AnimeLoom

Open-source character LoRAs and tooling for face-consistent anime video generation.

AnimeLoom is a text-to-anime-video pipeline focused on the hardest part of long-form anime generation: keeping the same character's face the same across every shot.

This organization hosts the character LoRAs used by the AnimeLoom pipeline on GitHub.

---

Character LoRAs

Every LoRA ships two adapters in one repo:

• sdxl/ — for Animagine XL 3.1, used by AnimeLoom's identity-keyframe stage
• sd15/ — for DreamShaper 8, for lightweight AnimateDiff workflows

All adapters are LoRA rank 32, trained with PEFT.

---

The pipeline

AnimeLoom is a Director-Orchestrated pipeline that turns a text story into a multi-shot anime video with consistent character identity.

text story
   ↓
Story Decomposer (Gemini → Claude)        →  shot script
   ↓
SDXL + character LoRA + IP-Adapter        →  identity keyframe per shot   (Phase 2)
   ↓
Wan2.2 I2V                                →  motion driving clip          (Phase 3a)
   ↓
Wan2.2-Animate face-lock                  →  face from keyframe pasted
                                              onto motion clip            (Phase 3b)
   ↓
RIFE + Real-ESRGAN + GFPGAN               →  temporal/spatial upscale,
                                              anime face restore           (Phase 4)
   ↓
Cross-dissolve assembly                   →  final 24fps video

Why face-lock matters

The "Animate" stage at Phase 3b is the moat. Per community testing, Wan2.2-Animate-14B achieves an 89% usable rate versus AnimateDiff's 71% and commercial APIs' ~78%. Combined with character-specific LoRAs trained for the SDXL identity stage, AnimeLoom can keep a character recognizably "themselves" across every shot of a multi-shot scene — the part most open-source pipelines struggle with.

---

Quick start

Train your own character LoRA, or use one from this org:

import torch
from diffusers import StableDiffusionXLPipeline
from peft import PeftModel
pipe = StableDiffusionXLPipeline.from_pretrained(
    "cagliostrolab/animagine-xl-3.1",
    torch_dtype=torch.float16,
).to("cuda")
# Load any AnimeLoom character LoRA
pipe.unet = PeftModel.from_pretrained(
    pipe.unet,
    "AnimeLoom/sakura-haruno",   # or AnimeLoom/denji, AnimeLoom/yuki-nagato
    subfolder="sdxl",
)
img = pipe(
    "1girl, sakura haruno, pink hair, cherry blossom forest, anime, masterpiece",
    num_inference_steps=28,
    guidance_scale=6.5,
    height=1024, width=1024,
).images[0]
img.save("out.png")

For the full text-to-video pipeline, see the AnimeLoom RunPod notebook.

---

License

All character LoRAs in this org are released under OpenRAIL++, inheriting from their base models (Animagine XL 3.1 / DreamShaper 8). The AnimeLoom pipeline code on GitHub is apache-2.0 licensed.

---

Contributing

Want to add a character? The training script is in agents/character/trainer.py. Recommended dataset: 15-30 clean anime reference images of the character at 512 or 1024 resolution.

Issues, PRs, and new character requests: GitHub Issues.

---

Links

• 🐙 GitHub: JoelJohnsonThomas/AnimeLoom
• 📦 Pipeline notebook: AnimeLoom_RunPod.ipynb
• 🤗 All models: huggingface.co/AnimeLoom