A light, fast Brain for MOSS-Image. Type a rough idea, let the built-in prompt enhancer write the full caption, and get a square, a phone and a wide picture from one prompt and one seed. Five LoRA slots, an enhancer of its own and overnight batches from a plain text file. Every part sits in a colour-coded group with a short note beside it.
READ THE FULL GUIDE HERE!
Why you will like it
- A prompt enhancer that is just a switch. Type tags, plain words or a half sentence. The enhancer rewrites it into a full MOSS-style caption. Turn it off and your words go through exactly as typed. A preview box shows the prompt that is really sent.
- Pick your own enhancer model. The enhancer has its own loader, so you can try different text models without touching the picture text encoder.
- Enhancer LoRA slot. A LoRA that changes only what the enhancer writes, not the picture model.
- Three shapes, draft then hi-res. Square, phone and wide lanes from the same prompt and seed. Mute a lane to skip it. Switch on the HI-RES group under a lane to enlarge and refine that picture 1.5 times. A simple model upscale group handles any picture without new detail.
- Five LoRA slots. Run top to bottom, one key to turn each on or off.
- Overnight batches. Put prompts in a text file, split them with a line of =====, turn the prompt file on and queue.
- Four files, pick your size. Pictures only, pictures + Wan, pictures + MiniMax H3, or video only. The Wan and H3 versions add a simple 3-part chain to extend a video. The video-only file has the full long chain: up to 10 scenes for Wan and H3, each H3 scene can start from and end on its own picture, and extra LoRA slots.
- Clean copies for posting. Hover a picture and press C. A copy is saved with the prompt hidden. Your model, LoRAs and seed stay, so the Resources box still fills in on Civitai.
## What is inside
| Group | What it does |
|---|---|
| MODELS | The MOSS model, the text encoder and the VAE. |
| LORA STACK | Five slots, run top to bottom. Ctrl+B turns a slot on or off. |
| ENHANCER ENGINE | The enhancer with its own text model loader, its system prompt, and an enhancer-only LoRA slot. |
| THE BRAIN | Prompt, enhancer switch, final prompt preview, negative, seed, steps, sampler, CFG. |
| BATCH: PROMPT FILE | Overnight batch from a text file. |
| SQUARE, PORTRAIT, WIDE lanes | Three pictures from one prompt. Mute a lane to skip it. |
| HI-RES lanes | Enlarge and refine a lane's picture 1.5 times with the same seed. |
| F. Simple Upscale | Model upscale for any picture, no new detail. |
| Video files | 2_MOSS_Brain_Images_plus_Wan.json (Wan), 1_MOSS_Brain_Images_plus_MiniMax_H3.json (MiniMax H3 with sound), and 4_Video_Only_Wan_and_H3.json (both, with the 10-scene long chain). |
## Quick start (4 steps)
1. Copy the anima_batch_loader folder from the download into ComfyUI/custom_nodes/, then copy the download's input folder contents into ComfyUI/input/. Restart ComfyUI. No other custom nodes are needed for pictures. ComfyUI-GGUF is needed for the Wan video groups only.
2. Put the model files in their folders (list below).
3. Drag the workflow JSON onto ComfyUI. Pick your files in the model boxes.
4. Type a prompt in YOUR PROMPT and press Run.
You need a ComfyUI recent enough to support MOSS-Image, so update it first.
## Models (from [Comfy-Org/MOSS-Image](https://huggingface.co/Comfy-Org/MOSS-Image))
| What | Folder |
|---|---|
| moss_image_sft_bf16.safetensors (the int8 and Base files work too) | models/diffusion_models |
| qwen3.5_9b_int8_convrot.safetensors | models/text_encoders |
| flux2-vae.safetensors | models/vae |
The picture text encoder must stay Qwen3.5 9B with type moss_image. A few-step MOSS LoRA (for example a 4-step speed LoRA) goes in models/loras.
Optional, for video: the Wan 2.2 GGUF models and lightning LoRAs, and the MiniMax H3 model files. The file names and folders are listed in the guide, section 4. The simple upscale group needs an upscale model such as 4x_foolhardy_Remacri.pth in models/upscale_models.
## Settings that work
Normal SFT or Base at about 30 steps, CFG 3, euler (the official guide suggests 50 or more steps for best quality). With a 4-step speed LoRA in Slot 1: 4 steps, CFG 1. With an 8-step model: 8 steps, CFG 1.
## Notes
- Swap enhancer models in the ENHANCER MODEL loader. If you swap the picture text encoder instead, the sampler fails with a "mat1 and mat2 shapes cannot be multiplied" error.
- The enhancer is optional. No enhancer file? Set its switch to OFF, select the Enhancer group and press Ctrl+M.
- Leave your computer awake for overnight runs, and ComfyUI open.
The full guide, with every group explained step by step, is in the article.
## Video
Pick the file that fits: 3_MOSS_Brain_Images_Only.json (pictures only), 2_MOSS_Brain_Images_plus_Wan.json (pictures + Wan), 1_MOSS_Brain_Images_plus_MiniMax_H3.json (pictures + MiniMax H3) or 4_Video_Only_Wan_and_H3.json (video only, both models). In a video file, load any picture you made in the Video Controls box, turn on one video switch (V2 for Wan, V4 and V4a for H3) and press Run. To extend a video turn on the chain group (V2x or V4x) and write what happens next in the Chain Part boxes. The video-only file goes to 10 scenes: pick how many in the VC2 panel and write PART 2: to PART 10: in your prompt. Each scene is another full run, so more scenes take longer.