CFG
CFG (Classifier-Free Guidance) is a guidance mechanism used in diffusion models. It computes both conditional and unconditional noise predictions simultaneously and amplifies generation guidance in a specified direction, controlling the degree to which the generated image matches the input prompt.
Core Facts
The implementation of negative prompting techniques in text-to-image scenarios relies on the Classifier-Free Guidance (CFG) mechanism.
A higher CFG Scale causes the generated result to adhere more strictly to the text description, but too high a value may lead to oversaturated colors and distorted details.
Timeline (last 90 days)
The SFG branch only needs to run during the first 20% of image sampling steps, and without using CFG its overhead is only approximately 1.2× the base NFE.
Classifier-free guidance (CFG) runs both a conditional branch and an unconditional branch simultaneously, amplifying the conditional signal by subtracting the unconditional direction from the conditional direction, trading a doubling of inference cost for quality improvement.
Turbo defaults to CFG=1, requiring only one forward inference pass per step.
CFG (Classifier-Free Guidance) requires running both conditional and unconditional forward passes simultaneously at each denoising step, doubling the computation per inference step when enabled.
On Qwen-Image 2.1, a lower CFG value (3) generally produces more natural output with less of an overfitting feel.
Supra2-IMG uses CFG 3.0 as the default display value, which is lower than the 7.0–9.0 commonly used by Stable Diffusion.
Dithering textures and dotted textures in AI image generation typically stem from insufficient sampling steps during the diffusion model denoising process, improper CFG guidance strength, or training quality issues.
CFG is a core hyperparameter in diffusion models, rooted in a 2022 research finding, and works by enhancing responsiveness to text descriptions through a weighted mixture of the predictions from the conditional and unconditional denoising paths.
The tester found that raising CFG from 1 to 4–4.5 yielded some improvement in output image quality.
CFG=1 means no interpolation amplification is applied at all, causing the generation process to degrade into pure unconditional diffusion, rendering the prompt ineffective as a guide.
14 more timeline events
All Facts (20)
The implementation of negative prompting techniques in text-to-image scenarios relies on the Classifier-Free Guidance (CFG) mechanism.
75%VerifiedA higher CFG Scale causes the generated result to adhere more strictly to the text description, but too high a value may lead to oversaturated colors and distorted details.
65%UnverifiedCFG值（Classifier-Free Guidance Scale）通常设置在5-12之间，采样步数一般在20-50步之间取得较好平衡
85%Unverified新一代模型如FLUX、Z-Image由于架构改进，CFG值设为1甚至更低即可获得最佳效果，而传统Stable Diffusion模型通常使用7-12的CFG值
80%UnverifiedFSM constraints can guarantee JSON format validity at the grammatical level, but cannot constrain the semantic quality of fields.
70%UnverifiedEven with the same prompt, different seeds produce completely different images; once the seed is fixed, even small changes to the number of sampling steps or the CFG guidance strength can cause differences in the output.
60%UnverifiedThe CFG value (Classifier-Free Guidance Scale) controls the degree to which the model adheres to the prompt; a higher value yields greater fidelity but may result in oversaturation.
60%UnverifiedClassifier-free guidance (CFG) runs both a conditional branch and an unconditional branch simultaneously, amplifying the conditional signal by subtracting the unconditional direction from the conditional direction, trading a doubling of inference cost for quality improvement.
50%UnverifiedThe SFG branch only needs to run during the first 20% of image sampling steps, and without using CFG its overhead is only approximately 1.2× the base NFE.
50%UnverifiedTurbo defaults to CFG=1, requiring only one forward inference pass per step.
50%UnverifiedCFG (Classifier-Free Guidance) requires running both conditional and unconditional forward passes simultaneously at each denoising step, doubling the computation per inference step when enabled.
50%UnverifiedOn Qwen-Image 2.1, a lower CFG value (3) generally produces more natural output with less of an overfitting feel.
50%UnverifiedSupra2-IMG uses CFG 3.0 as the default display value, which is lower than the 7.0–9.0 commonly used by Stable Diffusion.
50%UnverifiedDithering textures and dotted textures in AI image generation typically stem from insufficient sampling steps during the diffusion model denoising process, improper CFG guidance strength, or training quality issues.
50%UnverifiedThe tester found that raising CFG from 1 to 4–4.5 yielded some improvement in output image quality.
50%UnverifiedCFG is a core hyperparameter in diffusion models, rooted in a 2022 research finding, and works by enhancing responsiveness to text descriptions through a weighted mixture of the predictions from the conditional and unconditional denoising paths.
50%UnverifiedCFG=1 means no interpolation amplification is applied at all, causing the generation process to degrade into pure unconditional diffusion, rendering the prompt ineffective as a guide.
50%UnverifiedCFG=4–4.5 is the recommended starting range for many Flux and SD3 series models, while SDXL-type models commonly use 5–8.
50%UnverifiedExcessively high CFG values (e.g., above 10) can cause color saturation to blow out, over-sharpening of details, and the appearance of artifacts.
50%UnverifiedFor Qwen Image 2.1, it is recommended to start around CFG 4, paired with approximately 40 sampling steps.
50%