⚠️ IMPORTANT UPDATE — HaDeS MOSS Suite v0.9.2 ⚠️
With the release of HaDeS MOSS Suite v0.9.2, which is now available, I am currently updating all my MOSS workflows.
Over the coming days, older workflows will be gradually removed and replaced with updated versions to take advantage of the latest improvements and features introduced in the Suite.
Thank you for your patience and continued support!
HaDeS MOSS Suite Workflows
This page gathers my ComfyUI workflows for MOSS-Image, each one a new version here. They are built around the HaDeS MOSS Suite nodes and kept simple enough to load and run. Every setting in them comes from measurements, not guesswork, and the measurements are public in the suite's tools/RESULTS.md.
More workflows will follow (merging, LoRA tools, layout prompts...). Follow the page to get them as they come out.
Simple Basic Generation workflow
HaDeS MOSS Suite v0.9.2 - Basic Generation (MOSS-Image, 25 steps or 8 with Turbo v1.0)
The workflow from my video tutorial "Your first MOSS-Image workflow in ComfyUI, from scratch". It uses six nodes from the free HaDeS MOSS Suite, plus Suite Telemetry so the saved PNG carries its settings.
▶ Video tutorial (built node by node): https://www.youtube.com/@Devildonia
What's inside
MOSS Loader: MOSS-Image Base, the Qwen3.5 9B int8 text encoder and the FLUX.2 VAE, in one node.
MOSS Prompt Enhancer in "as written" mode: your sentence is the prompt. MOSS reads natural language, not tag lists.
MOSS Latent: 1024 × 1024 and the seed.
MOSS KSampler: 25 steps, CFG 2.5, euler / simple.
MOSS Turbo LoRA: my Turbo 8-steps v1.0, trained for MOSS-Image Base, at 8 steps, CFG 1, euler / simple. Switch it off to go back to 25 steps.
MOSS Image Saver with Suite Telemetry: the PNG gets the prompt, the settings and the model / LoRA hashes as A1111 parameters.
Times on my RTX 5090 (1024 × 1024, models already loaded, as ComfyUI recorded them in the video): 13.4 s at 25 steps, 3.2 s with the turbo at 8 steps.
You need
HaDeS MOSS Suite: ComfyUI Manager → Custom Nodes Manager → search "HaDeS MOSS". Or https://github.com/Devildonia/HaDeS-MOSS-Suite for the latest updates.
MOSS-Image Base, the text encoder and the VAE: https://huggingface.co/Comfy-Org/MOSS-Image. ComfyUI offers to download missing files from the workflow.
Turbo 8-steps v1.0, in models/loras: https://huggingface.co/DeVilDoNia/HaDeS-MOSS-Turbo
Tip: with my model HaDeS MOSS v1.0 Turbo (https://huggingface.co/DeVilDoNia/HaDeS-MOSS-v1.0-Turbo), leave the turbo input empty. The turbo is already baked in.
Everything is free. MOSS-Image is Apache 2.0, and so are my LoRAs.
HaDeS MOSS Suite – Compare Models
Compare up to four MOSS-Image models side by side, with the same prompt, seed and settings, on one labelled sheet.
How it works
MOSS Compare Sampler samples every model plugged into model_A to model_D with the same prompt, latent, seed, steps, CFG and sampler, exactly as the MOSS KSampler would, and saves them as one sheet.
Each column is labelled from what the model actually carries: its file name, the merge it came from or its LoRAs. The labels don’t depend on how you wired the nodes. Switch a Multi-LoRA off and that column shows the bare model.
Turbo: connect a MOSS Turbo LoRA and it applies to every column, so the comparison stays fair.
Blind mode: shuffles the columns and numbers them 1 to 4. The key goes to the node’s info and into the PNG, never onto the sheet. Pick your favourite before you look.
Compare Sheet (a muted group in the workflow) puts any images side by side, for example two seeds of the same model.
One cable to Suite Telemetry checks the settings of every branch.
Changes in v0.9.0
More honest footer. The footer under the sheet (seed, steps, prompt…) now only shows what every panel shares. Before, if one image came without info (loaded from disk or made with another pack), the other branch’s seed and prompt went in the footer as if the whole sheet used them. Now they stay on that panel’s label.
