Engineer, Mathematician, Computer Scientist
San Francisco, CA
Joined October 2007
- Tweets3K
- Following579
- Followers628
- Likes18.4K
Pablo Delgado retweeted
Today's AI models are a technical marvel. With enough expertise, you can steer them almost anywhere. But by default they have to serve many masters, and easy prompts give you bland results.
Magnific is not build for "anyone." We serve creatives, and the default mode on these models is not what our users want.
Months ago we started a project with these goals:
1) built on top of existing models. Leverage the existing AI models as the building block. Get better when models get better.
2) make the default generation by and for creatives. Easy to get great results.
3) trivial to stay on brand.
A new model from scratch will be something that will become obsolete in weeks / months. Our approach enhances them for our users, making a useful contribution that is long-lasting and differentiated.
After many false starts, we landed on a pre, post-processing, and routing layer on top of a few of today's state of the art models.
We tested the model first with our employees, and later with external users, on two tests: best image for a given prompt, and best image overall (blinding the prompt). The first checks prompt adherence, the last checks image aesthetics.
Magnific One came out first on both tests by a wide margin.
You can use it exclusively on @magnific, test it yourself and tells us what you think.
Pablo Delgado retweeted
Introducing Kardashev-0.7, the worldโs first trained swarm made of 32 distinct models
Trained with RL for Population Scaling (RLPS), 32 models organically develop specialization & complementary capabilities, delivering frontier performance at:
- 0.007x ~ 0.02x of the inference cost
- 0.03x of the required memory
Civilization advances through different minds specializing and working together. Weโre bringing that principle into AI
@BanburyRoadAI, weโre scaling intelligence by model count, toward civilizations of models that learn to build on one another
Pablo Delgado retweeted
@MicrosoftAIโs Suqiang (Jack) Song is speaking at #Reverie2026.
Hear how MAI is building the frontier AI data flywheel, from curation and deduplication to search and training, with different workloads running on the same underlying data.
Apply to attend โ na2.hubs.ly/H084LFl0
Pablo Delgado retweeted
See you at Neurips! Letโs talk world models!
stable-worldmodel got into neurips 2026!
it's basically everything we keep rebuilding for world model research (data loading, baselines, planning, evals) in one place so nobody has to do it again
arxiv.org/abs/2605.21800
Pablo Delgado retweeted
๐ Sol-Attn is here!
We present a training-free sparse attention method that accelerates video generation while better preserving quality.
Sol-Attn unifies dynamic routing, sparse computation, and approximate correction in a single online-softmax pass:
โข On-the-fly block thresholding for dynamic yet controllable budgets
โข Proxy-score reuse to approximate unselected blocks
Results (vs dense FlashAttention-3):
โข Wan 2.1-14B: 2.02ร end-to-end
โข HunyuanVideo-13B: 2.12ร end-to-end
โข LTX 2.3: up to 2.4ร end-to-end
When integrated into Sol-Engine (with kernel fusion + caching):
โข Wan 2.1-14B: 3.48ร end-to-end
โข HunyuanVideo-13B: 5.08ร end-to-end
Already available in Sol-Engine.
The B200 kernel is still under further optimization.
๐ฌ Project: nvlabs.github.io/Sana/Sol-Atโฆ
๐ Paper: arxiv.org/abs/2607.24027
๐ Code: github.com/NVlabs/Sana/tree/โฆ
Pablo Delgado retweeted
Very cool to see @netflix releasing video datasets and models on @huggingface. They have a strong video AI team and would be amazing to see them open-source more!
huggingface.co/netflix
Pablo Delgado retweeted
Data processing has become a GPU workload and is dominated by inference. A new architecture is required, and it is not Spark on GPUs. anyscale.com/blog/data-proceโฆ
Pablo Delgado retweeted
1/ World model research is fragmented: every paper reimplements its own data pipeline, baselines, and eval harness. Comparing two methods fairly is weeks of infra work.
๐๐๐ฎ๐ฏ๐น๐ฒ-๐๐ผ๐ฟ๐น๐ฑ๐บ๐ผ๐ฑ๐ฒ๐น is a new open-source platform that standardizes the whole thing: arxiv.org/abs/2605.21800
Pablo Delgado retweeted
in case you missed it @lancedb and HF are partnering up to unlock the next generation of large dataset storage on the Hub ๐ฅ
And it's fire!
- Supports storing embeddings (and their indexes) directly alongside the data
- Vector search / similarity search is built-in
- Large multimodal datasets (text, images, video)
just use the hf:// prefix:
db = lancedb. connect("hf://datasets/julien-c/hub-stats-lance")
๐ฅ๐ฅ
Pablo Delgado retweeted
Most video models are silent.
