axiomofmind/Moss-image-animated-lora

MOSS Image Animated Character LoRA

A broad animated-character LoRA for MOSS-Image. It is intended to add a polished, dimensional cartoon treatment across people, animals, action scenes, and varied environments rather than reproduce one narrowly defined visual style.

The adapter was trained locally against MOSS-Image-Base. Local checkpoint evaluation used MOSS-Image-SFT as the underlying inference model.

Checkpoint and prompting

• Published checkpoint: 2,000 steps
• Registered trigger: moss_ac_style
• Tested LoRA scales: 0.7, 0.9, and 1.0
• Tested category wording: prompts both with and without 3d animated character
• Recommended starting point pending final sample selection: MOSS-Image-SFT, scale 0.7, 50 steps, guidance 3.0

When using the trigger, place it at the beginning of the prompt:

moss_ac_style, an elderly woman beekeeper lifting a honeycomb frame in a rooftop garden above a busy city, bees glowing in late-afternoon sunlight

The trigger and the natural phrase 3d animated character were evaluated independently. Their usefulness can vary by subject and desired strength, so they are not both required.

ComfyUI compatibility

The root pytorch_lora_weights.safetensors uses Diffusers / PEFT keys such as transformer.layers.0.proj.q.lora_A.weight. ComfyUI's native MOSS-Image loader addresses the same modules under diffusion_model.layers..., so loading the root file directly can report lora key not loaded.

Use the native ComfyUI variant for the canonical step-2000 checkpoint:

File	Size	Tensors	SHA-256
MOSS-Image-Animated-LoRA-ComfyUI.safetensors	70.93 MiB	504	A99173F545271DCD43BBB79483C60FECEAB447B80841D2391936A60672E41934

1. Update ComfyUI to a version with native MOSS-Image support; commit b1a247c or newer contains the required loader code.
2. Put MOSS-Image-Animated-LoRA-ComfyUI.safetensors in ComfyUI/models/loras/.
3. Use ComfyUI's standard LoRA loader with model strength 0.7 and clip strength 0.0.
4. Use natural subject and scene descriptions. The optional 3d animated character wording and moss_ac_style trigger were tested separately and together; when using the trigger, place it at the beginning of the prompt. Scales 0.9 and 1.0 provide stronger alternatives.

Download only this variant with:

hf download axiomofmind/Moss-image-animated-lora \
  MOSS-Image-Animated-LoRA-ComfyUI.safetensors \
  --local-dir ./ComfyUI/models/loras

The conversion is namespace-only: transformer. becomes diffusion_model.. Tensor values, shapes, dtypes, rank, and alpha behavior are unchanged. All 504 converted tensors were checked against the 252 corresponding MOSS-Image-SFT modules. The reproducible converter is scripts/convert_moss_lora_for_comfyui.py.

This ComfyUI file corresponds to the canonical step-2000 root weight.

Evaluation samples

Prompt wording and trigger comparison

The first comparison keeps the content prompt and seed fixed while changing only the natural category wording and registered trigger. All three images use checkpoint 2000 with MOSS-Image-SFT, LoRA scale 0.7, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 5555.

Scene	Without phrase · no trigger	Without phrase · triggered	With phrase · triggered
Elderly woman archaeologist

Base content prompt:

an elderly woman archaeologist brushing sand from a colorful mosaic at a desert excavation, canvas tents and rocky cliffs behind her

The triggered prompts prepend moss_ac_style,. The with-phrase prompt additionally begins its content with a 3d animated character of.

Stock SFT and trigger comparison

This comparison keeps the scene and seed fixed while showing stock SFT alongside wording and trigger variants. The LoRA columns use checkpoint 2000 at scale 0.7; all four images use MOSS-Image-SFT, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 4444.

Scene	Stock SFT	Without phrase · no trigger	With phrase · no trigger	With phrase · triggered
Adult male park ranger

Content prompt:

a 3d animated character of an adult male park ranger with a shaved head lifting a baby owl from a fallen tree in a misty forest, soft morning light through the branches

The triggered LoRA prompt prepends moss_ac_style,. Stock SFT and the with-phrase no-trigger LoRA use the content prompt exactly as shown. The without-phrase LoRA prompt removes a 3d animated character of.

