CFG Scale — What Is It, What Does It Do, How To Use It?

AI-generated art isn’t just about prompts and style — it’s about control. Every time you enter a prompt in Midjourney, Stable Diffusion, or Leonardo, the model has to decide how closely to follow your words versus how much to rely on its own “creative instincts.” This decision-making balance is governed by a single, powerful variable: the Classifier-Free Guidance Scale (CFG Scale).

Think of it as the invisible hand steering your AI’s imagination. The CFG Scale determines whether your AI acts like a disciplined illustrator who follows every detail of your prompt — or like a visionary artist who interprets freely. Understanding how this parameter works is essential if you want consistent, high-quality, and visually intentional results.

In this article, we will discuss in detail what the CFG Scale is.

to content ↑

What is the Classifier-Free Guidance (CFG) Scale, and How Does it Control AI Generation?

AI image generation might look effortless — type a prompt, get an image — but under the hood, it’s a negotiation between your intent and the model’s learned distribution of the world. The CFG Scale is the numeric control that manages that negotiation. It tells the model how “obedient” to be toward your prompt.

A low CFG value gives the AI freedom to imagine, creating results that are often more abstract or painterly. A high CFG value forces it to stick closely to your instructions, producing highly literal and detailed outputs. Too low, and your prompt gets ignored; too high, and your image looks overprocessed or unnatural.

In short — CFG isn’t just about style preference. It’s a mathematical control over how the AI interprets and translates your words into pixels.

to content ↑

Why is a Technical Introduction to CFG Necessary?

Understanding CFG isn’t just for engineers. For AI artists, it’s the key to reproducible results — to knowing why one prompt looks cinematic while another looks cartoonish, even when using the same model.

AI art tools have evolved rapidly, but their underlying behavior is rooted in probability and controlled randomness. CFG sits at the center of that system. Without it, prompt-based creation would feel inconsistent and unpredictable.

How Did the AI Art Revolution Create a Control Paradox?

As AI art tools became more creative, they also became harder to control. Diffusion models like Stable Diffusion introduced near-infinite possibility — but also randomness. Users wanted both freedom and fidelity: the ability to generate new ideas without losing command over tone, color, or subject.

Key points to understand this paradox:

• Generative explosion: Models like Stable Diffusion and DALL·E can produce thousands of variations from a single idea.
• Loss of consistency: Without guidance, the model interprets prompts loosely — the same text can yield wildly different visuals.
• Creative tension: Artists want both novelty (AI freedom) and control (faithful adherence to their vision).
• Need for balance: This conflict between chaos and precision gave rise to the guidance problem — how to steer creativity without limiting it.

How do Generative Models create an inherent need for control?

Diffusion models work by learning the statistical structure of images — the patterns that make a face, sky, or shadow look real. But that same statistical freedom means that every output is a probability, not a guarantee. This makes AI both fascinating and frustrating: it can surprise you, but it can also ignore you.

Why this creates control issues:

• Probabilistic output: Every generation is a sample from a distribution, not a fixed result.
• Interpretive looseness: The AI recognizes what a “cat” might look like, not your specific cat.
• Prompt ambiguity: The model doesn’t reason about your intent — it interprets token probabilities, which can diverge from your mental image.
• Result: The more creative freedom the AI has, the more control you lose — a fundamental trade-off in generative design.

What is the CFG Scale’s primary function in this paradox?

The CFG Scale was designed as the answer to this creative tension. It gives users a way to bias the generative process toward their prompt — strengthening the link between human intention and machine imagination. The higher the scale, the stronger the bias toward your input prompt.

In simple terms:

• Acts as a steering wheel: CFG is the user’s control lever for prompt strength.
• Balances two forces:
  • Low CFG → Model creativity dominates (more abstract, unexpected).
  • High CFG → Prompt direction dominates (more faithful, literal).
• Core purpose: Reconnect human intention with machine imagination — ensuring that AI art feels guided, not random.

to content ↑

What is Classifier-Free Guidance (CFG) at its Conceptual Core?

Before you can tune CFG effectively, you need to understand what’s happening under the hood. Classifier-Free Guidance isn’t a “style” setting — it’s a mathematical process that modifies how diffusion models remove noise during image generation.

In essence, the model generates two versions of the same image — one guided by your prompt and one unguided — and then combines them in a way that balances realism and instruction. The scale you set determines how far it pushes that balance.