Most audio models donโt see.
LTX-2 learns the joint distribution of sound and vision, generating speech, foley, ambience, motion, and timing together not as a post-hoc pipeline.
Pablo Delgado retweeted
HiStream
Meta AI researchers introduce an efficient autoregressive framework for 1080p video generation. By eliminating spatial, temporal, and timestep redundancy, HiStream achieves state-of-the-art quality with up to 107.5ร speedup, making high-resolution video generation practical.
Pablo Delgado retweeted
Our session is on Day 2, Nov 5 at 4pm. Check out the agenda here: anyscale.com/ray-summit/2025โฆ
Pablo Delgado retweeted
Weโll walk through how Ray enables large-scale processing across hundreds of GPUs, while LanceDBโs columnar design provides efficient, intelligent curation and sampling. Together, theyโre producing smaller, more diverse, and higher-quality datasets for cutting-edge text-to-image and video-to-text research.
Pablo Delgado retweeted
Building and curating large-scale multimodal datasets has long been a complex, resource-heavy challenge. But thatโs changing fast. Lei Xu of LanceDB and Pablo Delgado of @netflix will be speaking at Ray Summit 2025 โ Scaling Multimodal Data Curation with Ray and LanceDB
Pablo Delgado retweeted
Now this is a @lancedb feature to make @Noahpinion proud. Introducing RabitQ: better compression, better recall, faster index build, higher throughput.
lancedb.com/blog/feature-rabโฆ
Pablo Delgado retweeted
๐ฅณ Welcome another #Lancelot at the Roundtable, Ethan Rosenthal ๐
On Ethanโs first day at @runwayml , he was tasked with building a multimodal ๐ฑ๐ฎ๐๐ฎ ๐๐๐๐๐ฒ๐บ ๐๐ต๐ฎ๐ ๐๐๐ฝ๐ฝ๐ผ๐ฟ๐๐ฒ๐ฑ ๐ฏ๐ผ๐๐ต ๐ฑ๐ถ๐๐๐ฟ๐ถ๐ฏ๐๐๐ฒ๐ฑ ๐ฑ๐ฎ๐๐ฎ๐น๐ผ๐ฎ๐ฑ๐ถ๐ป๐ด ๐ฎ๐ป๐ฑ ๐ฒ๐
๐ฝ๐น๐ผ๐ฟ๐ฎ๐๐ผ๐ฟ๐ ๐ฑ๐ฎ๐๐ฎ ๐ฎ๐ป๐ฎ๐น๐๐๐ถ๐. He said, โ๐๐ฉ๐ข๐ตโ๐ด ๐ข ๐ต๐ฆ๐ณ๐ณ๐ช๐ฃ๐ญ๐ฆ ๐ช๐ฅ๐ฆ๐ข. ๐ ๐ฐ๐ถ ๐ด๐ฉ๐ฐ๐ถ๐ญ๐ฅ ๐ฏ๐ฆ๐ท๐ฆ๐ณ ๐ต๐ณ๐บ ๐ต๐ฐ ๐ฅ๐ฐ ๐ต๐ฉ๐ช๐ด ๐ธ๐ช๐ต๐ฉ ๐ฐ๐ฏ๐ฆ ๐ด๐บ๐ด๐ต๐ฆ๐ฎ!". He then found #Lance and did exactly what he said not to do. ๐
Pablo Delgado retweeted
Video diffusion models struggle beyond training resolution โ artifacts & repetition.
๐ฅCineScale๐ฅ solves this with a novel inference paradigm:
โก Dedicated variants for video architectures
โก Extends T2I to T2V & I2V & V2V
โก 8K images & 4K video, tuning-free/minimal tuning
Expanding the frontier of generative video fidelity.
โ Kudos to the teamwork led by our intern
@qhnmoon at @eyelinestudios.
#AI #AIResearch #MachineLearning #AIGC #GenAI #videos #DiffusionModels #HighRes #fidelity #ComputerVision #internship
@eyelinestudios and @NTUsg's latest research paper, CineScale, showcases a new method for creating higher-resolution image and video content with novel adaptations for a variety of visual generative model architectures.
Unleash the resolution of text-to-video, image-to-video, and video-to-video generative diffusion models with minimal tuning! Watch our 4k video generation demo below. #CineScale
Kudos to the team: @qhnmoon, @realNingYu, @ziqi_huang_, @debfx, @liuziwei7
Paper: arxiv.org/pdf/2508.15774
Project: eyeline-labs.github.io/CineSโฆ
Code: github.com/Eyeline-Labs/Cineโฆ
***This is part of the ongoing research and development at @eyelinestudios, and we hope to see adoption of these techniques and workflows soon.