axiomofmind/Fibo-2.0-anime-lora

Fibo 2.0 Anime LoRA

An anime and seinen-style illustration LoRA for BRIA Fibo 2.0 Dev. This repository contains the step-400 checkpoint from the fibo-anime-v1-pilot-400 training run.

The adapter was trained with the trigger fibo_seinen_style. Use the trigger at the beginning of style_medium in a structured Fibo prompt. A LoRA scale of 0.7 is the recommended starting point; use 1.0 for a stronger learned-style effect.

License: Fibo 2.0 Dev is gated for non-commercial use under BRIA's Fibo terms. The base repository identifies CC BY-NC 4.0 and states that commercial use requires a separate BRIA agreement. Review the base-model license and access terms before use or redistribution.

Files

File	Description
pytorch_lora_weights.safetensors	Checkpoint-400 Diffusers / PEFT adapter
Fibo-2.0-Anime-LoRA-ComfyUI.safetensors	Checkpoint-400 adapter converted to Fibo's native ComfyUI layout
adapter_config.json	Diffusers adapter configuration
scripts/generate_fibo_lora.py	Ready-to-run structured-prompt Diffusers example
images/	Stock, no-trigger, and triggered checkpoint-400 comparison examples

Recommended settings

• Base model: briaai/Fibo-2.0-dev
• Trigger: fibo_seinen_style
• Preferred starting scale: 0.7
• Stronger scale: 1.0
• Inference steps: 50
• Guidance scale: 5.0
• Resolution: 1024 x 1024
• Prompt format: structured Thales VGL JSON

Stock versus LoRA without the trigger

These examples use the same structured JSON scene prompt, seed, dimensions, 50 inference steps, and guidance scale 5.0 in both columns. The stock column disables the adapter. The no-trigger column enables checkpoint 400 at LoRA scale 1.0 but leaves fibo_seinen_style out of style_medium.

Prompt	Fibo stock	LoRA 1.0 — no trigger
Rooftop runner, seed 42
Late-night diner, seed 314159
Older woman portrait, seed 8675309

The adapter produces a visible learned-style shift even without its trigger. The trigger remains recommended because it matches the training captions and makes the intended style explicit.

Triggered scale 0.7 examples

These pairs use the same structured JSON scene prompt, seed, dimensions, 50 inference steps, and guidance scale 5.0 in both columns. The LoRA column enables checkpoint 400 at the recommended scale 0.7 and begins style_medium with fibo_seinen_style.

Prompt	Fibo stock	LoRA 0.7 — triggered
Rooftop runner, seed 42
Late-night diner, seed 314159
Orbital greenhouse botanist, seed 20261024
Storm lighthouse keeper, seed 161803
Desert train courier, seed 424242

Triggered example

This pair uses the same structured prompt, seed 57721, 768 × 1344 resolution, 50 inference steps, and guidance scale 5.0. The LoRA image enables checkpoint 400 at scale 1.0 and begins style_medium with fibo_seinen_style.

Prompt	Fibo stock	LoRA 1.0 — triggered
Elderly watchmaker

Structured JSON prompt example

Pass the serialized JSON object as Fibo's positive prompt. Keep the trigger in style_medium:

{
  "short_description": "An adult swordswoman stands on a rain-slick rooftop above a neon city at night.",
  "objects": [
    {
      "description": "an adult swordswoman in a dark fitted jacket holding a sheathed katana",
      "location": "center foreground"
    }
  ],
  "background_setting": "A dense futuristic city skyline with glowing signs, antennas, and drifting rain.",
  "lighting": {
    "description": "Cool moonlight and magenta city light reflect across the wet rooftop."
  },
  "aesthetics": {
    "composition": "A dramatic low-angle full-body composition with layered city depth.",
    "color_palette": "midnight blue, violet, magenta, and silver",
    "mood": "tense, cinematic, and contemplative"
  },
  "style_medium": "fibo_seinen_style, detailed anime illustration",
  "context": "A quiet moment before a confrontation.",
  "artistic_style": "crisp linework, expressive anime character design, cinematic background detail",
  "text_render": []
}

The included example uses this structured negative prompt:

