---
base_model: cagliostrolab/animagine-xl-3.1
library_name: peft
license: openrail++
pipeline_tag: text-to-image
tags:
  - stable-diffusion-xl
  - stable-diffusion
  - lora
  - peft
  - anime
  - character-lora
  - animeloom
  - denji
  - chainsaw-man
instance_prompt: "1boy, denji, chainsaw man, blonde hair, yellow eyes"
widget:
  - text: "1boy, denji, chainsaw man, blonde hair, yellow eyes, anime, masterpiece, best quality"
---

# AnimeLoom — Denji LoRA

Character LoRA for **Denji** (Chainsaw Man), trained as part of the
[AnimeLoom](https://github.com/JoelJohnsonThomas/AnimeLoom) anime
character-consistency pipeline.

Two adapters are provided in this repo:

| Folder  | Base model                       | Rank | Steps | Images | Resolution |
|---------|----------------------------------|------|-------|--------|------------|
| `sdxl/` | `cagliostrolab/animagine-xl-3.1` | 32   | 2200  | 15     | 1024       |
| `sd15/` | `Lykon/dreamshaper-8`            | 32   | 800   | 15     | 512        |

The SDXL adapter is the primary one used by AnimeLoom's identity-keyframe stage
(SDXL + character LoRA + IP-Adapter). The SD 1.5 adapter is provided for
compatibility with lighter inference pipelines (e.g. AnimateDiff workflows).

## Trigger words

```
1boy, denji, chainsaw man, blonde hair, yellow eyes
```

Add booru-style descriptors as needed (e.g. `black suit`, `chainsaw`,
`shark teeth`, `smiling`).

## Usage — SDXL with PEFT / diffusers

```python
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")

# AnimeLoom's training output is a PEFT adapter — load via PeftModel
pipe.unet = PeftModel.from_pretrained(
    pipe.unet,
    "AnimeLoom/denji",
    subfolder="sdxl",
)

img = pipe(
    "1boy, denji, chainsaw man, blonde hair, yellow eyes, anime, "
    "masterpiece, best quality, absurdres",
    negative_prompt="blurry, low quality, deformed, extra fingers, 3d render",
    num_inference_steps=28,
    guidance_scale=6.5,
    height=1024, width=1024,
).images[0]
img.save("denji.png")
```

## Usage — SD 1.5 with diffusers

```python
import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "Lykon/dreamshaper-8",
    torch_dtype=torch.float16,
).to("cuda")

pipe.load_lora_weights(
    "AnimeLoom/denji",
    subfolder="sd15",
    weight_name="pytorch_lora_weights.safetensors",
)

img = pipe(
    "1boy, denji, chainsaw man, blonde hair, anime, masterpiece",
    num_inference_steps=28,
    guidance_scale=7.0,
).images[0]
```

## Recommended weights & prompting

- **LoRA scale**: `0.85 - 1.15` (SDXL), `0.6 - 0.8` (SD 1.5)
- Pair with anime base models: Animagine XL 3.1, Counterfeit-V3.0, AnythingV5
- For face consistency in video, combine with **IP-Adapter SDXL** and
  **Wan2.2-Animate** face-lock — see the AnimeLoom pipeline.

## AnimeLoom video pipeline integration

This LoRA is built to feed AnimeLoom's text-to-anime-video pipeline:

```
SDXL + this LoRA + IP-Adapter   →  identity keyframe (Phase 2)
        ↓
Wan2.2 I2V                       →  motion driving clip (Phase 3a)
        ↓
Wan2.2-Animate face-lock         →  face from keyframe pasted onto motion (Phase 3b)
        ↓
RIFE + Real-ESRGAN + GFPGAN      →  temporal/spatial upscale + face restore (Phase 4)
        ↓
Final 24fps anime video with consistent character identity across shots.
```

## Limitations

- **Anime-only.** Photoreal prompts will degrade quality.
- **InsightFace face-swap does not work on these outputs** (it is photoreal-only).
  For identity rescue use IP-Adapter-FaceID-SDXL or
  [CharacterFaceSwap](https://github.com/ArtBot2023/CharacterFaceSwap).
- Trained on 15 images; expect occasional drift on unusual angles, complex
  outfits, or transformation/hybrid forms (chainsaw devil hybrid).
- The SD 1.5 adapter is intentionally lower-fidelity (800 steps at 512 res)
  for use as a lightweight fallback.

## License

Released under [OpenRAIL++](https://huggingface.co/spaces/CompVis/stable-diffusion-license)
(`openrail++`), inheriting from the base model
[Animagine XL 3.1](https://huggingface.co/cagliostrolab/animagine-xl-3.1).

## Related models

Part of the [AnimeLoom](https://huggingface.co/AnimeLoom) character collection:

- [AnimeLoom/sakura-haruno](https://huggingface.co/AnimeLoom/sakura-haruno)
- [AnimeLoom/denji](https://huggingface.co/AnimeLoom/denji) (this card)
- [AnimeLoom/yuki-nagato](https://huggingface.co/AnimeLoom/yuki-nagato)

## Citation

If you use this LoRA in research or production, please credit AnimeLoom and
the base model authors (Cagliostro Lab for Animagine XL 3.1, Lykon for
DreamShaper 8).

---

*Trained with [PEFT](https://huggingface.co/docs/peft) as part of the
[AnimeLoom](https://github.com/JoelJohnsonThomas/AnimeLoom) project.*