How Does the Underlying Diffusion Process Work?

Every image generation process in diffusion models happens through two main stages: forward diffusion (adding noise) and reverse diffusion (removing it). Understanding these helps explain why CFG even exists.

What is Forward Diffusion (Noise Injection)?

During training, the model learns by gradually corrupting real images with noise until they turn into pure static. This process helps it understand what randomness looks like at every stage, teaching it how to rebuild images step by step.

Core idea:

• The model starts with a clean image (x₀).
• It adds tiny amounts of random Gaussian noise over many time steps (t₁, t₂, …, tₙ).
• Eventually, the image becomes indistinguishable from white noise.

Why this matters:

• Each step teaches the model a map between “slightly noisy” and “cleaner” images.
• This is how it learns to understand structure, edges, and color relationships statistically.

Example analogy:

• Imagine you’re teaching an artist by showing them a photo and then smudging it little by little.
• By the end, it’s pure blur — but if they remember every smudge step, they’ll know exactly how to reverse it later.

Forward diffusion is destruction with purpose — the model learns by watching order decay into chaos.

What is Reverse Diffusion (Denoising)?

In generation mode, the process runs backward — starting with random noise and slowly removing it. At each step, the AI predicts what the cleaner version of the image should look like, based on the given prompt. This is the “creative” phase of the model’s work.

Core process:

• The AI begins with random noise (xₙ).
• At each step, it predicts the noise that should be removed, using the U-Net network.
• Guided by your prompt (conditional input), it keeps denoising until the final image emerges (x₀).

Why it’s crucial:

• Reverse diffusion is what enables creativity — the AI uses statistical memory to “hallucinate” structure, texture, and light from randomness.
• The closer its denoising predictions align with your prompt, the more accurate the result.

Example analogy:

• Imagine sculpting a statue from a block of marble, but instead of chiseling, you’re “subtracting noise” until the shape appears.
• Each denoising step refines the detail — from broad composition to fine texture.

Reverse diffusion is creation through refinement — turning structured randomness into meaningful imagery through guided prediction.

What is the Role of the U-Net in the Denoising Process?

The U-Net is the neural architecture that performs the denoising. It’s responsible for predicting the exact amount of noise to remove at each timestep. CFG modifies how much weight is given to the “prompted” version of this prediction — effectively steering the U-Net’s output in your chosen direction.

How it works:

• The U-Net takes a partially noisy image as input (xₜ).
• It encodes visual information through multiple convolutional layers — understanding structure, texture, and depth.
• It then decodes that information back to predict the noise component (ε̂), essentially saying, “Here’s what should be removed next.”
• This process repeats across hundreds of steps until the noise disappears and the final image emerges.

CFG’s influence:

• Classifier-Free Guidance (CFG) alters how much the U-Net listens to your prompt.
• At low CFG, it mostly trusts its learned training patterns (freedom, creativity).
• At high CFG, it leans more heavily on the prompt — forcing the U-Net to emphasize exactly what you asked for.
• Essentially, CFG is the steering wheel that controls the U-Net’s decision-making strength toward your intent.

Example analogy:

• Imagine the U-Net as a painter restoring an old photo covered in dust.
• Each pass removes a layer of dirt — but depending on your direction (“make it more cinematic” vs. “keep it realistic”), CFG adjusts how boldly the painter cleans and enhances the image.
• Too cautious, and the picture stays dull. Too aggressive, and details get lost.

The U-Net is the mechanical artist that reconstructs your image from noise. CFG determines how assertively it follows your creative command versus its own learned instincts.

Where Did the Concept of ‘Guidance’ Originate in AI?

The term “guidance” predates diffusion models. Early generative systems needed ways to force randomness toward desired results.

What was Classifier Guidance (CG), the predecessor technique?

Before CFG, guidance was external — an additional model classified whether an image looked “right” (like a dog, a face, etc.) and nudged the generator accordingly. It worked but was slow and prone to overfitting.

Here’s how it worked:

• The generator created an image based on noise and random variation.
• A classifier model analyzed the image to see if it matched the desired concept (e.g., “dog,” “face,” or “car”).
• The classifier then pushed the generator closer to that target concept by applying feedback gradients.