The images output returns the whole batch. Before, it returned only the first image of each model, even when the sheet showed the full batch. Now model A’s batch comes first, then B’s, C’s and D’s. If one of your workflows uses this output, it will now receive more images.
New example workflow, Compare Models: SFT vs Base, ready to load from the Suite’s examples/ folder.
HaDeS MOSS · LoRA Compare x4 (MOSS Compare Sheet)
Compare up to four LoRAs (or four checkpoints of one LoRA) on MOSS-Image with the same prompt and seed, and get one labelled sheet: every panel is named automatically with the LoRA its branch used, and the seed and prompt go underneath.
Made to show the new MOSS Compare Sheet node from HaDeS MOSS Suite v0.8.0.
How it works
One MOSS Loader, one Prompt Enhancer and one MOSS Latent feed four branches: Multi-LoRA → MOSS KSampler, A to D.
The shared seed lives in MOSS Latent, so all four branches (and the prompt, if the Enhancer writes it) use the same one.
MOSS Compare Sheet reads each branch's info: the label is the LoRA name, with its strength when it is not 1. No typing, no stitch nodes, no scripts.
The sheet is saved as one PNG with the workflow inside, so you can drag it back into ComfyUI.
How to use
1. Pick a LoRA in each Multi-LoRA row (A to D). Same LoRA at different strengths works too: the label shows the strength.
2. Write your idea in the Prompt Enhancer (or set it to "as written" to use your prompt as typed).
3. Set the seed in MOSS Latent and queue. Change the seed to see the next comparison.
Tip: with a turbo model at CFG 1 the negative prompt is not used, so put "no text" or "no signature" in the positive prompt if you need it.
Requirements
HaDeS MOSS Suite 0.8.0 or newer (ComfyUI Manager / Comfy Registry, or github.com/Devildonia/HaDeS-MOSS-Suite)
A MOSS-Image model. The workflow is set for HaDeS MOSS v1.0 Turbo (8 steps, CFG 1): huggingface.co/DeVilDoNia/HaDeS-MOSS-v1.0-Turbo
Text encoder qwen3.5_9b_int8_convrot (or w4a8_gptq) and flux2-vae, from Comfy-Org/MOSS-Image
The LoRAs you want to compare
Feedback and comparisons welcome!
HaDeS MOSS Suite Surprise Generator: random prompts or your own idea
Hit Run and get a scene you didn’t plan. Leave the idea box empty and the MOSS Prompt Enhancer invents the prompt itself. The seed first draws a medium (26 of them, from documentary photo to ukiyo-e, manga or claymation), a subject and a place in time, then the model writes one scene inside that draw. Set the batch count to 20 and see what comes out.
Got an idea instead? Type it in the same box. Nothing else changes: it writes a long, detailed prompt from your idea, and every run gives a new take on it.
No extra model: the prompt is written by MOSS-Image’s own text encoder (Qwen3.5-9B).
One seed reproduces everything: the draw, the prompt and the image.
The written prompt is saved in the PNG metadata.
Needs: HaDeS MOSS Suite v0.7.5 (ComfyUI Manager / Comfy Registry), the MOSS-Image text encoder and the flux2 VAE. Made for HaDeS MOSS v1.0 Turbo (8 steps, CFG 1), but it works with any MOSS-Image model.
Easy ControlNet Workflow Quick Start Guide
Hey everyone! Here is the ControlNet workflow for MOSS-Image built around the latest v0.6.0 release of the HaDeS MOSS Suite.
This version introduces the dedicated MOSS Control (color / depth) node, specifically designed to handle MOSS's unique 384-channel concat adapter architecture properly.
Requirements & Downloads
Custom Node Suite:
HaDeS MOSS Suite (
v0.6.0+). Install via ComfyUI Manager or git clone into yourcustom_nodes/folder.
ControlNet Adapter:
Grab the Depth adapter by thedeoxen: MOSS-Image-depth-controlnet on Hugging Face. Put the
.safetensorsin yourmodels/loras/directory.
