Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

The model sometimes enters a self-critique loop by itself, but you can trigger this manually, and the model tunes the prompt for itself through self-conversation. [Add e.g., "Verify the image, if it's incorrect, write your own prompt, try again, and repeat the process." ]

Mostafa Dehghani

11,563 subscribers

20,234 views • 1 year ago •via X (Twitter)

Education Science & Technology

Anya Rossi• Live Now

Private livecam show

11 Comments

Mostafa Dehghani1 year ago

I've been having a lot of fun seeing all the examples from folks trying Gemini’s native image generation! So many impressive edits (yes, Gemini is the best at this), and people love the magical consistency & pixel-perfect tweaks.

Mostafa Dehghani1 year ago

But what really sets native image generation apart from stand-alone models is that it’s your multimodal inspiration engine. Here are a few fun ways to push interleaved text-image generation (which isn’t perfect yet but already very impressive) even further:

Mostafa Dehghani1 year ago

Use it as your "multimodal brainstorming partner." Ask for creative ideas & let the model generate its own prompt for the images. At least for me, they’re always better than mine!

Mostafa Dehghani1 year ago

If the model makes a mistake, don’t reset. Don’t even tell it how to fix it. Just tell the model that it's wrong and ask to fix it. It's fun to watch the model figuring it out. (I’ve found Gemini to be very polite in such interactions😀)

Mostafa Dehghani1 year ago

I picked up this trick for force triggering self-critique from @a7b2_3 🧠🧠🧠

Jordan Ross1 year ago

I’ve helped 800+ agencies generate over $200,000,000 in revenue. But I’m often asked this question: “How did you even get 800+ clients?” The answer is pretty simple. A clear offer and diversified acquisition system. Let me explain.

Aman Madaan1 year ago

neat! reminds me of this from the old times

James1 year ago

What am I doing wrong?

Mostafa Dehghani1 year ago

Looks like the image might have triggered a safety filter that prevents editing certain types of content. You can try a different image... we're actively working on improving safety filters to reduce false positives!

Hiranmay Darshane1 year ago

This is beautiful. Please take it further with such a verifier-based post-training approach. These abilities might emerge naturally later. Till then why not change the system prompt to include this self-critique ability?

PlayCraft.fun1 year ago

Tried to generate tilesets from our base ones for our games. It's almost there. I wish we could control more the output. Any advices?

Related Videos

In this week’s Stability Seconds, we’re showing you how to use prompt weighting, a technique that helps you control which parts of your prompt have more or less influence on the final image. It’s a fast way to steer your image toward your desired output without rewriting your prompt. Here’s how to do it: 1️⃣ Emphasize elements: Put parentheses around the part of your prompt you want to focus on, then add a colon and a number. For example, (trench coat:1.5) tells the model to emphasize the trench coat. The higher the number, the more it stands out. 2️⃣ De-emphasize elements: Put parentheses around the part of your prompt you want to make less prominent, then add a colon and a number below one. For example, (background:0.5). The lower the number, the less it stands out. You can try prompt weighting with Stable Image Ultra here 👉

In this week’s Stability Seconds, we’re showing you how to use prompt weighting, a technique that helps you control which parts of your prompt have more or less influence on the final image. It’s a fast way to steer your image toward your desired output without rewriting your prompt. Here’s how to do it: 1️⃣ Emphasize elements: Put parentheses around the part of your prompt you want to focus on, then add a colon and a number. For example, (trench coat:1.5) tells the model to emphasize the trench coat. The higher the number, the more it stands out. 2️⃣ De-emphasize elements: Put parentheses around the part of your prompt you want to make less prominent, then add a colon and a number below one. For example, (background:0.5). The lower the number, the less it stands out. You can try prompt weighting with Stable Image Ultra here 👉

Stability AI

13,818 views • 1 year ago

2. Use the generated image as an image prompt for Runway or any video model that supports End-Frames. Use the image as the final/last frame and create a prompt that reveals the image in an ink-style animation. Example prompt:

2. Use the generated image as an image prompt for Runway or any video model that supports End-Frames. Use the image as the final/last frame and create a prompt that reveals the image in an ink-style animation. Example prompt:

Halim Alrasihi

14,596 views • 1 year ago

ON THE APP: → Tap on 'Imagine' in the app → Choose 'Video' from the prompt box → Tap on 'Animate Photos' → Select up to 7 images → Write your prompt → Tap the reference image while the cursor is where you want it to appear in the prompt. Repeat for every image. → Check the settings before sending the prompt for aspect ratio, video duration, and resolution. → Hit send and wait. Now you can use the same character, object, scene, or anything else across different videos.