However, CG had some critical downsides:

• It required an additional external model, which increased computational cost.
• It was prone to overfitting, meaning it forced the generator to stick too rigidly to the category, often losing creativity or diversity.
• It made generation slower and less stable because of the constant cross-model communication.

How did Classifier-Free Guidance (CFG) achieve a major breakthrough?

CFG eliminated the need for that external classifier by training the same model to predict both “guided” and “unguided” outputs internally. This made generation faster, more stable, and easier to fine-tune. The model could now steer itself — no external supervision required.

Here’s what changed:

• The model learns in two modes:
  • Conditional mode: with a prompt (e.g., “a cyberpunk city at night”).
  • Unconditional mode: without any prompt (just pure imagination).
• During generation, the two outputs are compared to calculate a guidance vector — showing how much to “push” the image toward the prompt.
• This makes generation faster, more efficient, and more controllable, balancing between prompt adherence and creative freedom.

CFG removed the need for external judgment — making the diffusion model self-steering, efficient, and capable of producing more natural, nuanced images.

to content ↑

What is the Mathematical Core Driving the CFG Scale?

To truly understand CFG, we have to look at its equation. The mathematics might seem intimidating, but at its heart, it’s about calculating how much to nudge the model from randomness toward your desired output.

What are the Two Essential Predictions Required by CFG?

Every time the diffusion model denoises an image (removes noise step by step), it doesn’t make just one guess — it actually makes two distinct predictions:

• One with your prompt, representing what the model thinks you want.
• One without your prompt, representing what it thinks looks statistically natural.

By comparing these two, the model finds the “direction” in latent space that moves an image closer to your intent. This dual-prediction method is the foundation of Classifier-Free Guidance (CFG).

What is the Conditional Noise Estimate (ϵθ​(xt​,c))?

This is the guided prediction — made when your text prompt is included. Think of it as the model’s “best guess” at what kind of noise to remove to make the image align with your instructions.

• Meaning: The model listens directly to your prompt (“a cat in neon Tokyo”) and adjusts the denoising process accordingly.
• Purpose: It enforces your creative intent — ensuring the image reflects your described scene, style, or subject.
• Mathematically: This is expressed as ϵθ​(xₜ, c), where c = conditioning (your prompt).

💡 Example:

If you prompt “a portrait of a woman in golden light,” this estimate ensures the noise is removed in a way that highlights warm lighting and facial structure consistent with your prompt.

What is the Unconditional Noise Estimate (ϵθ​(xt​,∅))?

This is the unguided prediction — made without any prompt. Here, the model generates what it believes is a statistically plausible image from its learned data, purely based on internal patterns.

• Meaning: The AI produces a baseline guess, imagining what a “neutral” or “average” image might look like at that step.
• Purpose: It provides a reference point — allowing CFG to calculate the difference between what’s random and what’s guided by your words.
• Mathematically: This is expressed as ϵθ​(xₜ, ∅), where ∅ = no conditioning.

💡 Example:

Without any text, the model might denoise toward a generic landscape, portrait, or texture — whatever is statistically common in its training data.

How Do We Deconstruct the Core CFG Scaling Formula?

At its core, Classifier-Free Guidance (CFG) works like a balance system — it combines two competing forces:

1. The model’s creative intuition (its unguided prediction), and
2. The user’s intention (your prompted direction).

The model compares both, then merges them mathematically to form the final denoising prediction.

In simpler terms, CFG acts like a volume knob for how strongly the AI listens to your words versus how much freedom it gives itself.

This process happens at every denoising step in diffusion — meaning each “frame” of generation gets a tiny push toward (or away from) your prompt depending on your CFG Scale value.

When you set the CFG Scale higher, you’re amplifying that bias — the model moves further in the direction of your prompt. When you lower it, you allow more randomness and creative drift.

What is the Guidance Vector (G)?

The Guidance Vector (G) is the mathematical “difference” between what the AI wants to create on its own and what you are asking it to create.

It’s the directional signal that tells the AI how far, and in what direction, to move in its latent space to align with your text.

Mathematically:

G = \epsilon_{\theta}(x_t, c) - \epsilon_{\theta}(x_t, \emptyset)

Or, in simpler language:

“Take what I would generate naturally, subtract it from what I’d generate when given a prompt, and move in that direction.”