Stock SFT and scale-0.9 wording comparison

This comparison shows stock SFT and triggered checkpoint-2000 LoRA output in both wording states. All four images use MOSS-Image-SFT, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 9999. The LoRA images use scale 0.9.

Scene	Stock SFT · with phrase	With phrase · triggered	Stock SFT · without phrase	Without phrase · triggered
Frog postal worker

Base content prompt:

a tiny moss-green frog postal worker steering a paper boat through a flooded cobblestone street, holding a mail satchel above lantern reflections

The with-phrase prompts prepend a 3d animated character of to the content. The triggered prompts additionally prepend moss_ac_style,.

Stock SFT and triggered scale-0.7 comparison

Both images use MOSS-Image-SFT, the same category wording, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 7777. The LoRA image uses checkpoint 2000 at scale 0.7 with moss_ac_style.

Scene	Stock SFT · with phrase	With phrase · triggered, scale 0.7
Adult man storm chaser

Content prompt:

a 3d animated character of an adult man storm chaser with close-cropped hair anchoring a weather instrument beside a yellow truck as lightning forks over open plains, dramatic dusk light

The triggered LoRA prompt prepends moss_ac_style,.

Elderly chef stock and scale-0.7 comparison

Both images use MOSS-Image-SFT, the same category wording, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 2222. The LoRA image uses checkpoint 2000 at scale 0.7 with moss_ac_style.

Scene	Stock SFT · with phrase	With phrase · triggered, scale 0.7
Elderly East Asian chef

Content prompt:

a 3d animated character of an elderly East Asian chef carrying a steaming stockpot through a busy restaurant kitchen, copper pans overhead, warm evening light

The triggered LoRA prompt prepends moss_ac_style,.

Mars astronaut stock and scale-1 comparison

All three images use MOSS-Image-SFT without the natural category phrase, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 707. Both LoRA images use checkpoint 2000 at scale 1.0; only the final column adds moss_ac_style.

Scene	Stock SFT · without phrase	Without phrase · no trigger, scale 1.0	Without phrase · triggered, scale 1.0
Adult woman astronaut on Mars

Content prompt:

an adult woman astronaut in a weathered white suit planting a green seedling inside a glass dome on Mars at dawn, wide shot

The triggered LoRA prompt prepends moss_ac_style,.

Clockmaker stock and scale-1 comparison

Both images use MOSS-Image-SFT without the natural category phrase, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 808. The LoRA image uses checkpoint 2000 at scale 1.0 without moss_ac_style.

Scene	Stock SFT · without phrase	Without phrase · no trigger, scale 1.0
Elderly clockmaker

Content prompt:

an elderly clockmaker with a silver moustache reaching into a giant wall clock in a cluttered workshop, warm window light, full-body shot

The content prompt is identical for stock SFT and the LoRA image.

Robot musician stock and scale-1 comparison

Both images use MOSS-Image-SFT without the natural category phrase, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 909. The LoRA image uses checkpoint 2000 at scale 1.0 without moss_ac_style.

Scene	Stock SFT · without phrase	Without phrase · no trigger, scale 1.0
Orange robot street musician

Content prompt:

a small orange robot street musician playing a blue accordion beneath a rain awning in a bustling night market, colorful reflections on wet pavement

The content prompt is identical for stock SFT and the LoRA image.

Wildfire lookout stock and scale-0.7 comparison

Both images use MOSS-Image-SFT with the natural category phrase, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 1111. The LoRA image uses checkpoint 2000 at scale 0.7 without moss_ac_style.

Scene	Stock SFT · with phrase	With phrase · no trigger, scale 0.7
Adult woman wildfire lookout

Content prompt:

a 3d animated character of an adult woman wildfire lookout whose bright auburn curls blow in the wind as she raises binoculars from a wooden tower above a pine valley during a smoky orange sunset, dressed in a green jacket

The content prompt is identical for stock SFT and the LoRA image.

