MOSS Image SFT int8 workflow for clean, natural photos, with 2 free LoRA slots ready for your own LoRAs (off by default, Ctrl+B to enable).
Skip Layer Guidance on layer 22 for cleaner structure
Optional HD redraw with Ultimate SD Upscale 2x
Resolution picker: just choose a ratio + megapixels
Same clean design as my other workflows: numbered groups, group toggles, readable node titles
Built and tested on an RTX 4070 Laptop 8GB / 16GB RAM.
These workflows are built and improved with the help of Claude (Anthropic). The collection keeps evolving with it, so expect regular updates.
Name : 1.0 · MOSS × USD
--- Version changes or notes ---
Flow: generate (30 steps) -> save
-> USD 2x (group ②) -> color match -> save HD
Toggle each group on/off with the bars above the groups.
Settings:
Sampler: euler / sgm_uniform, 30 steps, CFG 3
Skip Layer Guidance: layer 22, from 0 to 50 %
Negative prompt: empty
Default size: 3:4 portrait, 1.0 MP (change it in “Resolution”)
USD: 2x with 4x Remacri, 20 steps, denoise 0.15, color match mkl 0.7
About the HD pass:
MOSS shines at its native 1024 px, so USD upscales straight from pass 1 for the cleanest result. A latent second pass is included but muted; press Ctrl+M on those nodes if you want to experiment with it.
Models:
mossImage_sft_pruned_int8.safetensors (UNET)
mossImage_sft_txt_int4.safetensors (text encoder, type moss_image)
flux2-vae.safetensors (VAE): https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy/resolve/main/vae/flux2-vae.safetensors
4x_foolhardy_Remacri.pth (upscaler)
Custom nodes:
rgthree-comfy: https://github.com/rgthree/rgthree-comfy
KJNodes: https://github.com/kijai/ComfyUI-KJNodes
ComfyUI_UltimateSDUpscale: https://github.com/ssitu/ComfyUI_UltimateSDUpscale
Works with any recent ComfyUI with MOSS Image support.
These workflows are built and improved with the help of Claude (Anthropic): testing recipes, cleaning the files, writing the notes. The collection keeps evolving with it, so expect regular updates.Description
DIRTY MOSS × SeedVR — real-photo realism on an 8 GB laptop
A workflow built around Dirty MOSS V2 (by Sateluco), the MOSS model that goes for imperfect, "found in a camera roll" photos instead of the polished AI look. I use it every day for my realistic images: harsh phone flash, crushed blacks, natural skin, and text that stays sharp (signs, badges, labels).
HOW IT WORKS
- Pass 1: Dirty MOSS, 25 steps, CFG 2.5, er_sde / simple (the author's recipe; shift 5 is already the MOSS default in ComfyUI)
- SeedVR2 upscale ×2 (7B int8, tiled 512 so it fits in 8 GB)
- Post FX: my CRT settings for realism (levels + light grain), with its own upscale off since SeedVR already does it
- Optional, bypassed: your own MOSS LoRA, an 8-step DMD LoRA (faster but loses texture), a latent HD pass and Ultimate SD Upscale for bigger GPUs
TIPS
- Don't go below 25 steps: at 15 the lights, text and small parts fall apart.
- Describe the photo like a real snapshot (phone flash, 3 a.m., handheld blur, a messy room) rather than a studio shot.
- The negative prompt works here (CFG 2.5), unlike turbo workflows at CFG 1.
- The model leans NSFW: describe clothing if you want to keep it SFW.
MODELS
- diffusion_models: Dirty-MossV2-CONVROT8.safetensors (Dirty MOSS V2, CONVROT8 version)
- text_encoders: mossImage_sft_txt_int4.safetensors
- vae: flux2-vae.safetensors
- diffusion_models: seedvr2_7b_int8_convrot.safetensors
- vae: seedvr2_ema_vae_fp16.safetensors
CUSTOM NODES
- rgthree-comfy
- CRT Post-Process Suite (crt-nodes)
- Optional groups only: ComfyUI_UltimateSDUpscale, ColorMatchV2
SeedVR2, Skip Layer Guidance and the Resolution Selector are native ComfyUI nodes (update ComfyUI).
Cover made with this workflow + my noirelle LoRA.
Model credit: Dirty MOSS by Sateluco.
