
Moss Image Workflow
12GB VRAM | INT8-ConvRot | ComfyUI
Moss INTRODUCTION
Native MOSS support needs newer than 0.41.3 for Comfyui.
Compared to my original krea2 workflow, this is faster. (depends on checkpoints used of course).
The default is using .safetensors - I have not tested GGUF.
The text encoder is only used when the prompt changes, so ComfyUI can offload it during sampling.
MOSS accepts natural language, tags and JSON layouts, and it tends to do what the prompt literally says. Make it as natural as possible.
Sampler settings: 8 steps · CFG 1 · ER SDE · Beta · 1024×1536. Keep CFG at 1. (if using Muse By Stable Yogi Moss V1.0 Civ INT8-ConvRot C90.safetensors)
Fully tested on Comfyui V0.45.
Adapted from my Krea2 workflow for Moss Image
New default settings and seed set to 'fixed' on default.
The workflow is designed to maximize a 12GB VRAM GPU w/ 64GB System RAM. (32GB RAM is possible but untested, probably extremely tight.)
ComfyUI workflow for Moss Image generation that works with 12GB VRAM.
The workflow includes a Universal Model Loader for both safetensors and GGUF checkpoints (GGUF untested for this workflow). Optional SeedVR2 single-image restoration and a dedicated SISR upscaler, allowing you to generate, restore, and upscale images independently or combine both enhancement stages.
Important note: Control After/Before Generate should be set to 'before' in Comfyui settings NOT 'after' to preserve the correct seed generated. Make sure you set 'randomize' to 'fixed'. Otherwise you will get different results in both upscalers.
Here's the annoying Comfyui side effect; if 'Control AFTER Generate' is selected instead of 'before':
You generate an image using Randomize.
ComfyUI immediately changes the displayed seed for the next run.
You like the image and switch to Fixed.
It locks the new seed, not the seed that produced your image.
This is why its always crucial to use 'Before'. Not just this workflow, but most workflows. Comfyui has many annoying quirks, this is just one of them. 'After' is the original behavior.
If all else fails and you have a different seed, you can just drag the img you generated into comfyui to get the correct seed, but Control Before Generate is the only way to go.
loosely based on @Li3sOfPaul workflow.
1. REQUIRED CUSTOM NODES
Install the following 7 custom node packages:
Required ComfyUI Settings
Make sure Modern Node Design (Nodes 2.0) is disabled in Settings.
For PreviewMonitor:
Lite Graph → Canvas → Live selection: Enabled
Comfy → Execution → Live preview method: Auto
2. REQUIRED UPSCALE MODEL
Download:
Installation destination:
ComfyUI\models\upscale_models\
This is the dedicated 4× SISR super-resolution model used in the workflow.
3. RECOMMENDED Moss Image CHECKPOINT
Recommended for 12GB VRAM
Muse By Stable Yogi Moss V1.0 Civ INT8-ConvRot C90.safetensors
Checkpoint: V1.0 INT8
Main page https://civarchive.com/models/3035057/muse-by-stable-yogi-moss?modelVersionId=3448836
Recommended Text Encoder/CLIP
Official VAE:
Performance & Output
RTX 3080 12GB: Approximately 8 seconds per image using the workflow's default settings.
Two-pass upscaling: Approximately 35s.
Output directory:
ComfyUI\output\txt2img\
Universal Model Loader Configuration
Change model_type:
diffusion_modelfor standard Safetensors checkpoint files.gguf_netfor GGUF checkpoint files.
Set clip_name and vae_name for both model types once. You can then switch between model types without having to configure them again.
Remember to change the checkpoint filename when switching model types. Use .safetensors or .GGUF as appropriate.
4. SEEDVR2 MODELS
The SeedVR2 DiT and VAE models should download automatically on the first run if they are not already installed.
Recommended models for 12GB VRAM
DiT Model:
seedvr2_3b_fp8_e4m3fn.safetensors
VAE Model:
seedvr2_ema_vae_fp16.safetensors
The 3B FP8 model is recommended for this workflow and performed better than the 7B variant in my testing. Though more testing may be needed. You can still use 7B model, as I had no memory issues at all.