{
  "short_description": "",
  "objects": [
    {
      "description": "the primary character",
      "appearance_details": "malformed anatomy, distorted hands, extra fingers, extra limbs, duplicated features, or an obscured face"
    }
  ],
  "background_setting": "unintended text, captions, logos, signatures, or watermarks",
  "lighting": {},
  "aesthetics": {
    "artifacts": [
      "blur",
      "low detail",
      "compression artifacts",
      "color banding",
      "clipped highlights"
    ]
  },
  "style_medium": "",
  "context": "",
  "artistic_style": ""
}

Diffusers usage

Access to the gated Fibo base model is required. Use a recent Diffusers checkout containing BriaFibo2Pipeline and BriaFibo2LoraLoaderMixin:

pip install -U git+https://github.com/huggingface/diffusers.git transformers accelerate peft safetensors
huggingface-cli login

import json
import torch
from diffusers import BriaFibo2Pipeline

base_model = "briaai/Fibo-2.0-dev"
lora_repo = "axiomofmind/Fibo-2.0-anime-lora"

prompt = json.dumps(
    {
        "short_description": "An adult swordswoman stands on a rain-slick rooftop above a neon city at night.",
        "objects": [{
            "description": "an adult swordswoman in a dark fitted jacket holding a sheathed katana",
            "location": "center foreground"
        }],
        "background_setting": "A dense futuristic city skyline with glowing signs, antennas, and drifting rain.",
        "lighting": {"description": "Cool moonlight and magenta city light reflect across the wet rooftop."},
        "aesthetics": {
            "composition": "A dramatic low-angle full-body composition with layered city depth.",
            "color_palette": "midnight blue, violet, magenta, and silver",
            "mood": "tense, cinematic, and contemplative"
        },
        "style_medium": "fibo_seinen_style, detailed anime illustration",
        "context": "A quiet moment before a confrontation.",
        "artistic_style": "crisp linework, expressive anime character design, cinematic background detail",
        "text_render": []
    },
    separators=(",", ":")
)

pipe = BriaFibo2Pipeline.from_pretrained(base_model, dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights(
    lora_repo,
    weight_name="pytorch_lora_weights.safetensors",
    adapter_name="anime"
)
pipe.set_adapters("anime", adapter_weights=0.7)

image = pipe(
    prompt=prompt,
    width=1024,
    height=1024,
    num_inference_steps=50,
    guidance_scale=5.0,
    generator=torch.Generator(device="cuda").manual_seed(42)
).images[0]
image.save("fibo_anime.png")

The command-line example includes the complete structured negative prompt:

python scripts/generate_fibo_lora.py --scale 0.7 --seed 42 --output fibo_anime.png

Pass --prompt-json path/to/prompt.json to use another structured prompt. Its style_medium field must include fibo_seinen_style.

ComfyUI compatibility

The root pytorch_lora_weights.safetensors uses separate Diffusers / PEFT projections. ComfyUI's native Fibo 2.0 model uses fused Q/K/V, K/V, and SwiGLU W1/W3 projections, so use Fibo-2.0-Anime-LoRA-ComfyUI.safetensors in ComfyUI.

1. Update ComfyUI to a version with native Fibo 2.0 support; commit 87ed5a0 or newer contains the required model support.
2. Install the Fibo Dev diffusion model, Qwen3-VL text encoder, and Flux.2 VAE from Comfy-Org/Fibo-2.0. Use the Dev model, not Edit Turbo.
3. Put Fibo-2.0-Anime-LoRA-ComfyUI.safetensors in ComfyUI/models/loras/.
4. Use the standard LoRA loader with model strength 0.7 and clip strength 0.0.
5. Use Fibo's structured prompt encoder and place fibo_seinen_style at the beginning of style_medium.

Download only the converted variant with:

hf download axiomofmind/Fibo-2.0-anime-lora \
  Fibo-2.0-Anime-LoRA-ComfyUI.safetensors \
  --local-dir ./ComfyUI/models/loras

Limitations

• Scale 0.7 is a good starting point, but the best value can vary with subject and prompt.
• The learned style can remain visible whenever the adapter is enabled; the trigger does not guarantee complete gating.
• Fibo works best with structured VGL prompts. Plain text may be less predictable.
• Base-model limitations, biases, and licensing terms continue to apply.