Known-good reference settings
| Configuration | Steps / CFG | Sampler details | Use it for |
|---|---|---|---|
| Base standalone | At least 50; CFG 3 default, 3–6 recommended | Use the official user guide or matching pipeline defaults | Foundation-checkpoint evaluation and custom tuning |
| Official HF-directory workflow | 50 / 3 | Euler, sgm_uniform, shift 5; SLG single layer 22 | A repeatable plugin-path test |
| Native ComfyUI workflow | Preserve the values inside the downloaded graph | Use native loaders and its matching model components | Recommended new ComfyUI installations |
| Turbo or few-step LoRA | Adapter-specific | Follow that adapter's model card exactly | Faster drafts after the standard workflow works |
There is no single universal Turbo row: a four-step adapter and a fifteen-step adapter publish different CFG and scheduler recipes.
Why Base can look weak with a short prompt
The official Base guide recommends CFG 3–6 and says shorter prompts may benefit from higher values. Start at 3, keep the seed fixed, and move upward in small increments. Higher guidance is not automatically better: stop when contrast and prompt adherence improve without harsh or brittle detail.
For a deeper explanation of Base prompts, sampling, and model selection, use the complete Base checkpoint guide.
Why a local run can produce only noise
The first check is workflow alignment. Native ComfyUI and the optional MOSSImageHFLoader plugin use different loader paths, and the official guide says their workflows are not interchangeable. Import the complete workflow for your chosen path instead of swapping only the loader.
- Confirm the checkpoint is Base or SFT and record its exact source.
- Confirm the graph is native or HF-directory.
- Restore the matching workflow's saved settings.
- Run one 1024 × 1024 image with a fixed seed before adding LoRAs or quantization.
Blur, previews, and sudden regressions
Open the saved image before judging a low-resolution ComfyUI preview. If previously working output changes after an update, test the last saved graph in a clean environment and inspect the console for missing nodes, precision warnings, or NaN values. Do not compensate for a broken environment by randomly increasing CFG or steps.
For official plugin defaults, negative-branch skip guidance uses single layer 22. Preserve that graph while diagnosing; disable it only as a controlled comparison.
Sampler, shift, SLG, and the VAE
The official Base model card publishes flow-matching Euler, 50 steps, CFG 3, shift 5, and negative-branch layer skipping at index 22 as its starting configuration. The bundled pipeline uses the FLUX.2 small decoder, with the transformer and text encoder in BF16 and the VAE in FP32.
Treat these values as a matched system. If you change the sampler, scheduler, shift, VAE, or layer guidance while also changing a checkpoint, you lose the ability to identify what caused the result. Restore a complete known-good graph before experimenting.
Precision, memory, and FP16
File size is not runtime VRAM. Resolution, offloading, component placement, and precision all change the peak. If plain FP16 produces overflow or NaN behavior in your setup, a community FP16-fixed transformer is available, but it is not an official OpenMOSS checkpoint. Verify its model card, hashes, and fixed-seed output before replacing the official weights.
Sources
Official defaults and community variants checked October 27, 2026.
Official Base user guideOfficial ComfyUI guideCommunity FP16-fixed model card