Dmitry Baranchuk

GenCV Researcher

[Dmitry Baranchuk @DmitryBaranchuk]

Excited to share our new open-source SFT dataset for text-to-image generation! Feel free to use it in your research or product.
🤗: https://t.co/cH93Rrgp1W

[Sergey Kastryulin @snk4tr]

Quality SFT data for text2image is key, but often closed. We used a pre-trained diffusion model to select optimal samples, creating Alchemist: an open, general-purpose SFT dataset, outperforming LAION-Aesthetics V2. Link to the method & 5 fine-tuned SD models below 👇(1/9)

[Dmitry Baranchuk @DmitryBaranchuk]

Closing thoughts: unlike cascaded DMs or continuous-token AR, which run diffusion separately at each stage, SwD unifies scaling and diffusion into a single process. Thus, SwD can be seen as a natural continuous variant of next-scale prediction models (VAR/Switti/Infinity) (9/9)

[Dmitry Baranchuk @DmitryBaranchuk]

I'd like to share our new diffusion distillation method, SwD, which produces few-step generators with progressive resolution scaling over the diffusion process. On SD3.5, SwD matches the speed of two full-size steps but with much better quality. Demo & models are released. (1/9)

[Dmitry Baranchuk @DmitryBaranchuk]

We release our models, code and demo. Feel free to try it out and share your feedback :)

Page: https://t.co/APLUfrGCSb

Paper: https://t.co/HunSBP2ZkO

Demo: https://t.co/eGXJBzG8zZ

Code: https://t.co/RZYTP4ydyn (8/9)

[Dmitry Baranchuk @DmitryBaranchuk]

@ab_testing53 flux is clearly in a higher league. We acknowledge that it is better. Switti should be considered as a step in developing the new paradigm but not as a production-grade model like flux, midjorney, ideogram, etc

[Dmitry Baranchuk @DmitryBaranchuk]

We introduce Switti, a scale-wise transformer for T2I generation. Our 2.5B model outperforms existing AR models and competes with SOTA diffusion models of the same size, while being up to 7x faster! We release our models and code, along with a demo to play around.(1/10)

[Dmitry Baranchuk @DmitryBaranchuk]

There is still much work to be done to achieve top-tier t2i model quality such as Midjorney, FLUX, etc. As the next step forward, we hope to integrate Switti with higher quality scale-wise image tokenizers, e.g., Infinity [1] released today.

[1] https://t.co/M0VY0FRaTo (10/10)

[Dmitry Baranchuk @DmitryBaranchuk]

We release our models and code, along with a demo to play around.

https://t.co/AcG39AD3P8

https://t.co/1kLJk9a2dn

https://t.co/U9JYPlEAlZ

https://t.co/JvSE4B4YjF (9/10)