# MOSS Aria-BF16
Cinematic atmosphere. Tactile detail. Images with a story to tell.
MOSS Aria is my full fine-tune of [OpenMOSS-Team/MOSS-Image-Base](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base), trained for 5,000 steps on an NVIDIA H200. I developed it around expressive subjects, coherent anatomy, layered environments, and the small textures and particles that give a scene its sense of depth.
The gallery reflects that range: intimate portraits, wildlife and macro nature, weathered science fiction, and playful fantasy. The goal is a frame that feels composed as a whole while still inviting a closer look.
## The BF16 release
This download is the original full fine-tune in Safetensors format. It contains BF16 transformer weights, with one small F32 auxiliary tensor. It is a complete fine-tune of the MOSS-Image transformer, not a LoRA. The text encoder and VAE are separate components and are not included in this file.
| Detail | Value |
| --- | --- |
| Base architecture | MOSS-Image |
| Base model | OpenMOSS-Team/MOSS-Image-Base |
| Training | Full fine-tune, 5,000 steps |
| Training hardware | NVIDIA H200 |
| Download | Safetensors, approximately 9.47 GiB |
| SHA-256 | 3f331a97fc184cec752ef0a7f1c46187726a907bc5ffcb4db8b19902282d9fc1 |
Use a workflow that supports MOSS-Image and provide the corresponding text encoder and VAE. The [official MOSS-Image-Base page](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base) links to its setup and ComfyUI guidance.
## About the gallery
I created the gallery images in Draw Things with the corresponding converted MOSS Aria model. They show the visual character of this training run. Because they were rendered with a Draw Things conversion, they are illustrative examples rather than direct BF16 inference tests. Generation prompts and settings are not listed here.
## Credit and change notice
MOSS Aria is derived from MOSS-Image-Base by OpenMOSS. The base model weights are released under [Apache License 2.0](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base/blob/main/LICENSE). The transformer weights in this release were modified through full fine-tuning. Please credit OpenMOSS for the base model.
Description
MOSS Aria V4 FP16 | Full Fine-Tune
MOSS Aria V4 is a fully fine-tuned version of MOSS Image Base, trained for 5,000 steps on 2,242 curated images.
This release focuses on improving visual coherence, cinematic lighting, atmospheric depth, rich textures, and artistic expression.
My goal was not simply to make technically perfect images, but to give the model a stronger visual identity, with more natural details, emotional presence, and a sense of storytelling.
What's new in V4:
Improved subject and scene coherence
Richer textures and finer details
More expressive cinematic lighting
Greater depth and atmospheric realism
More consistent anatomy and material rendering
Enhanced artistic character and visual storytelling
Technical details:
Base architecture: MOSS Image
Training: Full fine-tune, not a LoRA
Training steps: 5,000
Training dataset: 2,242 curated images
Precision: FP16 with numerical stabilization
Model type: Full checkpoint, not a Turbo or distilled model
Trigger words: None required
Generation notes:
The model works with standard MOSS Image workflows in ComfyUI.Recommended Generation Settings
Standard Quality (FP16 / INT8 ConvRot)
Steps: 30–50 (50+ for maximum detail)
CFG Scale: 3.0 (recommended range: 3.0–6.0)
Sampler: Euler
Scheduler: SGM Uniform
Flow Shift: 5.0
Skip Layer Guidance: Layer 22 (optional)
Resolution: 1024×1024 or higher, preferably 1–2 megapixels
For faster previews, you can also experiment with ER SDE + Simple, around 25–35 steps.
Fast Generation (Optional 8-Step LoRA)
For significantly faster generation, I recommend experimenting with a compatible MOSS Image 8-step acceleration LoRA.
Steps: 8 (10–12 can also be tested)
CFG Scale: 1.0
Sampler: Euler
Scheduler: Simple
Flow Shift: 5.0
Skip Layer Guidance: Disabled
Please note that MOSS Aria V4 is a full fine-tuned model, not a Turbo or distilled checkpoint. The acceleration LoRA is optional and is not included or merged into the model.
The INT8 ConvRot version is a quantized version of the full model, not an 8-step distillation. For standard generation, the same 30–50 step recommendations apply.
For the best evaluation of the model's native quality and detail, I recommend testing it without acceleration LoRAs first.
Feel free to experiment with different samplers, schedulers and CFG values. Results may vary depending on the prompt and workflow.