In plain terms:

• Purpose: It measures how your prompt changes the model’s perception of the image.
• Direction: It acts like a compass, pointing toward the version of the image that fulfills your prompt’s intent.
• Mechanics: The greater the difference between the guided and unguided predictions, the stronger the “push” in that direction.
• Effect: When visualized, this vector reshapes the generated image — altering structure, color, and detail in real time as it denoises.

💡 Example:

If your unguided prediction leans toward “a neutral landscape” but your prompt is “a city skyline at sunset,” the guidance vector tells the model exactly how to shift — adding buildings, warm light, and horizon lines while removing irrelevant elements.

How is the Final Denoising Prediction (ϵCFG​) calculated?

Before the model renders your final image, it must decide how much of your prompt’s influence should be applied at every denoising step. This is where the CFG Scale takes center stage — it acts as the mathematical dial between creativity and control, blending free generation with user intent.

ϵ_{CFG}(x_t, c) = ϵ_{\theta}(x_t, ∅) + w \cdot G

In simple terms:

• The AI begins with the unguided prediction — its natural sense of what looks visually plausible.
• It then adds the guidance vector (G), the directional difference between what it imagines freely and what your prompt demands.
• The multiplier w (the CFG Scale) controls how heavily your prompt influences the output.

Think of this like a creative mixing board: the unguided prediction is the background rhythm, while w × G amplifies your prompt’s melody. A balanced mix produces harmony — too low, and your prompt fades; too high, and the output becomes distorted or overly literal.

This formula means: start with the unguided estimate, then add a scaled version of the guidance vector. The variable w (your CFG Scale) controls how far the AI pushes toward your prompt.

What does the Scale Factor (w) represent mathematically?

The scale factor (w) is like a volume knob for your prompt’s influence. Mathematically, it multiplies the guidance vector (G) — the difference between what the AI would create on its own and what you’ve asked it to make.

• When w is small (e.g., 1–6), the model follows your prompt lightly, leaving more room for its own creative interpretation.
• When w is large (e.g., 12–20), your prompt dominates, forcing the model to adhere strictly — but this can also push it beyond what it was trained for, leading to over-saturated, noisy, or unnatural results.

In simple terms: w controls how obedient or imaginative the AI becomes. Too low, and it wanders; too high, and it overcorrects. The magic lies in finding the balance where both human intent and machine creativity coexist smoothly.

to content ↑

How Does the CFG Scale Practically Impact the Creator’s Final Image?

For most artists, the CFG Scale is the invisible lever that determines how “human” or “machine” their art feels. It directly affects the realism, diversity, and emotional balance of an AI-generated image. In practice, every small change in CFG alters how much freedom the model gives itself — whether it interprets your words like a creative collaborator or follows them like strict instructions.

The right CFG setting can make the difference between organic, cinematic results and rigid, artificial compositions. Understanding where your project sits on this spectrum is essential for professional consistency.

Where Does CFG Sit on the “Adherence vs. Diversity” Spectrum?

At its core, CFG defines a continuum between two creative extremes: adherence (accuracy to your text prompt) and diversity (creative deviation). The scale you choose decides which side the model prioritizes.

What are the results of Low CFG (1–6) on creativity and prompt deviation?

Low CFG values (1–6) give the model a lot of freedom. The AI explores visual space with minimal constraints, resulting in dreamlike or abstract compositions that may loosely relate to your prompt.

These settings are perfect for concept exploration, moodboarding, or generating painterly effects — but they’re risky when you need the image to follow a specific brief.

For example, a prompt like “a red-haired woman in a misty forest” might produce variations in color or setting under low CFG — maybe a blue-haired woman, or a glowing background that ignores your “misty” detail.

Why is Medium CFG (7–12) considered the “sweet spot”?

The medium range (7–12) balances structure and creativity — it’s where most professional artists and AI photographers operate.

Here, the model still listens to your prompt closely but preserves a touch of artistic interpretation. Images tend to feel coherent, detailed, and natural, with strong composition and controlled lighting.

Most diffusion models (like Stable Diffusion XL or Leonardo AI) default within this range because it consistently produces visually believable results that align with the input prompt.

What are the consequences of High CFG (13+) on precision and artifacts?

Beyond 13, the model starts to overcommit to your prompt — like an overzealous assistant trying too hard to please. This can create mechanical textures, color clipping, or over-sharpened edges.

