The rest of the set (Qwen 4-step, Z-Image, img2img, upscaling): tinytensor.gumroad.com
MOSS 2.5D Anime, text to image, 1024x1536 on an 8GB laptop card — about 62 seconds per image.
One workflow. Drag it onto the canvas, point the loaders at the right files, press Run.
WHAT YOU NEED
mossImage25DAnime_sftINT8.safetensors — 5.41 GB — models/diffusion_models
qwen3.5_9b_int8_convrot.safetensors — 9.75 GB — models/text_encoders
flux2-vae.safetensors — 0.34 GB — models/vae
ComfyUI current build. No custom nodes — everything in this graph is core.
32GB system RAM is not optional. The text encoder is 9.75GB and cannot share an 8GB card with a 5.4GB diffusion model, so ComfyUI keeps it in RAM. On 16GB it pages and crawls.
SETTINGS
steps 25 — CFG 2.5 — er_sde / simple — 1024x1536 — text encoder type moss_image
The latent node must be EmptyFlux2LatentImage, not EmptyLatentImage.
CFG is 2.5 and not 1.0, so the negative prompt is actually computed here. Keep the guards.
MEASURED, NOT ESTIMATED
RTX 3070 Ti Laptop 8GB / Ryzen 7 6800H / 32GB RAM / Windows 11 / ComfyUI 0.42.1
70.3s on the first run (loads the model), then 62.2s and 66.3s steady at 1024x1536.
TWO THINGS THAT GO WRONG
No outfit in the prompt, and the model removes the clothes. Give it only face and style words and it renders a nude figure. Name the garment — wearing a navy wool coat and cream scarf — and keep nude, nsfw, topless, revealing clothing, cleavage in the negative. Five words now, one regenerated batch later.
The style pulls hard. It lives in the checkpoint, not a LoRA, so 2.5D, semi-realistic anime, subsurface scattering moves the whole image. Want flat cel colour instead? Say so and negate the 2.5D words.
PROMPT STRUCTURE
Five layers, or the output drifts: shape (anime face proportions, small nose, pointed jawline, large eyes) · materials (subsurface scattering, band highlight on hair clumps) · rendering (2.5D, rim lighting, ambient occlusion, no lineart) · optics (shallow depth of field, volumetric light) · composition (close-up, upper body).
Without shape words you get a 3D-render pretty woman. Push shape too hard and the skin turns porcelain and the joints look like a figurine's. Keep porcelain skin, doll joints, mannequin, figurine, plastic sheen in the negative next to lineart, outline, cel shading.
Hide the hands when the frame allows it — hands out of frame, hands behind her back. MOSS returns a clean frame that way, and a hidden hand cannot be a broken one.
LICENCE
MOSS is Apache 2.0 (upstream OpenMOSS-Team/MOSS-Image-SFT) — the images you generate are yours, including commercially. The workflow file: use it freely, do not resell the file itself.
Models are not bundled; they are free to download and their licence is not mine to redistribute.
Setup notes are in SETUP.md inside the zip.
The other two MOSS workflows (image-to-image restyle, photoreal) plus the Qwen-Image and Z-Image ones are in the paid pack on Gumroad:
https://tinytensor.gumroad.com/l/comfyui-8gb-vram
Ask in the comments if something breaks — I read them.