图像生成：从杂点到画面Image Generation: From Noise to a Picture
班级想给春游画海报。小禾输入“绿色山坡上的红色风筝”。机器怎样把文字变成像素？The class wants a poster of a red kite on a green hill. How can text become pixels?
第 1 步：先学习图文模式Step 1: Learn visual and language patterns
模型并不是拿着人类画笔。The model is not holding a human paintbrush.
图像由像素组成，每个像素记录颜色。生成模型从训练资料学习图像的规律。文字条件能引导主题和关系。模型可能组合出新画面，也可能重复训练资料的细节。不能把生成简单理解成搜图拼贴，也不能保证它从不记忆样本。An image consists of pixels storing color. A generative model learns visual patterns from training data. Text can condition its subject and relationships. It may compose new images and can sometimes reproduce training details. Image generation is neither simply image-search collage nor guaranteed freedom from memorization.
第 2 步：逐步去噪是一条路线Step 2: Denoising is one approach
一片杂点怎样逐渐有了形状？How can random-looking dots gain structure?
扩散模型训练时学习处理加噪样本。生成时，从噪声开始逐步更新。文字条件参与引导更新方向。随机种子改变起始噪声，结果也会变化。它不是把某张藏好的原图擦干净。这里演示噪声与目标的混合，只帮助看过程，不运行真正扩散网络。Diffusion models learn from noisy examples. Sampling starts with noise and progressively updates it, guided by conditions such as text. A random seed changes the initial noise and can change the result. This is not simply cleaning a hidden original photo. Our activity mixes noise with a target as a visual analogy; it does not run a diffusion network.
第 3 步：生成后仍要审图Step 3: Review the generated image
画面好看，就能直接当事实图吗？Does an attractive image make a factual diagram?
手指、文字、数量和空间关系都可能出问题。小禾数清风筝线，核对海报文字。历史和科学配图还要查资料。生成画面不能证明某件事真的发生。发布时也要考虑人物授权、来源和使用条件。明确的描述能改善可控性，但不保证每个细节准确。Anatomy, text, counts and spatial relationships can be wrong. He checks the kite strings and poster lettering. Historical or scientific illustrations also need source checks. A generated picture is not evidence that an event occurred. Use also raises questions of permission, provenance and terms. Clear instructions can help without guaranteeing correctness.
🎮 去噪画布实验🎮 Denoising canvas experiment
拖动去噪步数并改变种子，观察形状与细节怎样变化，再完成四个观察任务。Adjust steps and seed, observe the changing picture, then complete four observation tasks.
一句话记住它：Remember: 模型按学到的规律生成画面，输出仍需逐项核对。Models generate from learned patterns; their outputs still need review.
扩散是一种生成方法，并非唯一方法。Diffusion is one method among several.随机种子影响初始噪声。The seed affects initial noise.生成图片不等于事实证据。A generated picture is not factual evidence.
Back to tree