High CFG is useful for enforcing clarity in complex prompts (“a futuristic city reflected in chrome water”), but it often sacrifices aesthetic softness. Instead of cinematic depth, you might see flattened contrast, excessive glow, or visual noise.

What causes Image Quality Degradation at high values?

Mathematically, high CFG extrapolates the guidance vector too far from the model’s training distribution.

In simple terms: the model starts generating patterns it was never trained to produce, causing color oversaturation, unnatural contrast, and a plastic-like texture.

Why is there Reduced Generalization with high guidance?

When CFG is too high, the model stops “thinking creatively.” It fixates on the literal elements of your prompt and ignores the subtleties that make images feel organic — background cohesion, atmospheric lighting, and tonal balance.

This is why overly high CFG outputs often look too perfect, lacking the imperfections that make realism believable.

How Does CFG Influence Image Frequencies and Structure in Latent Space?

Every image can be analyzed in terms of its frequency content — the large-scale structure (low frequencies) and the fine details (high frequencies). CFG directly affects how these frequencies are emphasized or distorted.

What is the Impact on Low Frequencies (Structure)?

At lower CFG values, the AI explores broad structures more freely — composition, layout, perspective — producing varied and open-ended results.

When you increase CFG, it locks down the major shapes and proportions of your subject, ensuring the model follows your spatial instructions precisely.

However, overly strong low-frequency guidance can lead to rigid framing, where everything looks centered or unnaturally aligned.

What is the Impact on High Frequencies (Detail)?

High CFG also amplifies detail — edges become sharper, surfaces more reflective, and textures more defined.

Yet beyond a certain point, it starts “hallucinating” detail — creating noisy grain or plastic surfaces that look hyperrealistic but uncanny.

That’s why many artists balance CFG with style-related negative prompts like “no HDR lighting,” “no plastic skin,” or “avoid oversharpened details.”

How is Prompt Complexity a Factor in Choosing the Scale?

The more complex your prompt, the higher CFG it typically needs.

For instance, “a knight riding a dragon under aurora lights” contains multiple compositional and thematic cues. Without enough CFG, the model might skip key elements — the dragon might be missing, or the aurora might be faint.

Conversely, a simple prompt like “portrait of a woman” doesn’t need high CFG because it’s already well-represented in training data. Excessive CFG here would reduce realism or introduce noise.

What Advanced Techniques Involve CFG Control?

As AI artistry evolves, creators are learning that CFG isn’t a static slider — it can be orchestrated dynamically across the generation process. Advanced workflows now combine CFG scaling with time-based modulation, negative guidance, and adaptive prompting for deeper creative control.

What is CFG Scheduling?

CFG Scheduling adjusts the guidance scale during the denoising process — starting high to establish structure, then lowering it to preserve natural detail.

For example: early steps (high CFG) define composition and subject clarity, while later steps (low CFG) restore soft lighting and texture.

This technique mimics how cinematographers block a scene before lighting it — a structured-to-creative progression that yields balanced realism.

How does the Role of Negative Prompts relate to CFG?

Negative prompts are the “counterweight” to CFG. While CFG pushes the model toward your desired concept, negative prompts pull it away from unwanted elements.

They work under the same guidance principle but in reverse — applying an anti-guidance vector that steers the output away from certain traits (like “no extra hands,” “no harsh light,” or “no blur”).

Professionals often fine-tune both together: high CFG for clarity, and precise negative prompts for refinement.

to content ↑

What Are the Current Limitations and Future Directions of Guidance Research?

CFG is powerful — but not perfect. As researchers push diffusion models further, they’ve discovered that extreme CFG scaling leads to instability and visual degradation.

This has led to a new wave of research into adaptive and context-aware guidance systems.

Why Does High CFG Cause Oversaturation and Visual Artifacts?

When CFG pushes too far beyond the model’s statistical boundaries, it amplifies patterns that aren’t part of its natural “knowledge.”

Mathematically, this causes the model to sample from regions in latent space that weren’t well-trained — producing blown highlights, color banding, or distorted geometry.

What is the Mathematical Root Cause of Artifacts?

Artifacts emerge because the extrapolation term (w·G) becomes dominant. As w increases, the denoising prediction deviates more from the unconditional baseline, amplifying high-frequency error.

Simply put, the AI starts imagining detail it can’t actually support, like an artist adding brushstrokes to a scene they can no longer visualize clearly.

