This is a suite of model development tools for ComfyUI, designed for use with MOSS-Image.

In this release, we are providing three tools: TestSampler, MOSS-Image Merger, and the MOSS_BF16 Converter.

TestSampler

( sampler/debug/TestSampler )

TestSampler is a sampler intended for testing purposes; it functions much like a simplified version of the standard ComfyUI KSampler. I had originally conceived of this back in the SD1.5 era, but I abandoned the project due to the frustrating limitations of AI models like ChatGPT and Claude at the time. Given how much more capable these AI tools have become recently, I decided to have "Dots" (an AI assistant) help build it.

• Currently limited to MOSS-Image.

• Fixed to "Euler" and "Simple" sampling methods.

• Supports BF16 format only.

• The dimensions will be multiples of 16.

It allows for the extraction of the sampler's internal state.

Note: This tool is not intended for general users.

Note: Due to its many limitations and optimization for low-memory operation, this sampler is not suitable for standard image generation tasks.

MOSS-Image Merger

( model/merging/MOSS-Image Layer Merge )

A node that enables merging at the layer level.

It includes the following modes: weighted_sum, add_difference, and train_difference.

Unlike the merge nodes in ComfyUI, the alpha value represents the proportion of modelB.

ver.1.1

The default module settings for Merger were inappropriate.

As a new feature, you can now set numerical values ​​by pressing number keys while hovering your mouse cursor over a slider:

0: 0.0

1: 1.0

2: 0.2

...

9: 0.9

-: -1.0

Additionally, you can adjust the values ​​in increments of 0.05 using the arrow keys. Pressing Shift-0 sets the all values to 0.0, and Shift-1 sets ithem to 1.0.

MOSS_BF16 Converter

( model/convert/MOSS-Image to BF16)

Converts models (such as INT8 models) into a "Fused BF16" format, matching the structure of the base model. This tool functions as a standalone utility and does not require any external connections. The converted files will be stored in output/diffusion_models.