This is a node in ComfyUI that merges two SD1.5 models at a fixed ratio. It takes the MODEL and CLIP inputs of the two models and outputs the merged result. The contents can be seen in init.py, but it primarily replaces the higher-level IN layers with those of Model 2, based on the contents of Model 1.Please place it as follows:<pre><code>ComfyUI\custom_nodes\sd15_fixed_merge\__init__.py</code></pre>and restart ComfyUI.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b521e1cf-6971-403b-b0fe-3f414cbcc3a4/original=true/b521e1cf-6971-403b-b0fe-3f414cbcc3a4.jpeg" />Q. What are the effects?A. It depends on the models used, but mainly affects prompt understanding, composition, and pose. It is particularly effective for images with action or NSFW content.Q. What models are recommended?A. Model 1 can be anything. Model 2 is assumed to be one of my models such as <a target="_blank" rel="ugc" href="https://civarchive.com/models/1967523/neosd">NeoSD</a>. However, there are no particular restrictions. A model that aims to produce a wide range of styles is preferable.Q. Will it affect the style of the image?A. It will have a slight effect. Especially with photographic models, using an anime-style model for Model 2 can sometimes result in a childish appearance. Please use a realistic model. However, in many cases, anime-style models have a greater effect.Simply put, many SD1.5 models are specialized for displaying a single standing girl and hardly follow prompts. SD1.5's ControlNet works well, allowing for fine-grained control through screen splitting, so some might think this isn't much of a problem. It's the environment where you can create the most detailed models if you put in the effort. However, I suspect very few people use it to that extent. This tendency is generally present in other SD1.5 models as well, to varying degrees, so most people probably think that as long as a picture of a girl appears, it's fine.That's a waste, so this node was created to make it easier to use with prompts while retaining the rendering tendencies of many models. Because it's a simple ratio-based layer merge, it's less likely to break down.Of course, NeoSD itself, which is the key to this, is incomplete, so the output isn't perfect. In some cases, it might be better not to use it (you can bypass it with Ctrl-B). Nevertheless, having more options isn't necessarily a bad thing.

Files with SHA256: 98cc01dee12cb803bc519f198398151fb423289aef54c27084ab09a389e7e6cb