ON THE APP: → Tap on 'Imagine' in the app → Choose 'Video' from the prompt box → Tap on 'Animate Photos' → Select up to 7 images → Write your prompt → Tap the reference image while the cursor is where you want it to appear in the prompt. Repeat for every image. → Check the settings before sending the prompt for aspect ratio, video duration, and resolution. → Hit send and wait. Now you can use the same character, object, scene, or anything else across different videos.

Dogan Ural

27,224 views • 4 months ago

Something we discovered by accident: what happens if we start Genie 3 from a video and a completely unrelated prompt? Turns out the model really, really wants to make it work, to the point where it emulates itself. The prompt in this one is about a trex on a tropical island.

Something we discovered by accident: what happens if we start Genie 3 from a video and a completely unrelated prompt? Turns out the model really, really wants to make it work, to the point where it emulates itself. The prompt in this one is about a trex on a tropical island.

Jakob Bauer

685,174 views • 11 months ago

The most important prompt Give me an answer. Then Destroy your own Answer. Add this to the end of every prompt you do. And see the result!

The most important prompt Give me an answer. Then Destroy your own Answer. Add this to the end of every prompt you do. And see the result!

Nuseir Yassin

15,910 views • 2 months ago

Is this the future of advertising? With Google Veo 3, you can prompt for a specific scene, character and dialogue. Imagine when you can add your own products to the mix.

Is this the future of advertising? With Google Veo 3, you can prompt for a specific scene, character and dialogue. Imagine when you can add your own products to the mix.

Jerrod Lew

69,217 views • 1 year ago

✨ New mini feature that helps you write prompts today that I got the idea from getimg .ai, an [ AI Improve ] button under the prompt box It takes your prompt, expands and improves it and adds detail One of ChatGPT's strengths is you ask it to make an image but it turns it into a really great big image prompt for its image model A detailed prompt gets better image results than a simple prompt generally: "model with green hair in magical forest" -> "a male model with striking green hair stands amidst a magical forest, where the canopy is alive with luminescent foliage. the sunlight filters through the mist, casting ethereal beams of light that dance upon the moss-covered ground. his hair appears to blend seamlessly with the vibrant emerald shades around him, almost glowing with its own mystical energy. the model wears an intricately detailed outfit that mirrors the forest's hues, adorned with delicate vines and leaves. in the background, the forest is alive with enchanted creatures fluttering about, their eyes luminescent and curious. the atmosphere is serene and otherworldly, evoking a sense of wonder and enchantment. the scene is captured with a high level of detail and a soft focus around the edges, enhancing the dreamlike quality."

✨ New mini feature that helps you write prompts today that I got the idea from getimg .ai, an [ AI Improve ] button under the prompt box It takes your prompt, expands and improves it and adds detail One of ChatGPT's strengths is you ask it to make an image but it turns it into a really great big image prompt for its image model A detailed prompt gets better image results than a simple prompt generally: "model with green hair in magical forest" -> "a male model with striking green hair stands amidst a magical forest, where the canopy is alive with luminescent foliage. the sunlight filters through the mist, casting ethereal beams of light that dance upon the moss-covered ground. his hair appears to blend seamlessly with the vibrant emerald shades around him, almost glowing with its own mystical energy. the model wears an intricately detailed outfit that mirrors the forest's hues, adorned with delicate vines and leaves. in the background, the forest is alive with enchanted creatures fluttering about, their eyes luminescent and curious. the atmosphere is serene and otherworldly, evoking a sense of wonder and enchantment. the scene is captured with a high level of detail and a soft focus around the edges, enhancing the dreamlike quality."

@levelsio

132,473 views • 1 year ago

Through the length of SDXL-Turbo you can see the image forming in the activations like an object taking shape in the afternoon clouds. Like David emerging from a block of granite. At two-thirds through the model you can see the image clearly. It's spatially crisp, refined. Then in the last third of the model something strange but predictable happens: the model translates this well-formed image back into noise. By the output of the U-net, the activations are again completely uninterpretable. Why does this happen? Because we're asking the model to predict the noise rather than the clean image. This is awkward. Like Michelangelo visualizing the chips of granite surrounding the David rather than the David itself. Like asking a friend to look at a cloud and imagine not a shape, but the extra bits of cloud that when removed will leave a shape remaining.

