THE MOSS-IMAGE FIELD NOTES
MOSS-Image LoRA and turbo guide: compatibility and ComfyUI
Understand MOSS-Image LoRAs, community 8-step DMD adapters and turbo terminology. Check the base checkpoint, precision, strength and workflow before loading an adapter.
Updated October 24, 2026 · Independent guide · Official sources linked belowA MOSS-Image LoRA is an adapter loaded alongside a compatible base checkpoint, not a replacement for the complete model. Some adapters change appearance or a specific artifact; others distill sampling into fewer evaluations. For a community “turbo” or 8-step model, use the author’s matched files and sampling instructions rather than only lowering the step count.
What does “MOSS Image turbo LoRA” mean?
“Turbo” in a community listing is a label to investigate, not enough information to select a sampler. OpenMOSS publishes Base and SFT; community acceleration projects must be identified by their actual repository, checkpoint and instructions. A regular style LoRA and a sampling-distillation adapter solve different problems.
| Adapter purpose | What to check | What not to assume |
|---|---|---|
| Style or subject adaptation | Training checkpoint, target modules and strength. | That it reduces the required number of steps. |
| Artifact control | Artifact being targeted and changes to texture or detail. | That removing one artifact improves every other quality. |
| Few-step distillation | The exact model, precision, schedule, guidance and adapter file. | That setting an ordinary SFT graph to eight steps reproduces the author’s method. |
Choose your starting checkpoint with the Base vs SFT comparison. The OpenMOSS community list links projects with different purposes; it does not turn their settings into one shared recipe.
Check compatibility before downloading an adapter
- Checkpoint: confirm Base versus SFT and the intended revision.
- Format: distinguish a LoRA file from a full transformer checkpoint or extra sampling weights.
- Precision: check the author’s guidance before using BF16, quantized or rescaled FP16 weights.
- Loader: confirm whether the file works with the built-in ComfyUI model LoRA loader or needs project-specific code.
- Sampling: record steps or evaluations, CFG/guidance, scheduler, shift and any layer skipping.
- Strength: use the value for that exact adapter filename, not another revision’s value.
- Source and license: keep the author’s model card and revision alongside your saved graph.
Start with the working ComfyUI baseline. If an adapter breaks that result, remove it before changing the prompt or VAE.
Load a compatible LoRA in ComfyUI
ComfyUI/models/loras/
your-compatible-moss-adapter.safetensors- Download the actual adapter from its author and place it in
models/loras/. - Refresh model discovery or restart ComfyUI.
- Add the matching model LoRA loader to your working graph and select the file. A model-only loader can be used when the adapter targets diffusion weights only; follow the author if text-encoder weights are also involved.
- Apply the adapter to the loaded model before the model-patching and sampling path, following the matched graph.
- Use the exact strength and sampling instructions from the adapter card. Save this as a separate workflow.
See the native node flow for the baseline and workflow downloads for native versus HF-plugin graphs. A LoRA does not replace the text encoder, VAE or missing loader nodes.
Example: the community MOSS-Image SFT 8-step DMD adapter
lodestones/MOSS-Image-SFT-8-steps-dmd describes a work-in-progress data-space distillation of SFT. Its current usage section links two LoRA files with different suggested strengths:
| Author-provided file | Suggested strength on the checked card | Checkpoint family |
|---|---|---|
| lora.safetensors | 1.0 | MOSS-Image SFT |
| lora_latest.safetensors | 0.7 | MOSS-Image SFT |
Checked October 24, 2026. Filenames ending in “latest” can change. Recheck the author’s usage section before running and record the file revision you downloaded. The card labels its older full-transformer implementation section as deprecated; do not combine that old precision recipe with the current LoRA instructions by accident.
The author describes an eight-evaluation method without CFG. This page has not independently GPU-tested that configuration and does not provide a guessed KSampler preset. Follow the author’s current complete workflow and runtime requirements. A service calling itself “8-step” may use a different sampler than the ordinary Base/SFT reference.
Compare speed and quality with the same brief
Keep the canvas, prompt and evaluation criteria stable. Record the checkpoint, adapter revision, precision, strength, sampler and total runtime for every attempt. Compare the exact lettering, layout, detail and failure rate as well as elapsed time. Fewer evaluations do not directly guarantee a fixed wall-clock speedup on your hardware.
Use the menu prompts to check exact item-price pairs or the two layout inputs to inspect composition. The hosted examples use ordinary SFT through the API; they are not evidence of this community adapter’s quality or speed.
Does this studio accept custom LoRAs?
The online studio uses its configured SFT API model and does not expose a custom-LoRA upload or a turbo switch. For adapter experiments, follow the local ComfyUI guide. To generate with the hosted model, open the studio.
Original sources
Adapter author’s usage and revision history, OpenMOSS loader and sampling notes, model prompting and training guide. Return to model selection when the required checkpoint or file format is unclear.