Lavender penguin stock and scale-0.7 comparison

Both images use MOSS-Image-SFT with the natural category phrase, 50 inference steps, guidance 3.0, 1024 × 1024 resolution, an empty negative prompt, and seed 6666. The LoRA image uses checkpoint 2000 at scale 0.7 with moss_ac_style.

Scene	Stock SFT · with phrase	With phrase · triggered, scale 0.7
Lavender penguin librarian

Content prompt:

a 3d animated character of a plump lavender penguin librarian pushing a rolling ladder through an enormous circular library while paper airplanes drift between the shelves

The triggered LoRA prompt prepends moss_ac_style,.

Recommended settings

• Underlying model: OpenMOSS-Team/MOSS-Image-SFT
• Inference steps: 50
• Guidance scale: 3.0
• Negative prompt: empty
• Typical resolution: 1024 × 1024
• LoRA scale: begin at 0.7; increase toward 1.0 for a stronger learned treatment

Loading with Diffusers and PEFT

MOSS-Image uses a custom Diffusers pipeline. Attach the adapter directly to the transformer with PEFT:

import json

import torch
from diffusers import DiffusionPipeline
from huggingface_hub import hf_hub_download
from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
from peft.tuners.lora.layer import LoraLayer
from safetensors.torch import load_file

repo_id = "axiomofmind/Moss-image-animated-lora"
lora_scale = 0.7

config_path = hf_hub_download(repo_id, "adapter_config.json")
weights_path = hf_hub_download(repo_id, "pytorch_lora_weights.safetensors")

with open(config_path, encoding="utf-8") as file:
    config = json.load(file)

pipe = DiffusionPipeline.from_pretrained(
    "OpenMOSS-Team/MOSS-Image-SFT",
    custom_pipeline="pipeline_moss_image",
    trust_remote_code=True,
    torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
)
pipe.scheduler = type(pipe.scheduler).from_config(pipe.scheduler.config)

inject_adapter_in_model(
    LoraConfig(
        r=config["r"],
        lora_alpha=config["lora_alpha"],
        target_modules=config["target_modules"],
    ),
    pipe.transformer,
)

state = {
    key.removeprefix("transformer."): value
    for key, value in load_file(weights_path, device="cpu").items()
}
_, unexpected = set_peft_model_state_dict(pipe.transformer, state)
if unexpected:
    raise RuntimeError(f"Unexpected LoRA tensors: {unexpected[:5]}")

for layer in pipe.transformer.modules():
    if isinstance(layer, LoraLayer):
        layer.set_scale("default", lora_scale)

pipe.transformer.to(torch.bfloat16)
pipe.to("cuda")

prompt = (
    "moss_ac_style, an adult woman explorer crossing a rope bridge "
    "above a misty jungle canyon at sunrise"
)

image = pipe(
    prompt,
    negative_prompt="",
    width=1024,
    height=1024,
    num_inference_steps=50,
    guidance_scale=3.0,
    uncond_skip_layers=[22],
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]

image.save("moss_animated_character.png")

Training details

• Training base model: OpenMOSS-Team/MOSS-Image-Base
• Training images: 113
• Resolution: 1024 with aspect-ratio buckets
• Total training steps: 2,000
• LoRA rank / alpha: 16 / 16
• Target modules: transformer attention q/k/v/o projections and feed-forward up/down projections
• Optimizer: AdamW
• Learning rate: 1e-4
• Scheduler: constant with warmup
• Batch size / gradient accumulation: 1 / 1
• Weight decay: 0.01
• Caption dropout: 0
• Random horizontal flip: disabled
• Training seed: 42
• Training environment: local hardware

Intended use and limitations

This adapter is intended for lawful creative image generation involving fictional animated characters and environments. It may produce malformed anatomy, hands, faces, text, mechanical objects, or multi-character relationships. Results vary with prompt construction, seed, adapter scale, aspect ratio, and the underlying MOSS model.

License

Apache 2.0, matching MOSS-Image. This is an independent community LoRA and is not an official OpenMOSS release.