Krea2 EchoRC 是一个面向 Krea2 / FLUX 架构模型 的 ComfyUI 推理增强节点。
Krea2 EchoRC is a ComfyUI inference enhancement node designed for Krea2 and FLUX-based diffusion models.
EchoRC 不修改模型权重，而是在 Krea2 采样过程中加入一个动态记忆系统，让模型在生成过程中保留部分历史状态信息，从而增强不同采样阶段之间的信息传递。
主要用于提升：
复杂提示词理解
场景一致性
多物体关系
细节稳定性
EchoRC does not modify model weights. Instead, it introduces a lightweight dynamic memory system during Krea2 sampling, allowing information from previous denoising stages to influence later generation steps.
Designed to improve:
Complex prompt consistency
Scene coherence
Object relationships
Detail stability
参数说明
Reservoir Size
控制动态记忆容量。数值越大，保存的信息越多，但计算量增加。
Rho (Spectral Radius)
控制记忆状态的稳定性和持续时间。越高，历史信息保留越久。
Leak Rate
控制新信息与旧状态的混合比例。越高，记忆衰减越慢。
Input Scale
控制输入信息进入 EchoRC 的强度。
Strength
控制 EchoRC 对最终生成结果的影响程度。
Start Block / End Block
控制 EchoRC 作用于 Krea2 Transformer 的范围。
Seed
控制 EchoRC 动态状态初始化，使结果具有可复现性。
Parameters
Reservoir Size
Controls the memory capacity of the reservoir. Higher values store more information but increase computation.
Rho (Spectral Radius)
Controls memory stability and retention length.
Leak Rate
Controls the balance between previous memory states and new information.
Input Scale
Controls the strength of input signals entering the reservoir.
Strength
Controls the overall influence of EchoRC on generation.
Start Block / End Block
Defines which Krea2 Transformer blocks are affected.
Seed
Controls reservoir initialization for reproducible results.
Description
如果有问题欢迎私聊，或者发送邮件到[email protected]。
本节点包由Vibe coding生成。
Comments (4)
That's a very interesting node! Thanks for that. Is it worth experimenting with the values, or are the default values already the best? If it makes sense to experiment, which setting should you change first? Does it matter whether the seed is different from or the same as the KSampler's seed?
Haha, first of all, thank you for your love. However, version v1 did not achieve the effect I wanted at all. It only achieved about 0.1% of my goal. I released version v2. You can use the default parameters to set it. If you need more freedom, you can refer to the parameter introduction in this section. Using different parameters can sometimes produce many unexpected results!The execution order of nodes and kmesamplifiers is such that nodes are injected into each step of kmesamplifiers. Therefore, adjusting the seed of each will produce different effects. In short, we welcome your suggestions and ideas!
哥，我眼拙，实在找不到什么 很明显的进步点...您举例那些完全可以理解为抽卡不同种子的偶发性
虽然按理来说在同一个seed里面产生的图片应该是一样的，第一版产生的效果是因为节点永远仅在第一步注入Reservoir，还有其他的问题，不过要谢谢你的评论让我回头看了一下代码并且发现了这些bug，我处理了这些bug，现在本节点会带来较为显著的处理，您可以查看参考图，他们都是在同一个seed下产生的，并且产生了很多我认为好的优化，欢迎体验v2版本的节点，并且把你的感受评论下来