5. ATTENTION OPTIMIZATION (OPTIONAL)
Attention Backends
SeedVR2 supports several attention backends:
Backend: SDPA
RTX 3080: Supported
Backend: SageAttention 2
RTX 3080: Supported
Backend: SageAttention 3
RTX 3080: Not supported (RTX 50 Series)
Backend: Flash Attention 2
RTX 3080: Supported
Backend: Flash Attention 3
RTX 3080: Not supported (Hopper / H100+)
Only one backend is needed at a time.
SDPA is the default and requires no additional installation. SageAttention 2 or Flash Attention 2 may improve processing speed on compatible GPUs.
If no optimized backend is installed, SeedVR2 automatically uses SDPA. Processing may take longer.
No additional packages are required if you do not have an optional Sage or Flash backend. Though I recommend installing one for this and future use in all workflows.
6. HOW TO USE
Fast Groups Bypasser
The workflow uses four groups to control image generation and enhancement stages.
First run: Enable Groups 0 + 1.
Once happy with the generated image, disable Group 1 to prevent duplicate saves.
Set Seed to 'Fixed'. (Control Before Generate) (prevents re-running Group 0)
Enable Group 2 (SeedVR2 Upscaler).
Enable Group 3 (SISR Upscaler).
Or leave them all on, it's designed to run every node from start to finish, it will just take longer on a 12GB Card. This is why it's designed like this.
Keep Group 0 enabled throughout. It contains the shared image source used by the other groups.
Upscaling (2 Passes)
If SeedVR2 smooths the image more than you prefer, you can disable it and go straight to SISR.
This lets you choose between SeedVR2 restoration, direct SISR upscaling, or combining both.
7. SEEDVR2: SINGLE-IMAGE RESTORATION
SeedVR2 uses a diffusion-transformer model to restore and enhance a single image, reconstructing fine details, textures, and clarity while preserving composition.
This works incredibly well with Moss Image.
Model: SeedVR2 3B FP8
VAE: EMA VAE FP16
Downloads automatically on first run. (Check CMD prompt)
If you find the smoothing too strong, simply disable Group 2 in the Fast Groups Bypasser.
8. SISR UPSCALER
Single-Image Super-Resolution (SISR) enhances a single input image by increasing its resolution while reconstructing fine details.
Upscale Model: DF2K.pth
This 4× super-resolution model enlarges the image by four times in each dimension, reconstructing additional detail and improving clarity.
The final output scale can be adjusted using the Upscale Image By node.
When skipping SeedVR2, you can either:
Bypass the Scale By node (CTRL+B) for the full 4× upscale.
OR
Increase Scale By from 0.50 to 1.0 for the full 4× upscale.
Output Resolution
The default setting of 0.50 produces 2× the input dimensions.
Increasing the setting to 1.0 produces the full 4× upscale.
Actual output resolution depends on the image entering the upscaler.
For example, assuming a 4K image from SeedVR2:
0.50 = approximately 8K output.
1.0 = approximately 16K output.
If SeedVR2 is bypassed, the original image is used instead, so the resulting resolution will be lower.
Full 4× upscaling can be overkill, depending on what you're generating.
Important: File Sizes
File sizes can become very large at full 4× resolution.
Avoid switching the final Image Saver (3rd Image Saver node) to PNG. At these resolutions, PNG files can exceed 200 MB.
This is why WEBP format is used for the final image, offering considerably smaller file sizes.
9. ADDITIONAL WORKFLOW OPTIONS
Control Before/After Generate
Seed: The workflow ships with the seed set to fixed. Group 2 and Group 3 reuse the image from Group 0, so a changing seed gives you different results in each stage. If you do switch it to randomize, set ComfyUI’s control mode to “before” first.
Sharpen Image
You can increase or decrease Sigma and Alpha to adjust the sharpening effect to your preference.
Feel free to disable the Sharpen Image node if you don't like its results but the default settings are very minor. Select the node and press CTRL+B to bypass it.
Have fun experimenting! The workflow is designed to maximize a 12GB VRAM GPU w/ 64GB RAM. (32GB RAM is possible but untested, probably extremely tight.)
Description
initial release