What Research Solutions (CFG++, APG, etc.) are correcting high-CFG issues?

Recent innovations like Adaptive Projected Guidance (APG) and CFG++ aim to preserve the benefits of strong guidance while minimizing artifacts.

APG, for instance, dampens the component of the guidance vector that runs parallel to the model’s noise distribution — effectively softening the overshoot. The result is cleaner contrast and truer color fidelity even at higher scales.

Where Else Are Guidance Concepts Being Applied?

The philosophy behind CFG — balancing user intent with model freedom — is spreading beyond image generation.

How is CFG being adapted for Text Generation (LLMs)?

Large Language Models (like ChatGPT or Gemini) now integrate similar forms of stylistic guidance, adjusting tone and adherence to user input dynamically. The same concept of “conditional vs. unconditional” sampling applies to text-based creativity.

Why is there a Need for Adaptive Control in Future Models?

Future AI systems will likely include content-aware CFG, automatically tuning the scale based on scene complexity, timestep, or even emotion. Instead of manually setting CFG, artists might describe their intent (“balanced realism” or “expressive freedom”), and the system will self-adjust its guidance strength accordingly.

to content ↑

What is the Final Synthesis of CFG Knowledge?

At its core, CFG is the language bridge between human intention and machine imagination. It converts text meaning into visual discipline — ensuring the AI doesn’t just generate images, but your images.

What is the CFG Scale’s essential function?

The CFG Scale amplifies the difference between guided and unguided predictions, letting users control how strongly their prompt directs the output. It’s the balance knob for creativity versus accuracy.

Why is mastering the CFG Scale crucial for professional AI artists?

Mastering CFG is mastering visual control. It turns AI from a random generator into a responsive creative instrument — allowing you to design consistent, cinematic, and intentional art that aligns perfectly with your vision.

to content ↑

Mastering the Art and Science of CFG

Classifier-Free Guidance (CFG) is more than a technical parameter — it’s the invisible hand that shapes the dialogue between human intent and machine creativity. Understanding it allows you to move from “trying prompts” to directing a model like an artist. The CFG scale governs how obedient or imaginative your AI becomes, defining whether an image looks cinematic, abstract, or artificially constrained.

For creators, it’s the difference between accidental beauty and intentional design. Once you understand how CFG interacts with the model’s internal guidance — from noise prediction to latent structure — you gain precision that rivals traditional digital artistry.

Ultimately, mastering CFG isn’t just about achieving the “best image.” It’s about controlling how the image emerges — balancing fidelity, emotion, and realism in a way that makes AI art truly human-directed.

to content ↑

FAQs About the CFG Scale in AI Generation

1. What is the purpose of the CFG Scale in AI image generation?
The CFG Scale (Classifier-Free Guidance Scale) controls how closely an AI model follows your text prompt versus its own learned creativity. A higher CFG makes outputs more faithful to the prompt but may introduce artifacts; a lower CFG gives more diversity and natural variation.

2. What’s the best CFG value to use in Stable Diffusion or similar models?
Most creators find the “sweet spot” between 7 and 12. This range keeps prompts clear and detailed without causing harsh contrast or over-saturated colors. Lower than 5 often feels random, while higher than 13 risks unnatural, noisy textures.

3. Why do high CFG values cause oversaturation or visual noise?
When CFG is set too high, the model exaggerates its response to your prompt, stretching the guidance vector beyond its trained distribution. This overcorrection results in blown-out highlights, plastic skin tones, or dense artifacts.

4. How is CFG different from using negative prompts?
CFG controls how strongly the AI listens to your overall instruction, while negative prompts specify what to avoid. In practice, they work together — high CFG enforces direction, and negative prompts clean up unwanted outcomes like “no distortion” or “no orange tint.”

5. Does CFG affect the structure of the image or just colors and details?
Both. At low levels, CFG mainly affects composition and large forms (structure). At higher values, it begins to push texture, lighting, and tone, sometimes exaggerating details or distorting fine patterns.

6. Can I change CFG during generation for better results?
Yes — this technique is known as CFG Scheduling. Some advanced users vary CFG dynamically: higher at early steps to build structure, then lower later for finer textures and realism. It’s an emerging method for balancing detail and artistic fidelity.

Sources:

Arxiv

Open Review

Apple

AI Summer