Introduction

Put together a simple text to image workflow after experimenting with MOSS-Image for the better part of the day today.

This model has actually been astounding me with its capabilities thus far. It essentially has built-in regional prompting which you can actually use without having a PHD in computer science, can generate realism or anime, supports native generation up to 4 megapixels, and supports both sfw and nsfw.

I think that the model may be on par with the likes of Anima and Krea 2 in some aspects. Let me know your guys' thoughts in the comments, as well as any suggestions or feedback you have on the workflow.

Overview

Started off with this official Comfy-Org MOSS workflow as a base: (Comfy-Org Workflow)

1. Replaced vae decode with tiled vae decode (I always do this as it's faster on my setup, if you want non-tiled you'll have to manually replace it).

2. Added support for turbo LoRA with toggle.

3. Changed saving path to promote more organization, saving to a separate 'moss' subfolder and grouping by date.

4. Other more minor changes.

Links to everything that you need is included within the workflow itself on the left-most markdown note.

My recommendations:

Diffusion model:

moss_image_sft_int8_convrot.safetensors, moss_image_sft_bf16.safetensors if your PC can handle it. After experimenting with moss image base and experiencing a large number of anatomical errors and choppiness, I realized that SFT outperforms it in almost every way, with the exception of prompt adherence.

Text encoder:

qwen3.5_9b_int8_convrot.safetensors, qwen3.5_9b_bf16.safetensors if your PC can handle it. No further comments.

Vae:

full_encoder_small_decoder.safetensors, always. Comfy-Org recommends flux2-vae_small.safetensors, which may be the same/around the same, however I've used full encoder small decoder for previous flux workflows and trust its efficacy. Without delving too deep into the science, full encoder small decoder allows for much faster vae encoding and decoding for negligible if not imperceptible degradation. If you don't trust this you can use the main vae which is flux2-vae.safetensors.

Usage

You need to make sure your comfy is updated to version v0.42.0 or higher, as native MOSS support was not added until then. Comfy just released v0.42.1 which fixed crashes pertaining specifically to MOSS so I highly recommend updating to that.

It's always recommended to use enable sage attention or CK (Comfy Kitchen) attention for faster generation times.

Next because the bounding box editor node only works with Nodes 2.0, you will need to explicitly enable it in settings if not enabled already.

Since most of the workflow was re-used, there's not really any additional knowledge I can provide.

As far as the workflow itself, Comfy-Org provided a detailed run-down as to both the workflow itself and MOSS architecture, so kudos to them.

The only part they were incorrect on is that they specify the default settings recommended by MOSS are 50 steps, cfg 3, Euler sampler, default scheduler, denoise 1, and shift 5 which is correct for the most part but MOSS recommends sgm_uniform for the scheduler, not default. I took this directly from the official comfy MOSS-Image workflow: (MOSS-Image workflow)

The settings I use also reflect the exact manner in which the above workflow is set up, with the exception of flux shift because Comfy built that into the diffusion model loader.

The MOSS user guide can also be found here: MOSS User Guide

The MOSS recommended default negatives of: nsfw, worst quality, bad anatomy, watermark, jpeg artifact appear to be sufficient in most instances. Obviously if you're attempting to do nsfw don't include nsfw in the negatives. Of course as more users refine their workflows better negative combinations may reveal themselves.

As far as positive prompting goes it's not really much different that Anima. MOSS states the following:

Recommended Tag Order: For the best results with tag sequences, we recommend the order Quality -> Score -> Meta -> Year -> Character -> Series -> Artist -> General

where artist tags are prepended with the @ symbol.

When using the bounding boxes you can leave the main prompt as is and simply describe what you want shown in that area and the workflow will automatically convert it into valid json.

Based on my limited experience overlapping bounding boxes appears to be fine and even may grant better results in some instances.

Simply toggle the turbo button to disable/enable turbo:

My recommendations are always Euler sampler, sgm_uniform scheduler, cfg 3, and SLG on, as is consistent with what MOSS themselves recommend.

When using turbo I recommend 6 steps, without it I recommend 50 steps. I was averaging about 6 minutes and 50 seconds for a 4 megapixel generation using the full BF16 diffusion model and text encoder, while enabling turbo brought me down to about 45-50 seconds.

The quality with and without turbo appears to be around the same, though of course turbo will be more prone to errors and won't have as good of prompt adherence. If you want the best quality result it's always recommended to disable it and go back to the MOSS recommendation of 50 steps.