Faster MOSS-Image: is there a Turbo version?

MATLOWAI merged lodestones' 8-step DMD LoRA (at 0.35) and the SFT stripe fix into one int8 model, so it loads like the official file.

1. Download moss_image_sft_deband_fast_dmd035_int8_convrot.safetensors (5.4 GB) into ComfyUI/models/diffusion_models/.
2. Use the qwen3.5_9b_int8_convrot.safetensors text encoder and the full flux2-vae.safetensors VAE, as the model card does.
3. Pick the new file in Load Diffusion Model and set 10 steps, or drag in the card's moss_deband_int8.json workflow, which sets the sampler for you.

The same repo has DeBand and DeBand+ without the speed-up, for 25 steps.

lodestones distilled SFT from 50 steps with CFG down to 8 steps without CFG; darask0 trained a 4-step adapter on top of it. Both run through darask0's rapid-moss node, which uses the sampler the adapters were trained with.

1. Copy the comfyui/rapid-moss folder from darask0's Hugging Face repo to ComfyUI/custom_nodes/rapid-moss/ and restart ComfyUI.
2. Put each adapter's three files (lora.safetensors, time_embedding.safetensors and dataspace.safetensors) in its own folder under ComfyUI/models/rapid_moss/. For lodestones' 8-step version, rename lora_latest.safetensors to lora.safetensors.
3. Load the rapid-moss-4step-native.json example workflow. It uses the core loaders: moss_image_sft_bf16.safetensors in Load Diffusion Model, the Qwen3.5-9B text encoder (type moss_image) in Load CLIP, and the FLUX.2 VAE.
4. Keep the size near 1 megapixel: 1024×1024, 1216×832, 832×1216, 1344×768 or 768×1344. Larger images break into a grid of small repeated subjects; detailed prompts help.

ComfyUI/
├── custom_nodes/
│   └── rapid-moss/
└── models/
    └── rapid_moss/
        ├── rapid-moss-4step/
        │   ├── lora.safetensors
        │   ├── time_embedding.safetensors
        │   └── dataspace.safetensors
        └── lodestones-8step-latest/
            └── (the same three files)

• lodestones' card says its LoRA also works in ComfyUI's normal LoRA loader (strength 1, or 0.7 for lora_latest). darask0 found that the core LoraLoaderModelOnly adds these fp32 adapters into bf16 weights with an 81% error on the change: you get a picture, but not the one the adapter was trained to make. For 4 to 8 steps, use the rapid-moss node.
• Noise, a black image or “ModuleDict … has no attribute q”: update the rapid-moss node, which now handles the Comfy-Org files with fused q/k/v. Older preset workflows needed a split q/k/v copy of SFT instead (10.2 GB, from MATLOWAI or nynxz).

ComfyUI-Spectrum-MOSS forecasts part of the transformer's work from earlier steps instead of computing every step. With its default settings it is about 2–2.5× faster, and it keeps your model, sampler and step count. The defaults lean toward quality; raise max_consecutive_forecasts for more speed.

• Two nodes: Manual, where you choose the forecast steps, and Adaptive, which is experimental and picks them from an error estimate.
• Shown with OpenMOSS's 50 steps of Euler + sgm_uniform and the template's 25 steps of er_sde + simple; other samplers may not run well.
• With skip-layer guidance, set its start and end to 0 and 1: partial ranges leave fewer steps to forecast.
• License: GPL-3.0.

OpenMOSS's user guide says: “We are actively working on few-step distillation methods such as DMD. Stay tuned.” As of Oct 22, 2026 it has not given a date.