Uploaded October 2026 | Updated November 2026, 2 days ago
I tested MOSS-Image in ComfyUI: image editing with a character reference LoRA, layout control with Create Bounding Boxes, inpainting, and the Color/Depth Control LoRA. In my test, the skin looked better than Qwen Image 2.1 and FLUX 3, but dancing hands and inpainting failed.
In this video, I show where MOSS-Image is great, where it fails, and the settings that gave me my best results. I compare Base and SFT, and I test realistic portraits, a tea stall scene, product photos, anime and posters. Then I try the advanced workflows: Create Bounding Boxes for layout and text control, a character reference LoRA, and the Color/Depth Control LoRA to copy a pose from a photo. At the end, I test a 4-step model that is much faster.
MY VERDICT
Great for: realistic photos, product shots, layout and text control, copying a pose
Use another model for: hard hand poses (like dancing) and inpainting (editing part of an image)
Want speed? Try the 4-step model.
MY BEST SETTINGS
SFT: sampler er_sde, scheduler simple or sgm_uniform, up to 30 steps, CFG 2.5
Base: sampler Euler, scheduler sgm_uniform, 50 steps, CFG 3 (same as the official settings)
For the most realistic images, try Base at 50 steps.
Negative prompt I used: plastic skin, waxy skin
4-step model or 4-step LoRA: 4 steps, CFG 1, sampler Euler, scheduler simple
At CFG 1, the negative prompt does nothing, so you do not need one.
Bigger than 1024 x 1024? The model page says 6 steps for about 1536 px wide, and 8 steps for 2048 x 2048.
WHAT YOU NEED
ComfyUI 0.42.1 or newer (older versions crashed in the MOSS-Image templates)
Model: MOSS-Image SFT in BF16 or INT8 (Base is good for training your own LoRA)
Text encoder: Qwen3.5 9B BF16
VAE: FLUX.2 VAE
LoRAs: MOSS-Image Color Depth Control, MOSS-Image CharRef PoC
Custom node: ComfyUI-MOSS-Image-Ostris-Edit
4-step model: SFT 4-step v2.0 in FP32, about 20 GB (the normal BF16 model is about 10 GB). If 20 GB is too big for your graphics card, use the 4-step LoRA instead.
LINKS
Workflows: patreon.com/ComfyuiWorkflowBlog/posts/moss-image-6-and-173093243
MOSS-Image models (SFT and Base): huggingface.co/Comfy-Org/MOSS-Image/tree/main/diffusion_models
Qwen3.5 9B text encoder: huggingface.co/Comfy-Org/MOSS-Image/tree/main/text_encoders
FLUX.2 VAE: huggingface.co/Comfy-Org/MOSS-Image/tree/main/vae
Color Depth Control LoRA: huggingface.co/Comfy-Org/MOSS-Image/tree/main/loras
CharRef PoC LoRA: huggingface.co/pleiades7/MOSS-Image-CharRef-PoC
ComfyUI-MOSS-Image-Ostris-Edit node: github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit
My updated nodes.py file:
4-step model (FP32) and 4-step LoRA: huggingface.co/akhaliq/MOSS-Image-SFT-4-steps
#MOSSImage #ComfyUI #AIImageGeneration
I tested MOSS-Image in ComfyUI: image editing with a character reference LoRA, layout control with Create Bounding Boxes, inpainting, and the Color/Depth Control LoRA. In my test, the skin looked better than Qwen Image 2.1 and FLUX 3, but dancing hands and inpainting failed.
In this video, I show where MOSS-Image is great, where it fails, and the settings that gave me my best results. I compare Base and SFT, and I test realistic portraits, a tea stall scene, product photos, anime and posters. Then I try the advanced workflows: Create Bounding Boxes for layout and text control, a character reference LoRA, and the Color/Depth Control LoRA to copy a pose from a photo. At the end, I test a 4-step model that is much faster.
MY VERDICT
Great for: realistic photos, product shots, layout and text control, copying a pose
Use another model for: hard hand poses (like dancing) and inpainting (editing part of an image)
Want speed? Try the 4-step model.
MY BEST SETTINGS
SFT: sampler er_sde, scheduler simple or sgm_uniform, up to 30 steps, CFG 2.5
Base: sampler Euler, scheduler sgm_uniform, 50 steps, CFG 3 (same as the official settings)
For the most realistic images, try Base at 50 steps.
Negative prompt I used: plastic skin, waxy skin
4-step model or 4-step LoRA: 4 steps, CFG 1, sampler Euler, scheduler simple
At CFG 1, the negative prompt does nothing, so you do not need one.
Bigger than 1024 x 1024? The model page says 6 steps for about 1536 px wide, and 8 steps for 2048 x 2048.
WHAT YOU NEED
ComfyUI 0.42.1 or newer (older versions crashed in the MOSS-Image templates)
Model: MOSS-Image SFT in BF16 or INT8 (Base is good for training your own LoRA)
Text encoder: Qwen3.5 9B BF16
VAE: FLUX.2 VAE
LoRAs: MOSS-Image Color Depth Control, MOSS-Image CharRef PoC
Custom node: ComfyUI-MOSS-Image-Ostris-Edit
4-step model: SFT 4-step v2.0 in FP32, about 20 GB (the normal BF16 model is about 10 GB). If 20 GB is too big for your graphics card, use the 4-step LoRA instead.
LINKS
Workflows: patreon.com/ComfyuiWorkflowBlog/posts/moss-image-6-and-173093243
MOSS-Image models (SFT and Base): huggingface.co/Comfy-Org/MOSS-Image/tree/main/diffusion_models
Qwen3.5 9B text encoder: huggingface.co/Comfy-Org/MOSS-Image/tree/main/text_encoders
FLUX.2 VAE: huggingface.co/Comfy-Org/MOSS-Image/tree/main/vae
Color Depth Control LoRA: huggingface.co/Comfy-Org/MOSS-Image/tree/main/loras
CharRef PoC LoRA: huggingface.co/pleiades7/MOSS-Image-CharRef-PoC
ComfyUI-MOSS-Image-Ostris-Edit node: github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit
My updated nodes.py file:
4-step model (FP32) and 4-step LoRA: huggingface.co/akhaliq/MOSS-Image-SFT-4-steps
#MOSSImage #ComfyUI #AIImageGeneration










