🔗 RunningHub Workflow:
https://www.runninghub.ai/zh-cn/post/2117070971680444418/?inviteCode=rh-v1497
▶️ YouTube Tutorial:
https://youtu.be/dyHyUmQfnoQ
New to RunningHub? Sign up through my link to receive 1,000 free credits and try the workflow.
I've put together two MOSS-Image workflows with three prompt input options and SeedVR2 upscaling. Here are the setups I enjoyed most after testing realistic portraits, stylized artwork, and sci-fi scenes.
Two generation setups
Workflow A — Single-pass sampling with two LoRAs
Produces a softer, more atmospheric look with less pronounced detail. I like its dreamy, grainy aesthetic, although prompt adherence was weaker in my tests.
Workflow B — SFT + Base sampling
Starts with the SFT model and finishes with Base. This was my preferred setup for natural skin texture, facial expressions, and more defined subjects. Most of the examples in my video were made with this workflow.
My default is 50 steps: 35 with SFT and 15 with Base. An 8-step option is also included, split evenly between the two models. It is faster, but I noticed a clear drop in image quality.
Three ways to provide a prompt
• Manual input: Use a prompt you've written or reviewed yourself.
• Automatic expansion: Turn a short idea into a detailed prompt.
• Image-to-prompt: Analyze a reference image and generate a prompt using the included MOSS-focused template.
For my examples, I used DeepSeek V4.1 Flash 1017 for image analysis and prompt generation.
Model setup and upscaling
The demonstrated setup uses MOSS INT8, Qwen 3.5 9B as the text encoder, and the FLUX.2 VAE. You can also try BF16 if you have enough VRAM.
SeedVR2 upscaling is included. Select the output dimensions separately for each generation workflow—the reference image dimensions are not applied automatically.
The prompt-processing nodes shown in the tutorial use RunningHub's built-in tools. For local use, replace them with compatible nodes for your chosen model or API.
The original generation workflows are by their respective creators; my combined setup adds the prompt input options and upscaling path demonstrated here.
I was particularly impressed by the Midjourney-like atmosphere MOSS can produce. Try both setups and let me know which style you prefer!
💬 Discord Community:
https://discord.gg/NmBhk9Vk8b