Andrew Ng announces short course AI Agents for Image and Video Generation developed with Google Cloud through DeepLearning.AI

Original post

New course: Build AI agents that generate images and videos -- an under-explored frontier. A key to performance is having the agent evaluate its own output, and iterate to improve quality. This short course is built together with @googlecloudtech and taught by Katie Nguyen and Wafae Bakkali.

You'll learn three evaluation techniques and combine them in an agent: image-text similarity scoring to check the output matches the prompt, an LLM judge that scores against custom criteria like brand consistency, and structured rubrics that break a prompt into verifiable yes/no questions like "is the subject in the frame?" and "does the camera motion match?"

Skills you'll gain: - Learn image and video prompt engineering - Build an image agent that turns brand guidelines into UI mockups - Build a video agent that plans multi-scene explainers and animates reference frames with synchronized audio

Join and build agents that create images and video!

https://www.deeplearning.ai/courses/ai-agents-for-image-and-video-generation

10:08 AM · May 20, 2026 · 89K Views

VIEWS337BOOKMARKS1LIKES1RETWEETS1

Zhou Yu@Zhou_Yu_AI

Evaluation is one of the hardest parts of building agents that work at scale, especially when there's no single correct answer. If you're thinking about this for conversational agents, we built ArkSim for that. It simulates realistic interactions, evaluates outputs turn-by-turn, and catches failures before deployment.

Give it a try: https://github.com/arklexai/arksim

Andrew Ng@AndrewYNg

Join and build agents that create images and video!

https://www.deeplearning.ai/courses/ai-agents-for-image-and-video-generation

40d33711

REPLIES3

rashad@rashad1791400

@AndrewYNg @GoogleCloudTech The self-evaluation loop is the real breakthrough here -- most agents just generate, but iterating on quality is what separates useful tools from toy demos. Looking forward to seeing how enterprises apply these techniques.

40d36

jon flaig@jonrobflaig

@AndrewYNg @GoogleCloudTech CC-Suite can do this out of the box.

40d31

Tanmay Saboo@TanmaySaboo

@AndrewYNg @GoogleCloudTech Self-evaluating agents that score their own image output against rubrics is a smart design. Most skip that feedback loop and wonder why quality stays inconsistent. This one is worth the time. 🧠

40d88

Suyash Joshi@suyashcjoshi

@AndrewYNg @GoogleCloudTech I'm teaching a similar course in London next week: https://www.meetup.com/creativecodeai/events/314645727/?eventOrigin=group_upcoming_events

40d78

𝕺𝖋𝖋𝖎𝖈𝖎𝖆𝖑 𝕯𝖆𝖛𝖎𝖉💫@DaveJoOpinion

@AndrewYNg @GoogleCloudTech

40d211

𝕺𝖋𝖋𝖎𝖈𝖎𝖆𝖑 𝕯𝖆𝖛𝖎𝖉💫@DaveJoOpinion

@AndrewYNg @GoogleCloudTech

40d181

Thomas Tao@Thomas_Tao_1

@AndrewYNg @GoogleCloudTech Self-eval loops are probably where most of the quality jump comes from. I’ve seen them drift fast though when the rubric is vague.

40d53

kayvon! (building murray ai)@MurrayAI

Self-evaluation + iteration is the real unlock for agentic systems.

Whether it’s generating images/video or operating in physical environments, the agents that actually ship are the ones that can critique their own output against clear criteria and improve (or hand off) instead of just firing once.

Excited to see more courses focused on this.

40d37

AgenticSource@AgenticSourcing

@AndrewYNg @GoogleCloudTech Does the course cover when an agent should retrieve instead of generate? Real logos, product photos, brand assets agents shouldn't fake. That's the gap we're working on at AgenticSource. Bookmarked.

40d37

Oleg Karakash@olegk_ai

@AndrewYNg @GoogleCloudTech Self-evaluation works here because image quality is actually measurable. It goes wrong when teams copy the loop into domains with no ground-truth signal, and the agent just iterates confidently toward its own bias.

40d34

Jahanzaib Ahmed@jahanzaibai

@AndrewYNg @GoogleCloudTech That eval loop is probably where it gets expensive. You're not scoring tokens, you're running a VLM judge on every frame and the latency stacks fast.

40d31

hodisr@hodisra

@AndrewYNg @GoogleCloudTech That eval loop converges fast in image-gen because the signal is visual and immediate — the credit-DD analog is harder for the exact same reason. Ground truth lives in human exception reviews, label velocity is weeks not seconds, and the loop has to be designed around that gap.

40d31

varun@vazuzu_varun

@AndrewYNg @GoogleCloudTech seems very interesting, im gonna build the next Higgsfield

40d25

TomAIdaily@TomAIdaily

@AndrewYNg @GoogleCloudTech bet the agent will fail the brand consistency check on its first try just like we all do with coffee

40d23

Postspark | AI Content Repurposing@Michael43981

@AndrewYNg @GoogleCloudTech Most people still think AI is just about generating content faster. The real shift is AI systems that evaluate, refine, and improve their own creative output. That’s where the next generation of content tools will win. 🔥 Huge opportunity for builders creating AI-native creative.

40d61

Agung Hermawan@herzmwn_agung

The self-eval loop is the real unlock here. Qwen3.7-Max just launched as "The Agent Frontier" and Forge showed guardrails take an 8B model from 53% to 99% on agentic tasks — both converging on the same thesis. Most people focus on generation quality. The production edge is the agent catching its own mistakes before humans see them.

40d16

Arthur Tse@xiaofengxie

The structured rubric approach (breaking prompts into verifiable yes/no questions) is the underrated part here. Most teams stop at similarity scoring and LLM judges, but it's the rubric layer that catches subtle failures — wrong number of objects, incorrect composition, brand color mismatches.

Anyone trying this at scale running into cost/latency tradeoffs with VLM judges on every frame?

40d15

Eclipse 🌖@ECLresearch

@AndrewYNg @GoogleCloudTech The self-evaluation loop is the real unlock—most agent pipelines treat generation as fire-and-forget. Curious if the course covers reward model design for visual quality scoring.

40d13

Charlotte Ketelaar@CZJPKW

@AndrewYNg @GoogleCloudTech self-evaluation loops are the unlock i keep pushing founders on. building our own agent stack at Capwave, eval-and-iterate doubled output quality more than any model swap we tried. underrated lever.

40d12