Speed / Turbo LoRA:
HaDeS MOSS Turbo — drops inference down to 8–10 steps without killing quality:
Available on Civitai or Hugging Face.
Base Setup:
MOSS-Image UNet (
moss_image_sft_int8_convrot.safetensorsor bf16/fp8).Text encoder (
qwen3.5_9b_w4a8_gptq.safetensors/ typemoss_image).VAE (
flux2-vae_small.safetensors– 128 channels).Any depth preprocessor of your choice (e.g.
DepthAnythingV2).
How to Use the Workflow
Load your reference photo into the
Load Imagenode.Depth map: Passed from the preprocessor straight into the
depth_mapsocket ofMOSS Control (color / depth).Color map & Reproducing 1:1 vs New Looks:
Near 1:1 Replica: Plug the original image into
color_mapand setcolor_modetoas given. It encodes the real photo’s VAE latent into the color channels, keeping the original lighting, contrast, and tones intact (no post-processing LUTs or color grading needed!).Palette / Layout Transfer: Keep the original image plugged into
color_map, but setcolor_modetoflat regions(8colors,4.0blur). This flattens the reference into color zones, letting the prompt reshape the clothing and textures while preserving the global palette.Depth Only (Fresh generation): Leave
color_mapunplugged and setcolor_modetooff (depth only). MOSS will only follow the geometry and generate everything else from scratch based on your prompt.
Control Strength:
Start at
1.0(the nominal trained weight). You rarely need to push it beyond1.0–1.1.
Turbo Sampling:
If using HaDeS MOSS Turbo, set your sampler to 8–10 steps, Euler, and Beta / Simple scheduler with CFG 1.0. (Push CFG to 1.1/1.2 if you need negative prompt).
Check the map preview directly out of the node's maps output to see exactly what’s being fed to the VAE. Have fun experimenting!
Simple Turbo Image to Image Workflow Example with Prompt Enhancer.
Check out this example workflow for HaDeS-MOSS-Suite: Simple Turbo Image to Image Workflow Example with Prompt Enhancer. It provides an at-a-glance look at how versatile the Prompt Enhancer can be for guiding your image-to-image pipeline, with full telemetry and HaDeS Turbo LoRA integration shown in the preview.
Introducing HaDeS MOSS Suite LoRa Toolkit v0.1
LoRa Merger Node
LoRa Extractor
LoRa Resize
Especially designed for MOSS IMAGE.
Simple Layout Editor v0.1
Simple workflow example of how to use the MOSS Layout Editor and the MOSS Layout Node.
Simple Turbo 8 steps + Upscale x2 v0.2
NEW NODES ADDED!!
MOSS Multi LoRa
MOSS Prompt Enhancer
MOSS Image Saver
Simple Turbo 8 steps + Upscale x2 v0.1
A simple MOSS-Image workflow: generate at 8 steps with our turbo LoRA, then optionally double the size.
Each step is a group that you switch on or off with the Fast Group Bypasser. Groups 2 and 3 are off by default, so you only render the 1024 image until you turn one of them on.
The turbo LoRA used here: HaDeS MOSS Turbo 8-steps v0.7 → [Civitai link to the Turbo model page], (also on Hugging Face: https://huggingface.co/DeVilDoNia/HaDeS-MOSS-Turbo).
What you need
- HaDeS MOSS Suite (ComfyUI Manager or https://github.com/Devildonia/HaDeS-MOSS-Suite)
- rgthree-comfy for the Fast Group Bypasser
- Models:
- diffusion_models: moss_image_sft_bf16 (or int8_convrot)
- text_encoders: qwen3.5_9b_int8_convrot
- vae: flux2-vae
- loras: HaDeS-MOSS-Turbo-8steps-v0.7
- upscale_models: RealESRGAN_x2plus.pth
Description
Check out this example workflow for HaDeS-MOSS-Suite: Simple Turbo Image to Image Workflow Example with Prompt Enhancer. It provides an at-a-glance look at how versatile the Prompt Enhancer can be for guiding your image-to-image pipeline, with full telemetry and HaDeS Turbo LoRA integration shown in the preview.