Through the length of SDXL-Turbo you can see the image forming in the activations like an object taking shape in the afternoon clouds. Like David emerging from a block of granite. At two-thirds through the model you can see the image clearly. It's spatially crisp, refined. Then in the last third of the model something strange but predictable happens: the model translates this well-formed image back into noise. By the output of the U-net, the activations are again completely uninterpretable. Why does this happen? Because we're asking the model to predict the noise rather than the clean image. This is awkward. Like Michelangelo visualizing the chips of granite surrounding the David rather than the David itself. Like asking a friend to look at a cloud and imagine not a shape, but the extra bits of cloud that when removed will leave a shape remaining.

Rudy Gilman

11,641 views • 1 year ago

What you’re seeing Is the truffle learning To express itself Through light Every input (user keystroke, prompt, query) Produces embeddings That are live-diffused To the LED’s The truffle runs a diffusion model In parallel To express itself to the world

What you’re seeing Is the truffle learning To express itself Through light Every input (user keystroke, prompt, query) Produces embeddings That are live-diffused To the LED’s The truffle runs a diffusion model In parallel To express itself to the world

simp 4 satoshi

19,542 views • 9 months ago

Anthropic research lead: "99% of our engineers are running swarms of 300+ self-improving agents. close the agent loop. Give the model a way to verify its own output" in a 20-minute session, Anthropic team member explains how to build a model that improves itself. Claude + loops + plan mode + dynamic workflows -that’s the secret. Watch the talk, then save the playbook below.

Anthropic research lead: "99% of our engineers are running swarms of 300+ self-improving agents. close the agent loop. Give the model a way to verify its own output" in a 20-minute session, Anthropic team member explains how to build a model that improves itself. Claude + loops + plan mode + dynamic workflows -that’s the secret. Watch the talk, then save the playbook below.

Codez

662,787 views • 26 days ago

Literally, the first transcription model that you can steer with a prompt. Send audio + a prompt. The model does what you ask. Universal 3 Pro from AssemblyAI is the world's first production model to support this. 2026 is off to a great start!

Literally, the first transcription model that you can steer with a prompt. Send audio + a prompt. The model does what you ask. Universal 3 Pro from AssemblyAI is the world's first production model to support this. 2026 is off to a great start!

Santiago

10,147 views • 5 months ago

Stop trading manually like a caveman. If you can type, anyone can copy my Polymarket high-frequency bot. I didn’t write a single line of code. I built a top-tier strategy just by “chatting” with AI,one that self-debugs, self-repairs, and even uses “large model decision-making + small model monitoring”. This is a life-changing, dimension-dropping opportunity for ordinary people: Let AI write the script (working > perfect) Let AI review losses (the key to evolution) Let AI fix bugs on its own (24/7 passive income) Machines have no emotions, only execution. Copy this system, and you become the market’s harvester. 🛠️💸

Stop trading manually like a caveman. If you can type, anyone can copy my Polymarket high-frequency bot. I didn’t write a single line of code. I built a top-tier strategy just by “chatting” with AI,one that self-debugs, self-repairs, and even uses “large model decision-making + small model monitoring”. This is a life-changing, dimension-dropping opportunity for ordinary people: Let AI write the script (working > perfect) Let AI review losses (the key to evolution) Let AI fix bugs on its own (24/7 passive income) Machines have no emotions, only execution. Copy this system, and you become the market’s harvester. 🛠️💸

Satori 🎴 💀

73,545 views • 5 months ago

Did you know you can add a single file to your project and dramatically improve Copilot's anwers while *at the same time* reduce the size of the prompt you have to write?

Did you know you can add a single file to your project and dramatically improve Copilot's anwers while at the same time reduce the size of the prompt you have to write?

Visual Studio Code

122,015 views • 1 year ago

An extra layer of clarity of surety for you to validate with your own research never hurt nobody. If you want to get the best experience out of your AI, tell it what you are trying to do and have it develop its own prompt to help you with your problem. After you use that prompt and you get an answer, ask it to critique itself for an even better answer. Especially with your own Astrology and energy!

