Agneet Chatterjee

CS PhD @ASU | Controllable Media Generation

[Agneet Chatterjee @agneet42]

As video generative models reshape how we consume visual media, we ask: what controls are essential for them to meaningfully support professional workflows?

Stable Cinemetrics will be presented at NeurIPS 2025. Please visit our (interactive) webpage for detailed findings, insights and more : https://t.co/3oJSBlYQ3z

In this project, we had the unique opportunity to receive feedback from an Academy Award winner. While projects may come and go, that is one experience I will always cherish!

[Varun Jampani @jampani_varun]

🎬 Introducing Stable Cinemetrics, to be presented at NeurIPS 2025.

We present the first taxonomy of professional controls to systematically study and control video generative models through the lens of filmmaking.

Interactive webpage with paper link: https://t.co/Eh4Hw3hBZl
🧵

[Zhiqiu Lin @ZhiqiuLin]

Before AI can generate professional videos, it needs to see like a professional.

We spent a year with 100+ content creators teaching AI to describe video like a filmmaker would.

Introducing CHAI: Critique-based Human-AI Oversight for Building a Precise Video Language [CVPR'26 Highlight, Top 3%].

Try prompting a video generator for a dolly zoom, dutch angle, point of view, or camera roll. Most fall back to the same bland defaults: a push-in, a level shot, a third-person view. Why? These techniques require a language of cinema that current models rarely speak.

We built that language:

1️⃣ Precise specification: 5-aspect structured captions co-designed with professional cinematographers covering subject, scene, motion, spatial, and camera dynamics

2️⃣ Scalable oversight: LLMs draft captions, humans critique what's wrong and how to fix it

3️⃣ Post-training recipes: Qwen3-VL-8B surpasses Gemini-3.1 and GPT-5

4️⃣ Video generation: fine-tuned Wan follows 400-word cinematic prompts with precise control

Here's how each works 🧵

Work led by CMU and Harvard with @chancharikm, @du_yilun, and @RamananDeva.

📄 Paper: https://t.co/wCwEtvrntM
🌐 Site: https://t.co/oAAQklGrfF

[Agneet Chatterjee @agneet42]

Amidst today's tsunami of papers, it's quietly validating as an early career researcher to see your past work cited by multiple reviewers in their @iclr_conf reviews!

[Varun Jampani @jampani_varun]

Looking forward to ICCV’25! 3 workshop talks, 1 panel and 7 papers.

Keynote talk on “Crafting Video Diffusion: Precise Inputs and Rich Outputs” in 3 workshops:
https://t.co/LekqWM8Xil
https://t.co/jdNDbjVSxY
https://t.co/vMJqiph7M3

Panel discussion at https://t.co/WYyrnUMCR0

[Agneet Chatterjee @agneet42]

@RisingSayak Not an expert in preference datasets: But is there also merit in scoring images? Chosen/Rejected is a form of hard scoring whereas a soft score (e.g. between 1-10) might also be beneficial to the community (and for models) ?

[Agneet Chatterjee @agneet42]

@Dazitu_616 Very cool work! Is there an ETA on the code release?

[Agneet Chatterjee @agneet42]

Feels good to be recognized!

[EMNLP 2026 @emnlpmeeting]

We're kicking off the awards session at #EMNLP2024 by announcing our (many) **Outstanding Reviewers**!

[Agneet Chatterjee @agneet42]

@sourajitCS @trgokhale @wacv_official Congratulations to both you and @trgokhale !

[Sayak Paul @RisingSayak]

Looking forward to our SPRIGHT poster session at #ECCV2024 today with @agneet42.

It's #213 and we will be there from 4:30 PM to 6:30 PM (CEST).

But I am here at ECCV today almost for the day. Looking forward to chats 🇮🇹

https://t.co/DY4DPnEkVZ

[Agneet Chatterjee @agneet42]

Looking forward to @eccvconf and Milan! I'll be presenting 2 papers:

REVISION (https://t.co/rr6i97qiZB): 10/1, 10.30-12.30, #140

SPRIGHT (https://t.co/jFt26Eu4Hm) w @RisingSayak: 10/2, 16.30-18.30, #213

Let's talk generative models -- or even better, risotto recommendations!

['YZ' Yezhou Yang (杨叶舟) @prof_yz]

Dear friends, from Sun Sep 29th, @ApgAsu ers will present 👇papers, DC poster, organize tutorial and workshop paper @eccvconf with a focus on semantically precise #T2I, secure #GenAI and a survey of Recent Event Camera Innovations.

Please chat with our talented members! 🙏🤠

[Agneet Chatterjee @agneet42]

Joint work with @FPSLuozi @trgokhale @prof_yz @cbaral

Both REVISION and SPRIGHT (https://t.co/yIiP6grOth) will be presented at @eccvconf and are efforts towards improving the 3D understanding of current vision-language models.

[Agneet Chatterjee @agneet42]

While we conduct most of our experiments on Stable Diffusion 1.4 and 1.5, our pipeline can be extended to incorporate larger models, for increased control over backgrounds, colors or style.

[Agneet Chatterjee @agneet42]

We also develop the RevQA benchmark, to evaluate spatial reasoning abilities of multimodal LLMs. RevQA is a question-answering benchmarking which has 16 diverse question types and their adversarial variations consisting of negations, conjunctions, and disjunctions.

[Agneet Chatterjee @agneet42]

REVISION elevates the ability of existing T2I models to generate spatially accurate images. Given an input prompt, we generate a synthetic image using REVISION, which is used as additional guidance during image generation.

[Agneet Chatterjee @agneet42]

REVISION parses a prompt into assets (objects) along with the spatial relationship between them and synthesizes a symbolic image in Blender, placing the respective object assets at coordinates corresponding to the parsed spatial relationship.

[Agneet Chatterjee @agneet42]

Happy to share our latest work, REVISION, which will be presented at #ECCV2024.

Project Page : https://t.co/tI9My1aBP9

With REVISION, we combine the controllability of graphics rendering engines and the photorealism of T2I models to improve spatial fidelity.

🧵

[Brian Bartoldson @bartoldson]

Adversarial attacks jailbreak models. Existing defenses don’t even safeguard simple models. In our ICML 2024 paper on "Adversarial Robustness Limits”, we show how scaling helps defense, up to the point where attacks start to fool humans (take quiz: https://t.co/mmLDnvTErY). 🧵1/n

[Sayak Paul @RisingSayak]

Very pleased to see this accepted at #ECCV2024. See you in Milan to talk about diffusion models and other adjacent areas of work. 🇮🇹

Kudos to our truly global team 😉

[Agneet Chatterjee @agneet42]

@saakur3 Thanks Sathya!

[Agneet Chatterjee @agneet42]

Happy to share that 2/2 papers are accepted to #ECCV2024, with 1 of them being SPRIGHT. 👇

Congratulations to all the co-authors. See you in Milan!

[AK @_akhaliq]

Getting it Right

Improving Spatial Consistency in Text-to-Image Models

One of the key shortcomings in current text-to-image (T2I) models is their inability to consistently generate images which faithfully follow the spatial relationships specified in the text prompt. In