MOSS-Image-Turbo

A fifteen-step acceleration LoRA for MOSS-Image-SFT, with merged BF16 and scaled FP8 diffusion-model exports. Created by Coffmeteor.

The accelerator was trained as a rank64 LoRA on the original SFT model. The merged files bake that same adapter into the diffusion model at strength1. They do not include skin, lighting, character or style adapters.

Files

File	Size	Use
MOSS Image Turbo LoRA.safetensors	0.595 GB	Original SFT + Turbo LoRA at1.0; FP32 adapter weights
MOSS Image Turbo BF16.safetensors	10.165 GB	Merged diffusion model
MOSS Image Turbo FP8.safetensors	5.408 GB	Scaled E4M3FN block weights; BF16 activations and other layers

Hashes and exact sizes: SHA256SUMS, release.json.

Recommended settings

• 15 steps, CFG3, Euler, sgm_uniform, flow shift 5.
• Negative-branch skip guidance: single layer22, start0, end1; no double layers.
• Turbo LoRA strength 1.0; no trigger word.
• Do not add the Turbo LoRA to either merged model: it is already included.

The original SFT + bypass LoRA path is the reference configuration. Merged models can change pose and facial details through finite-precision arithmetic. CFG1 is an optional lower-cost setting but was weaker on dense text; CFG3 was selected for the final mixed photography/structure/text evaluation.

ComfyUI setup

1. Install current ComfyUI and the official MOSS-Image ComfyUI extension.
2. Download the original MOSS-Image-SFT repository to ComfyUI/models/diffusers/MOSS-Image-SFT, including its transformer, text_encoder, tokenizer and vae folders. The merged variants reuse its text encoder, tokenizer and VAE.
3. Put the Turbo LoRA in ComfyUI/models/loras/moss-image/, or put a merged model in ComfyUI/models/diffusion_models/moss-image/. Keep the filenames above.
4. Open a matching UI workflow from workflows. Fifteen selects the original base plus the Turbo LoRA; BF16 and FP8 select the merged models. Start ComfyUI from its root or set the HF loader's model_path to your SFT folder.
5. Enter your prompt and seed. No optional custom skin or lighting adapter is required.

The LoRA workflow uses ComfyUI's LoraLoaderBypassModelOnly to preserve the reviewed adapter execution. The FP8 workflow intentionally selects bf16 in the MOSS loader: embedded per-layer quantization metadata keeps FP8 weights quantized while selecting BF16 activation/matrix computation. Do not force a blanket unscaled FP8 cast over this checkpoint.

Evaluation

The reference LoRA configuration was reviewed on 100 random SFW prompts across 15 resolutions and six aspect-ratio groups, plus dedicated structure, Chinese/ English text and character-card cases. Visible effects from independent Gouache, Psychedelic and Beauty content LoRAs were retained in the reviewed stacking cases. The three export comparisons below use identical prompts, seeds and settings.

Export precision and provenance

BF16 merging uses FP32 matrix multiplication and accumulation with TF32 disabled, followed by one final cast. FP8 is quantized directly from the FP32 merged weights, without an intermediate BF16 rounding. Normalization, modulation, input and output layers retain their original precision. Per-layer scales were checked against the high-precision merged weights; the final choice does not increase weight MSE relative to the initial export.

Base revision: 2c7954437269461767f533b006823e8f85973a60. Training used rank64, alpha64 and64 updates with teacher CFG3 and the actual fifteen-step inference grid. The original model and architecture are by OpenMOSS. This derivative is distributed under the included Apache2.0 license.

Full 100-image gallery

Browse all 100 original-resolution images, prompts and numbered previews.