An extra layer of clarity of surety for you to validate with your own research never hurt nobody. If you want to get the best experience out of your AI, tell it what you are trying to do and have it develop its own prompt to help you with your problem. After you use that prompt and you get an answer, ask it to critique itself for an even better answer. Especially with your own Astrology and energy!

Inspirenaire

32,398 views • 6 months ago

Just added Google Gemini 3.1 Pro to Base44 Gemini 3.1 Pro is really good for gaming and design workflows, and it's fast. If that's what you're building, this is your model. + we're adding something a lot of you were asking for - you can now pick your AI model from the very first prompt. No more starting with a default and switching halfway through. Checkout this Pokemon game I built with the new model, and let me know what you're building

Just added Google Gemini 3.1 Pro to Base44 Gemini 3.1 Pro is really good for gaming and design workflows, and it's fast. If that's what you're building, this is your model. + we're adding something a lot of you were asking for - you can now pick your AI model from the very first prompt. No more starting with a default and switching halfway through. Checkout this Pokemon game I built with the new model, and let me know what you're building

Maor Shlomo

23,662 views • 4 months ago

Dreamina Seedance 2.0 is by far the best video model I've tried, and Dreamina makes it even better. You can try it in the link below. It's not just about the quality of the output, but about how much control you have over your video's look. Watch my video here. I'm attaching reference images and asking the model to generate a video using them. I can reference each image using the @ symbol to instruct the model which image to use. You can even upload a clip and use its camera movement, styles from an image, and audio vibe from a track. By the way, you can take an existing video and replace, remove, or add elements to it while the model preserves everything else. This is the closest we've gotten to "editing videos like photos".

Dreamina Seedance 2.0 is by far the best video model I've tried, and Dreamina makes it even better. You can try it in the link below. It's not just about the quality of the output, but about how much control you have over your video's look. Watch my video here. I'm attaching reference images and asking the model to generate a video using them. I can reference each image using the @ symbol to instruct the model which image to use. You can even upload a clip and use its camera movement, styles from an image, and audio vibe from a track. By the way, you can take an existing video and replace, remove, or add elements to it while the model preserves everything else. This is the closest we've gotten to "editing videos like photos".

Santiago

45,143 views • 3 months ago

🔗Watch the full series for free: 🪢 Whether you’re a rigger or a model, self-tying can be more than you might expect. It’s a chance to slow down, ground yourself, and reconnect with what really matters in rope: your body and the sensations moving through it. 📚 In the free session linked above, dive deeper into the pracitice of self-tying. It can create space for presence, self-expression and introspection, reminding us that self-tying isn’t just a substitute, but a powerful rope experience in its own right.

Sensitive content

🔗Watch the full series for free: 🪢 Whether you’re a rigger or a model, self-tying can be more than you might expect. It’s a chance to slow down, ground yourself, and reconnect with what really matters in rope: your body and the sensations moving through it. 📚 In the free session linked above, dive deeper into the pracitice of self-tying. It can create space for presence, self-expression and introspection, reminding us that self-tying isn’t just a substitute, but a powerful rope experience in its own right.

Shibari Study

46,933 views • 3 months ago

Anthropic engineer: "Over 90% of Anthropic engineers use a self-improving agentic loop for coding. The key is to close the loop: give the agent a way to verify and improve its own work" In this 12-minute talk, an Anthropic engineer explains how to build an agent that continuously improves itself. Claude + routines + loops + CLAUDE.md - that's the foundation. Skip Netflix tonight. Save this post!

Anthropic engineer: "Over 90% of Anthropic engineers use a self-improving agentic loop for coding. The key is to close the loop: give the agent a way to verify and improve its own work" In this 12-minute talk, an Anthropic engineer explains how to build an agent that continuously improves itself. Claude + routines + loops + CLAUDE.md - that's the foundation. Skip Netflix tonight. Save this post!

unicode

61,445 views • 18 days ago

Websites now self-improve! Our AI agent (trained on thousands of tests) comes up with website variations and then launches A/B tests. When tests reach stat sig, we use the data collected to improve your model and repeat the process. Try it:

Websites now self-improve! Our AI agent (trained on thousands of tests) comes up with website variations and then launches A/B tests. When tests reach stat sig, we use the data collected to improve your model and repeat the process. Try it:

Fab

6,793,206 views • 1 year ago