Instructions to use hinablue/1111-MOSSImage-Asian-Utopian with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use hinablue/1111-MOSSImage-Asian-Utopian with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("OpenMOSS-Team/MOSS-Image-Base", dtype=torch.bfloat16, device_map="cuda") pipe.load_lora_weights("hinablue/1111-MOSSImage-Asian-Utopian") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
1111 - MOSS Image Asian Utopian ยท R4
A MOSS-Image derivative by hinablue, focused on portraits of Asian women, with adjustments to anatomy, poses, facial expressions, and portrait styles. It combines LoRA fine-tuning and layer-wise weight merging to provide starting points for different portrait aesthetics while retaining the original model's flexibility.
This repository provides R4 full checkpoints in BF16 and INT8 ConvRot formats, along with standalone LoRAs for the original MOSS-Image-Base and MOSS-Image-SFT models, respectively. The original Civitai page currently features the UBW - UNLOCK version. The original training background and R4 checkpoint metadata are described separately below to avoid mixing information from different versions.
Model Files
R4 Full Checkpoints
| File | Format | Size | Recommended Use |
|---|---|---|---|
| 1111_hina_MOSSImage_R4_bf16.safetensors | BF16, with original F32 tensors preserved | 9.47 GiB | General use and compatibility-first setups |
| 1111_hina_MOSSImage_R4_int8-convrot.safetensors | INT8 ConvRot, mixed with BF16/F32 | 5.04 GiB | Loaders that explicitly support this ConvRot format |
Standalone LoRAs for the Original Base/SFT Models
| LoRA File | Corresponding Original Model | Size |
|---|---|---|
| hina_MOSSImageBase_asianMix_v3-R4_lora.safetensors | OpenMOSS-Team/MOSS-Image-Base | 302.91 MiB |
| hina_MOSSImageSFT_asianMix_v3-R4_lora.safetensors | OpenMOSS-Team/MOSS-Image-SFT | 302.93 MiB |
Both LoRAs use rank 64 and alpha 32 and are intended to be loaded with their corresponding original models. You do not need to download the R4 full checkpoints from this repository. A LoRA cannot generate images on its own; pair it with Base or SFT according to the table above.
How R4 Was Built
According to the metadata embedded in the BF16 checkpoint, R4 starts from the corrected UBW bf16-fix checkpoint. The Base and SFT R4 LoRAs are first fused using complementary layer-wise weights, then baked into the checkpoint at an external strength of 0.6:
R4 = UBW bf16-fix + 0.6 ร ฮ(fused Base/SFT R4 LoRA)
The LoRA delta matrix operations and additions are performed in FP32, then cast back to the original tensor precision. Unmodified tensors retain their original values. The INT8 ConvRot version is quantized from the R4 BF16 checkpoint provided in this repository.
Original Training Background
Approximately 8,000 curated samples were used, with training data prepared to address anatomy, movement, facial expressions, and related aspects. LoRA fine-tuning was performed on the Base model, followed by layer-wise weight merging back into the model. The detailed training section specifies 7,780 samples, a batch size of 2, and 4 epochs.
The training tool was KohakuMossTrainer.
R4 Training Background
R4 was retrained using a new set of 1,105 images, with detailed facial descriptions covering four countries, resulting in a total of 4,420 training samples. The training tool was KohakuMossTrainer.
Usage
Included ComfyUI Workflow
Download MOSS Image R4 with Lora.json and drag it into ComfyUI, or open it through the workflow menu.
This workflow uses the R4 INT8 ConvRot checkpoint with an 8-step acceleration LoRA. It is not an example of applying the standalone portrait LoRAs to the original Base/SFT models.
| Component | File | ComfyUI Directory |
|---|---|---|
| Diffusion model | 1111_hina_MOSSImage_R4_int8-convrot.safetensors |
models/diffusion_models/ |
| Acceleration LoRA | MOSS-Image-Base_8-steps_adapter_lora.safetensors (upstream LoRA) |
models/loras/ |
| Text encoder | qwen3.5_9b_w4a8_gptq.safetensors | models/text_encoders/ |
| VAE | flux2-vae_small.safetensors | models/vae/ |
The acceleration LoRA filename above is the local name saved in the workflow; the upstream file is named lora.safetensors. Select your downloaded file in LoraLoaderModelOnly if its name differs. Also reselect the R4 checkpoint, text encoder, and VAE in their loader nodes. Use a ComfyUI installation that supports INT8 ConvRot and the GPTQ text encoder.
