.eth 👾🥨@nickstracke_·
Why does withdrawing from @Blast_L2 require a 14 Day Withdrawal Period when there are no fraud proofs anyway and it's basically just a multisig? Cries in Stage 0 L2 @l2beat ...
English
1
0
4
280
Nick | KLIYΞR.eth 👾🥨 retweetledi
Miguel Angel Bautista
Miguel Angel Bautista@itsbautistam·
New paper! Check out Nick's work on efficient adapters for text-to-image diffusion, this was a fun collaboration between @Apple MLR and LMU. Enabling fine-grained and disentangled control for diffusion models is really interesting! LoRA-style conditional adapters are powerful!
Nick | KLIYΞR.eth 👾🥨@nickstracke_

Can we condition LoRAs to use their great performance for 0-shot adaptations? 🤔 So far, LoRAs have been used extensively but their adaptions remain static after training. We introduce LoRAdapter: efficient and flexible conditioning of LoRAs, which we validate on SD. 🧵👇

English
0
7
24
2.6K
Tanishq Mathew Abraham, Ph.D.
Tanishq Mathew Abraham, Ph.D.@iScienceLuvr·
CTRLorALTer: Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models abs: arxiv.org/abs/2405.07913 project page: compvis.github.io/LoRAdapter/ This paper from the CompVis group demonstrates a novel LoRA adapter approach that can perform image generation based on both structure and style conditioning in a single formulation. "we propose a LoRA-based conditioning mechanism whose behavior changes based on conditioning provided at inference time, enabling zero-shot generalization."
Tanishq Mathew Abraham, Ph.D. tweet media
English
3
22
68
10.1K
Nick | KLIYΞR.eth 👾🥨
Nick | KLIYΞR.eth 👾🥨@nickstracke_·
Overall, LoRAdapter is very flexible and efficient, which we show by gradually reducing the rank of a LoRA conditioned on Style. Even a tiny LoRAdapter can still achieve decent performance!
Nick | KLIYΞR.eth 👾🥨 tweet media
English
1
0
2
209
Nick | KLIYΞR.eth 👾🥨
Nick | KLIYΞR.eth 👾🥨@nickstracke_·
Can we condition LoRAs to use their great performance for 0-shot adaptations? 🤔 So far, LoRAs have been used extensively but their adaptions remain static after training. We introduce LoRAdapter: efficient and flexible conditioning of LoRAs, which we validate on SD. 🧵👇
Nick | KLIYΞR.eth 👾🥨 tweet media
English
1
5
21
3.9K

Keşfet

@opinionsdotfun @levelsio @Blast_L2 @l2beat @Apple @DigThatData @iScienceLuvr @StefanABaumann