DMD2 Models

Some community NPU models in the model collections have DMD2 LoRA merged in. They trade a small amount of image quality for a large speedup and much less device heat.

Identifying DMD2 Models

A model with _dmd2 in its filename has DMD2 LoRA merged into the checkpoint. For example:

anythingxl_dmd2_qnn2.28_8gen3.zip

If the suffix is not there, treat the model as a regular SDXL/SD1.5 checkpoint and use its normal recommended settings.

Recommended Parameters

DMD2 models are tuned for very low step counts and CFG 1. Use:

Parameter	Value
Scheduler	LCM
CFG	1
Steps	8

DMD2 will not behave correctly with higher CFG values or with DPM++ / Euler schedulers — use the combination above.

Why It's Fast

Two effects compound:

1. Fewer steps — 8 instead of the typical 20–30 means several times fewer UNet passes.
2. CFG = 1 optimization — on the NPU path, CFG exactly equal to 1 skips the unconditional UNet pass, halving the per-step compute (see CFG note).

In practice, a DMD2 model ends up around 4× faster end-to-end than the same base model with typical settings. Exact speedup depends on the comparison baseline (step count, CFG, and whether the base model also uses CFG 1).

Real-World Numbers (SDXL)

• On a 16 GB+ device with Low RAM Mode turned off, an SDXL DMD2 generation typically completes in ~6–7 s per image.
• Because each generation is short, the device stays cool — you can generate many images in a row without throttling.

Quality Tradeoff

DMD2 output is roughly 80–90% of the original model's quality at default settings. The drop is usually visible on fine textures and small details rather than overall composition.

A practical workflow when you want both speed and quality:

1. Generate candidates fast with the DMD2 model.
2. Pick the one you like.
3. Run img2img with the original (non-DMD2) model at a moderate denoise strength to refine details while keeping the composition.

This is similar in spirit to the SD1.5 Highres.fix workflow — DMD2 produces the structure, the full model finishes the detail.

When to Pick DMD2

• You want to iterate quickly through prompts and seeds.
• You want to generate many images in a session without the device heating up.
• You are on a device where SDXL is borderline-usable and the full model is uncomfortably slow.

If you specifically need maximum quality on a single image, use the non-DMD2 model directly, or finish with an img2img refinement pass as above.