Imagine asking for a purple car beside an orange house. The generated image contains a car, a house, purple, and orange—but the car is orange and the house is purple. An object-presence check would pass. A color-presence check would also pass. The failure is in the assignment between them.想象我们要求一辆紫色汽车停在橙色房子旁。生成图里汽车、房子、紫色和橙色都出现了,汽车却是橙色,房子成了紫色。只检查对象是否存在会通过,只检查颜色是否存在也会通过;错的是两者之间的归属关系。

A final image cannot tell us when that assignment was made. In CrystalBind, we study the denoising path itself. Starting from the same noise, we let one prompt guide the process and then swap the two attributes at different steps. The resulting image gives a counterfactual test: can later text still redirect which object owns which color?仅看最终图像,无法知道归属何时确定。在 CrystalBind 中,我们研究去噪过程本身:固定初始噪声,先用原提示词去噪,再在不同步骤交换两个属性。最终图像提供一个反事实检验:后续文本还能否改变“哪种颜色属于哪个对象”?

Separate reading from changing区分读到变化与改变结果

A failed late switch does not identify a single cause. The denoiser might not respond to the new condition. It might respond, but the sampler may write too little of that difference into the latent. Or the change may be written and then fade before the final image. We analyze these as read, write, and propagate: three different links in the path from text to ownership.一次较晚的提示词切换没有成功,并不能说明只有一个原因:去噪器可能没有响应新条件;也可能响应了,但采样器写入 latent 的变化太小;还可能写入了变化,却未能保留到最终图像。我们把它拆成读取、写入、传播三个环节,分别检查文本如何影响属性归属。

This distinction keeps the method honest. Stronger attention is not automatically stronger control, and a model noticing a new description does not prove that the final image can still be changed.这样拆解,才能避免把“注意力更强”直接等同于“控制更有效”,也避免把模型能辨别新描述,误当成最终图像仍可改写的证据。

Guide the assignment, not just the object引导属性归属,而不只是对象出现

CrystalBind turns that diagnosis into an inference-time design. It focuses updates in an early window, refreshes noun regions from the model's current cross-attention, and uses separate terms for object presence and directional attribute ownership. The region is a live target because the scene layout may still be moving; the ownership term asks not only for the right pairing, but also against the competing pairing.CrystalBind 将诊断转化为推理时设计:把更新集中在较早的窗口,从当前 cross-attention 动态刷新名词区域,并分别优化对象出现与属性归属。区域目标需要持续更新,因为场景布局仍可能变化;归属目标不仅要增强正确配对,还要抑制竞争配对。

No new generator weights are needed for this approach. The larger lesson is methodological: before increasing the strength of a guidance loss, find out which decision it is trying to change, when that decision remains reversible, and whether the intervention survives to the output.这一方法不需要训练新的生成模型权重。更一般的方法论是:在增强引导损失之前,先弄清它想改变什么决策、这个决策何时仍可逆,以及干预能否影响最终输出。