MOSS-Image Panel Workflow — one ComfyUI panel, six kinds of picture, optional free-prompt box.

Pick a TYPE (people, poster with text, sale set, object scene, illustration, video first frame), pick a SIZE, press Queue. Prompt presets, the adult-people guard, file names and output folder are built in. SFW. Everything you touch is in the group called CONTROL PANEL.

🖼️ Made with this panel (SFT, one workflow, no edits)

People

Poster with text

Object / scene

Special (illustration)

Free-prompt box

🧾 What I tested (RTX 16 GB, ComfyUI 0.44)

• TYPES 1, 2, 4 and 5 ran through this exact panel (the files on this page): 2:3 images in 84–97 s each, peak VRAM 14.0 GB, no errors. TYPES 3 (sale set) and 6 (first frame) and all sizes ran in the earlier own-use version of the same graph, not through these exact files. The example images on this page were made with these files, and each PNG contains the workflow, so you can drag it into ComfyUI.

• Version SFT is the one to use. Version Base is the pre-training checkpoint, here for comparison and LoRA work. It is lighter and shows less detail than SFT (see below).

• SeedVR2 upscale is OFF by default. In my one SeedVR2 test (2:3 → 2132×3198, 140 s) the system RAM reached 99 % on a 64 GB PC with other programs open. Turn it on only when you have a lot of free RAM.

🆚 SFT vs Base, same prompt, same seed

Same prompt, same seed. In each pair: Base on the left, SFT on the right.

People shots: SFT gives a bigger, clearer subject with a visible face; Base often makes the person small and far away. In first-frame crops Base made faces look older. One Base test (a kawaii whale) came out as a blob. This comparison is from my earlier test of the same graph, not a benchmark.

⚡ Quick start

1. Download the file for the version you want (SFT recommended). It is only the graph, no model weights.

2. Put the model files in the usual ComfyUI folders (names below), install the custom nodes, restart ComfyUI, open the workflow.

3. In CONTROL PANEL set TYPE and SIZE (and SEED), leave FREE at 0, Queue. One Queue = one image.

4. Images are saved to ComfyUI/output/MOSS_Image/ as moss_<type>_<time>.png.

🎛️ The panel

• TYPE: 1 people · 2 poster with text · 3 sale set · 4 object scene · 5 special (art) · 6 first frame

• SIZE: 1 = 4:5 · 2 = 2:3 · 3 = 1:1 · 4 = 9:16 raw · 5 = 16:9 raw (4 and 5 are video first-frame sizes, never upscaled)

• SEED: fixed or random

• FINISH: 0 = raw (default) · 1 = SeedVR2 upscale (RAM heavy)

• extra words: text added to the preset. People presets say "adult … early 20s"; keep that wording

• FREE = 1: your own prompt in the FREE PROMPT box (no preset). STYLE = 1 adds a short "photorealistic quality" phrase. FREE SIZE: 1 = 2:3 · 2 = 1:1 · 3 = 3:2. A standard negative prompt is used, output is raw, file name moss_free_. STEPS 50 and CFG 3 are the MOSS defaults

• Free mode is your own prompt: it must follow the Civitai rules and the model licenses. Keep "adult … early 20s" for people

📦 Required models (this page does not host them)

• UNet: moss_image_sft_int8_convrot.safetensors (SFT version, 5.4 GB) or moss_image_base_int8_convrot.safetensors (Base version) → models/diffusion_models

• Text encoder: qwen3.5_9b_int8_convrot.safetensors (9.8 GB), loader type "moss_image" → models/text_encoders

• VAE: flux2-vae.safetensors → models/vae

• Optional upscale: SeedVR2 7B (seedvr2_ema_7b_fp16.safetensors) + ema_vae_fp16.safetensors

Pick the files in the loader nodes if your file names differ.

🧩 Required custom nodes

ComfyUI-Impact-Pack · was-node-suite-comfyui · ComfyUI-Custom-Scripts (pysssss) · comfyui_essentials · ComfyUI-Easy-Use · comfyui_layerstyle (Seed node) · ComfyUI-Image-Saver · rgthree-comfy (the SeedVR2 on/off switch) · ComfyUI-SeedVR2_VideoUpscaler (only for FINISH = 1)

⚙️ Settings used

50 steps · Euler / sgm_uniform · CFG 3 · model shift 5 · skip-layer guidance on layer 22. Prompts and seeds are saved in every PNG.

💡 Tips

• Keep FINISH at 0 while you explore.

• Open the workflow from the ComfyUI workflow list or drag in one of the example PNGs.

• Text inside posters is hit and miss; check the spelling.

⚠️ Limits

• SFW. No real people, no celebrities.

• Tested on a 16 GB card only.

• In my runs the Base version is weaker than SFT.

📝 AI disclosure

Every example image on this page is AI-generated with this workflow.

⚖️ Licenses and credits

• MOSS-Image (SFT and Base) by OpenMOSS: Apache-2.0 according to the model card.

• Qwen3.5-9B text encoder: Apache-2.0 on its Hugging Face card (https://huggingface.co/Qwen/Qwen3.5-9B).

• FLUX.2 VAE: two sources say different things, so read them yourself before you use it. The BFL GitHub README (https://github.com/black-forest-labs/flux2) says the FLUX.2 autoencoder is released under Apache 2.0 and links the file ae.safetensors in the FLUX.2-dev repository on Hugging Face. That repository (https://huggingface.co/black-forest-labs/FLUX.2-dev) is labeled FLUX Non-Commercial License. I make no claim about commercial use of the VAE, and I do not forbid it either. This page shares only the workflow graph, not the VAE.

• SeedVR2 by ByteDance (optional upscale): follow its license.

• The workflow graph itself is free to use.

⭐ If it saved you time, a like or a comment with your result helps others find it.