Srikar Yellapragada

PhD student @SBUcompsc

[Srikar Yellapragada @ymsrikar]

Excited to share that our paper "ZoomLDM: Latent Diffusion Model for multi-scale image generation" will be presented at @CVPR 2025 #CVPR2025 !

Huge thanks to our collaborators - @algreco_ @KostasTriaridis @PrateekPrasana & others

https://t.co/OyUNfyYraH

[Srikar Yellapragada @ymsrikar]

@photoroom Impressive results! How did you condition the model?

[Srikar Yellapragada @ymsrikar]

@sainingxie Nice paper! I'm curious to know why researchers still use the SD-1.5 VAE for imagenet training, given that superior alternatives like the FLUX VAE (or RAE ?) exist.

[Srikar Yellapragada @ymsrikar]

@jo_schb Really nice work! We explored similar ideas in our CVPR’24 and '25 papers [1,2], where we condition LDMs on SSL embeddings (for pathology and satellite domains). Later extended in PixCell [3]

[1] arxiv: 2312.07330
[2] arxiv: 2501.04784
[3] arxiv: 2506.05127

[Srikar Yellapragada @ymsrikar]

@PACELab_SBU @ACMSigmetrics unbiased variance = ∞

[Srikar Yellapragada @ymsrikar]

@CVPR

[Srikar Yellapragada @ymsrikar]

Why won't they let us stand inside the workshop rooms? :( @CVPR #cvpr2025

[Srikar Yellapragada @ymsrikar]

@CVPR Update - there is unrest among the crowd

[Srikar Yellapragada @ymsrikar]

@CVPR It doesn't make sense because we selected our choice of workshops during registration. I just left the room to get a coffee :(

[Prateek Prasanna @PrateekPrasana]

Introducing PixCell — the first generative foundation model for digital histopathology!

🖼️ trained on 30M patches, released 1M curated H&E patches (SBU‑1M)
📏 Progressive diffusion‑transformer training
🧩 Conditioned on foundation model embeddings for style‑aware synthesis
📦 Open weights for PixCell‑256 and PixCell‑1024

Use cases (more to follow):
�� SOTA realism for targeted data‑augmentation
• Cell‑mask‑guided generation 🔬
• H&E ↔ IHC stain translation (no IHC training!) 🎨

Preprint, code, and models: https://t.co/9qGFmFfFph

@ymsrikar @algreco_ @SaltzJoel @varunbel @SaarthakKapse @KostasTriaridis @sbubmi @sbucompsc
#PixCell #PathologyAI #GenAI #diffusion #opensource

[Srikar Yellapragada @ymsrikar]

🚀Check out the code and trained models for ZoomLDM: https://t.co/07E8MGcIh9

We'll also present our poster at CVPR in Paris! Details: https://t.co/4kel8DYCs2

[Srikar Yellapragada @ymsrikar]

Excited to share that our paper "ZoomLDM: Latent Diffusion Model for multi-scale image generation" will be presented at @CVPR 2025 #CVPR2025 !

Huge thanks to our collaborators - @algreco_ @KostasTriaridis @PrateekPrasana & others

https://t.co/OyUNfyYraH

[Srikar Yellapragada @ymsrikar]

@cloneofsimo What's the issue with training a dinov2 on imagenet-1k, then using it for diffusion training?

[Srikar Yellapragada @ymsrikar]

@gabriberton Would they overtake CVPR if merged?

[Srikar Yellapragada @ymsrikar]

We propose a joint multi-scale sampling approach, allowing us to efficiently synthesize histopathology images of up to 4096x4096 pixels.

[Srikar Yellapragada @ymsrikar]

Excited to share that our paper "ZoomLDM: Latent Diffusion Model for multi-scale image generation" will be presented at @CVPR 2025 #CVPR2025 !

Huge thanks to our collaborators - @algreco_ @KostasTriaridis @PrateekPrasana & others

https://t.co/OyUNfyYraH

[Srikar Yellapragada @ymsrikar]

We present the first multi-scale conditional LDM that generates images in histopathology and satellite image data at multiple scales.

[Srikar Yellapragada @ymsrikar]

@docmilanfar It's just Pawbability and catistics

[Srikar Yellapragada @ymsrikar]

@giffmana Based DoTA-2 reference

[Srikar Yellapragada @ymsrikar]

If you're a guy in your early 20s, buy an Nvidia H100 GPU cluster.

Go into debt if you have to

[Steven @StevenPulteFam]

If you’re a guy in your early 20s, buy a Rolex.

Go into debt if you have to

[Srikar Yellapragada @ymsrikar]

@avni321 LGTM 🙏🫡

[Srikar Yellapragada @ymsrikar]

Copilot hit its limit, I guess I'll have to read the documentation... like it's 2022

[Srikar Yellapragada @ymsrikar]

@ashleymcnamara Feeling blessed 🫡