Seed hunting workflow for MOSS-Image SFT on Mac. Native nodes only, no custom packs (16 GB Macs need one, see below). Tested on an M5 Pro 48 GB and an M1 MacBook Air 16 GB.
What it does well:
Legible text in scene, English and Chinese. Signs, menus, posters, neon, multi-line.
Long prompts, photoreal skin, phone selfies, older faces, anime.
Settings I settled on:
25 steps, cfg 2.5, er_sde, simple. Official 50/cfg 3 gave same sharpness at double the time.
Negative prompt barely matters at cfg 2.5.
SLG (layer 22): +50% time, no visible gain. Off.
Prompt enhancer: ~15 min per prompt on Mac, no gain. Off.
8-step LoRA makes skin waxy (sharpness 3-10x lower). Kept as turbo_lora_hunt_mode for scanning comps at ~50 s, then re-render the pick with it off at the same seed.
Prompting: write a caption of the finished image. Visible text in quotes with position. No camera talk, no quality words, the model renders those literally.
Example: "A front-camera phone selfie of a man in his forties, arm extended, two days of stubble, cluttered kitchen behind him, flat window light."
Seed hunting: batch_size 2-4, or seed on increment and queue a few times.
Downloads on HF Comfy-Org/MOSS-Image (links in the workflow note): moss_image_sft_bf16 (10.2 GB), qwen3.5_9b_w4a8_gptq (7.3 GB), flux2-vae_small. Optional: lodestones 8-step DMD LoRA.
Memory: 10 GB DiT + 7 GB TE. 48 GB comfortable, 32 GB fine at 1 MP batch 1, 24 GB should work at 1 MP batch 1 (untested), 16 GB works with the int8 DiT (below). Steps don't change memory, resolution barely does. M5 Pro 48 GB: ~3.5 min per 1 MP image.
16 GB Macs (tested on an M1 MacBook Air 16 GB):
Swap the DiT for moss_image_sft_int8_convrot (5.4 GB, same HF repo).
Keep qwen3.5_9b_w4a8_gptq. The int8_convrot text encoder renders pure noise on MPS.
Install ComfyUI-AppleSilicon-FP8 to run int8 on MPS. The only custom pack, and only for 16 GB.
832x1216, 25 steps: 68.5 min per image, peak 15 GB, swap under 7 GB.
10 steps: 19.8 min, same composition, slightly softer skin and hair. Hunt at 10, re-render keepers at 25.
GGUF doesn't work yet: ComfyUI-GGUF rejects the moss_image architecture.