A ready-to-use ComfyUI text-to-image workflow for MiaoMiao Harem MOSS, tuned and measured on an 8GB laptop GPU. Drag the JSON into ComfyUI and it runs.
WHAT IT SOLVES
The model card says to use the MOSS pipeline but not how. MOSS Image is not an SDXL checkpoint: it needs a different text encoder, a different VAE, a flow-matching shift and roughly 50 steps. Wire any of those wrong and the output is mush.
REQUIRES
• MiaoMiao Harem MOSS int8, 5.04 GB (the fp16 file is 9.47 GB and will not fit in 8 GB)
• Text encoder: qwen3.5_9b_int8_convrot.safetensors, loaded with CLIPLoader where type = moss_image
• VAE: flux2-vae.safetensors
SETTINGS IN THE WORKFLOW
• Steps 50, CFG 3.0, sampler euler, scheduler simple
• Shift 5.0 via ModelSamplingAuraFlow (node 20)
• Resolution 1024 × 1408 — both values must be multiples of 16
• SkipLayerGuidanceDiT included but bypassed by default; select it and press Ctrl+B to enable (layers 22, scale 3.0)
MEASURED
RTX 3070 Ti Laptop, 8 GB VRAM, 32 GB RAM: 140–148 seconds per image at 1024 × 1408, 50 steps, clean queue.
GOTCHAS
1. CLIPLoader type must be moss_image. Any other type errors out or produces noise.
2. Width and height must be multiples of 16.
3. Keep ModelSamplingAuraFlow connected. Without the shift the compositions collapse into tight bust shots.
4. The model leans toward uniform-and-schoolyard wardrobe; name the clothing explicitly if you want something else.
INCLUDED
• miaomiao-harem-moss-t2i.json — the workflow
• 6 sample images with the exact prompts in the PNG metadata
Free to use and redistribute. Model weights belong to their author.
BASE MODEL
MiaoMiao Harem MOSS by MIAOKA — https://civarchive.com/models/3015023
Built for the int8 variant (5.04 GB). The fp16 file is 9.47 GB and will not fit in 8 GB of VRAM.
OTHER 8GB WORKFLOWS
Qwen-Image 2.5D Material Style System — https://civarchive.com/models/3031008
NoobAI + OpenPose Pose Control — https://civarchive.com/models/3031015
ALSO FREE ON GUMROAD
No account needed, direct download — https://tinytensor.gumroad.com/l/ygrzk
Description
Initial release. 50 steps / CFG 3.0 / euler / shift 5.0, 1024x1408. Measured 140-148s per image on an RTX 3070 Ti Laptop (8 GB). SLG node included but bypassed by default.











