MOSS-Image SFT in one clean node: flip between 8-step turbo and full 50-step quality, with an optional built-in prompt enhancer that also writes your negative prompt.

What's inside

• Turbo switch. One toggle swaps everything at once:

  • ON: turbo LoRA, 8 steps, CFG 1. About 20 s per image on a 16 GB card.

  • OFF: no LoRA, 50 steps, CFG 3, negative prompt active. About 80 s.

• Prompt enhancer switch. Qwen3.5 9B (the same file Moss already uses as its text encoder, so nothing extra to download) turns a short idea into a full detailed prompt. It also writes a negative prompt tailored to that image, which gets used automatically when turbo is off.

• Same seed = same composition in both modes. Explore fast with turbo, then flip turbo off on the seed you like for the final render.

• VRAM-friendly. The text encoder is unloaded after prompting, so 50-step runs stay at full speed on 16 GB cards even with the enhancer on.

• Built as a subgraph, Klein-template style. Double-click the node to see the enhancer, its system prompt, and the final prompts it sent.

• The options on the node are tagged [turbo ON] / [turbo OFF], so you can see which ones each mode uses.

Downloads

• Diffusion model: MOSS Image SFT (Civitai), or moss_image_sft_int8_convrot_attn.safetensors / bf16. Goes in models/diffusion_models.

• Text encoder: qwen3.5_9b_w4a8_gptq.safetensors. Goes in models/text_encoders.

• VAE: flux2-vae.safetensors, or optionally the FLUX.2 small decoder. Goes in models/vae.

• Turbo LoRA: lora_latest.safetensors by lodestones, at strength 0.7. Goes in models/loras.

• Custom node: comfyui-deno-custom-nodes, for the text-encoder unload step. If you don't use the enhancer, you can bypass it.

Needs an up-to-date ComfyUI with native MOSS-Image support.

Settings used

• Sampler euler, scheduler sgm_uniform, 1 MP (832×1248 portrait by default).

• Skip-layer guidance on block 22 for the uncond pass (the official setting).

• Shift: no ModelSampling node is needed, because ComfyUI already applies shift 5 to MOSS-Image. I tested it, and adding it produced pixel-identical images.

Tips

• Be concrete. Moss takes prompts very literally. If something keeps going wrong (feet, poses, object size), describe the exact position instead of using a negative ("both feet resting on the wire floor of the basket").

• Put text you want rendered in "quotes". It handles posters and signs really well.

• The negative prompt only matters with turbo off. CFG 1 ignores it.

Tested on: AMD RX 9060 XT 16 GB (ROCm, Windows).

• Turbo: about 20 s per image, or about 30–55 s with the enhancer.

• Full 50 steps: about 80 s, or about 2 min with the enhancer.

Credits: MOSS-Image by OpenMOSS, ComfyUI repack by Comfy-Org, 8-step DMD turbo LoRA by lodestones, unload node by Deno.