Do not reuse the standard workflow unchanged
Four-step, eight-step DMD, and fifteen-step adapters publish different samplers, guidance, and loading requirements.
Loading a few-step LoRA into a 25–50 step, CFG 2.5–3 graph can overcook or smear an adapter trained for four steps and CFG 1. The eight-step DMD release uses no classifier-free guidance and a specialized sampling path, while the verified fifteen-step adapter recommends CFG 3. Start from the exact maintainer instructions for the downloaded release.
4-step, 8-step, and 15-step recipes
| Adapter | Published starting recipe | Important note |
|---|---|---|
| akhaliq four-step v2 | 4 steps, CFG 1, stock flow-matching Euler, LoRA strength 1 | v2 changed the usage from earlier releases; above roughly one megapixel, the author suggests testing 6–8 steps. |
| lodestones eight-step DMD | 8 network evaluations, no CFG; LoRA strength 1, or 0.7 for the latest named file | Work in progress with a data-space DMD sampling method. Do not treat it as a normal Euler drop-in. |
| Coffmeteor fifteen-step | 15 steps, CFG 3, Euler, sgm_uniform, shift 5, LoRA strength 1 | Uses negative-branch skip guidance on single layer 22. Do not add the LoRA to a merged model that already contains it. |
These are community adapters. Their maintainers can update files and instructions, so recheck the source page before downloading.
Connect the LoRA without losing your baseline
- First save a working standard SFT graph and fixed-seed output.
- Add the LoRA with a model-only loader or use the maintainer's matching merged file.
- Apply the exact recipe for that release.
- Keep prompt, seed, dimensions, and VAE fixed for the comparison.
- Label saved outputs with adapter version, strength, and settings.
Do not compare a native workflow against an HF-directory graph by swapping only the loader. Establish the normal SFT result first, then add acceleration as the single changed variable.
Compare speed and quality honestly
Use the same prompt, seed, dimensions, and output review criteria for standard 50-step SFT and each accelerator. Record wall-clock time on your own hardware instead of publishing a universal speed multiplier. Score text readability, composition, prompt adherence, and the number of retries required to reach a usable image.
When the faster result is not the useful result
Few-step generation is valuable for drafts, prompt exploration, and composition searches. Dense typography, small labels, unusual aspect ratios, and final delivery images deserve a comparison against standard SFT. Proofread all generated text and keep failed samples: they reveal whether the speed tradeoff fits your actual work.
Adapter sources
Recipes checked October 27, 2026.
Four-step adapter model cardEight-step DMD model cardFifteen-step adapter model card