Resolve missing custom nodes before running. The workflow includes ResolutionSelector, SeedNode, and post-processing nodes from comfyui_layerstyle, comfyui-kjnodes, and Atomic-Trash/ComfyUI-ColoristFilmGrain.
| Setting | Saved Workflow Preset |
|---|---|
| Steps / CFG | 8 / 1 |
| Sampler / scheduler | er_sde / simple |
| Acceleration LoRA strength | 1.0 |
| Resolution selector | 3:4 portrait, 1.6 megapixels |
SkipLayerGuidanceDiTSimple |
Bypassed |
These are the workflow's accelerated settings, separate from the non-accelerated starting settings below.
R4 Full Checkpoints (ComfyUI)
- Use a ComfyUI version with MOSS-Image support and follow the upstream installation and workflow guide.
- Place the R4 BF16 file in
ComfyUI/models/diffusion_models/and select it in Load Diffusion Model. - Prepare a compatible text encoder and VAE according to the workflows in Comfy-Org/MOSS-Image. Set the text encoder's CLIP Loader type to
moss_image. - Start without acceleration, then add a compatible acceleration LoRA if needed. Before using the INT8 version, confirm that your loader supports INT8 ConvRot.
Native ComfyUI and the older HF directory plugin use different loader nodes and timestep settings. Use the matching workflow rather than carrying over the old plugin's scheduling nodes unchanged.
Original Base/SFT + LoRA (ComfyUI)
- Prepare the original MOSS-Image-Base or MOSS-Image-SFT model, along with the text encoder and VAE required by its workflow.
- Place the corresponding LoRA file from the table above in
ComfyUI/models/loras/. - After loading the original model, insert a compatible LoRA loader node between the model loader and sampler. Select the matching Base or SFT LoRA, then pass the resulting model to the sampler.
- Adjust the LoRA strength to suit your prompt and results. The merge strength used to build R4 is not a recommended inference strength for these standalone LoRAs.
Suggested Starting Settings
The resolution follows Civitai's recommendation. The remaining settings use the non-accelerated defaults from upstream MOSS-Image-Base as starting points, not experimentally established optimal settings for R4.
| Setting | Starting Value |
|---|---|
| Resolution | Approximately 1.2โ1.6 megapixels, such as 1024 ร 1280 or 1024 ร 1536 |
| Sampler | Euler (flow matching) |
| Steps | 50 |
| CFG | 3 |
| Shift | 5 |
| Negative prompt | Empty |
| Layer skipping | Negative branch, layer index 22 (zero-based) |
Acceleration LoRAs
The R4 full checkpoints do not have an 8-step or 4-step acceleration LoRA baked in, preserving flexibility in how they are used and configured. Choose a compatible acceleration LoRA for your workflow and follow its recommended step count and guidance settings rather than simply lowering the step count in the table above.
DMD LoRAs recommended on the original Civitai page:
Prompts
The trigger words listed on the original Civitai page are Asian woman and air. Use natural language to describe the subject, makeup, clothing, lighting, and setting. The included caption_templates.txt also provides Japanese, Korean, Taiwanese, and Thai style descriptions.
The following is a starting prompt, not the complete generation parameters for the showcase images below:
Asian woman, an adult Taiwanese woman with soft oval facial features,
natural makeup and airy curtain bangs, wearing a cream linen blouse,
seated beside a window in a quiet cafe, soft natural daylight,
realistic skin texture, relaxed expression, portrait photography.
Style and nationality terms are creative prompts, not fixed representations of any ethnic group.
R4 Example Images
License and Usage Notes
Released under the Apache License 2.0, as a derivative of upstream weights also licensed under Apache-2.0. When redistributing, retain applicable upstream licenses and notices and state your modifications.
Changes in this release include portrait-focused LoRA fine-tuning and merging, as well as an INT8 ConvRot quantized version. Review generated images for anatomical errors, detail issues, and potential biases, and respect likeness rights, privacy, and applicable laws.
Sources and Acknowledgments
- Downloads last month
- -
Model tree for hinablue/1111-MOSSImage-Asian-Utopian
Base model
OpenMOSS-Team/MOSS-Image-Base



