AnimeGen-I2V

AnimeGen I2V

日本語版モデルカード

Overview

AnimeGen I2V is a image-to-video generation model specialized for Japanese anime-style video generation.

AnimeGen I2V was developed by AIdeaLab, Inc. as part of its effort to build high-quality generative AI models for animation and video production. The model was developed with support from GENIAC — Generative AI Accelerator Challenge, a Japanese government-backed project led by the Ministry of Economy, Trade and Industry of Japan (METI) and the New Energy and Industrial Technology Development Organization (NEDO), whose goal is to strengthen domestic capabilities in generative AI development.

AnimeGen I2V is based on Wan 2.2 I2V A14B and further trained to improve anime-style motion, character rendering, color composition, and visual consistency for animation-oriented use cases.

The model is released under the Apache-2.0 license.

<video src="https://huggingface.co/aidealab/AnimeGen-I2V/resolve/main/FutureLoveSong.mp4" width="100%" controls loop></video>

How to Get Started

Requirements

AnimeGen I2V is intended to be used in a Python environment with PyTorch and Diffusers.

A high-end NVIDIA GPU is recommended. For local inference, an NVIDIA RTX 4090 or higher is recommended. Larger VRAM GPUs may provide a more stable experience, especially for longer videos or higher resolutions.

Install the required libraries:

pip install torch torchvision transformers diffusers accelerate peft ftfy imageio imageio-ffmpeg safetensors

Example I2V Inference Code

sample.png

import torch
from diffusers import WanImageToVideoPipeline, WanTransformer3DModel, FlowMatchEulerDiscreteScheduler, AutoencoderKLWan
from diffusers.utils import export_to_video, load_image
import numpy as np

scheduler = FlowMatchEulerDiscreteScheduler(shift=3.0)
transformer_high = WanTransformer3DModel.from_pretrained(
    "aidealab/AnimeGen-I2V",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
)

transformer_low = WanTransformer3DModel.from_pretrained(
    "aidealab/AnimeGen-I2V",
    subfolder="transformer_2",
    torch_dtype=torch.bfloat16,
)

vae = AutoencoderKLWan.from_pretrained("Wan-AI/Wan2.2-I2V-A14B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
pipe = WanImageToVideoPipeline.from_pretrained("Wan-AI/Wan2.2-I2V-A14B-Diffusers", transformer=transformer_high, transformer_2=transformer_low, scheduler=scheduler, vae=vae, torch_dtype=torch.bfloat16)
pipe.load_lora_weights('lightx2v/Wan2.2-Lightning', weight_name='Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/high_noise_model.safetensors', adapter_name='high')
pipe.load_lora_weights('lightx2v/Wan2.2-Lightning', weight_name='Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/low_noise_model.safetensors', adapter_name='low', load_into_transformer_2=True)
pipe.set_adapters(["high", "low"], adapter_weights=[1.0, 1.0])

transformer_high.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
transformer_low.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)

pipe.enable_model_cpu_offload()

width=832 # 1280
height=480 # 720

secs=5 # 2

image = load_image("sample.png")
max_area = height * width
aspect_ratio = image.height / image.width
mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
image = image.resize((width, height))

prompt = "blink her eyes."
negative_prompt = "3d, cg, photo, stop, wait"

generator = torch.Generator("cuda").manual_seed(42)

output = pipe(
    image=image,
    prompt="Japanese anime style, "+prompt,
    negative_prompt=negative_prompt,
    height=height,
    width=width,
    num_frames=int(16*secs+1),
    guidance_scale=1.0,
    num_inference_steps=4,
    generator=generator
).frames[0]
export_to_video(output, f"output.mp4", fps=16)

Example Frame Interpolation Inference Code

sample_1.png

sample_2.png

import torch
from diffusers import WanImageToVideoPipeline, WanTransformer3DModel, FlowMatchEulerDiscreteScheduler, AutoencoderKLWan
from diffusers.utils import export_to_video, load_image
import numpy as np

scheduler = FlowMatchEulerDiscreteScheduler(shift=3.0)
transformer_high = WanTransformer3DModel.from_pretrained(
    "aidealab/AnimeGen-I2V",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
)

transformer_low = WanTransformer3DModel.from_pretrained(
    "aidealab/AnimeGen-I2V",
    subfolder="transformer_2",
    torch_dtype=torch.bfloat16,
)

vae = AutoencoderKLWan.from_pretrained("Wan-AI/Wan2.2-I2V-A14B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
pipe = WanImageToVideoPipeline.from_pretrained("Wan-AI/Wan2.2-I2V-A14B-Diffusers", transformer=transformer_high, transformer_2=transformer_low, scheduler=scheduler, vae=vae, torch_dtype=torch.bfloat16)
pipe.load_lora_weights('lightx2v/Wan2.2-Lightning', weight_name='Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/high_noise_model.safetensors', adapter_name='high')
pipe.load_lora_weights('lightx2v/Wan2.2-Lightning', weight_name='Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/low_noise_model.safetensors', adapter_name='low', load_into_transformer_2=True)
pipe.set_adapters(["high", "low"], adapter_weights=[1.0, 1.0])

transformer_high.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
transformer_low.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)

pipe.enable_model_cpu_offload()

width=832 # 1280
height=480 # 720

secs=5 # 2

image = load_image("sample_1.png")
max_area = height * width
aspect_ratio = image.height / image.width
mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
image = image.resize((width, height))

image_2 = load_image("sample_2.png")
image_2 = image_2.resize((width, height))


prompt = "Make this character press both hands together in prayer and close their eyes."
negative_prompt = "3d, cg, photo, stop, wait"

generator = torch.Generator("cuda").manual_seed(42)

output = pipe(
    image=image,
    last_image=image_2,
    prompt="Japanese anime style, "+prompt,
    negative_prompt=negative_prompt,
    height=height,
    width=width,
    num_frames=int(16*secs+1),
    guidance_scale=1.0,
    num_inference_steps=4,
    generator=generator
).frames[0]
export_to_video(output, f"output.mp4", fps=16)

For ComfyUI

high_noise

low_noise

Model Details

Model Description

AnimeGen I2V is a text-to-video model fine-tuned from Alibaba’s Wan 2.2 video generation model. By building on Wan 2.2, AnimeGen I2V benefits from the existing Wan ecosystem, including inference pipelines, model conversion tools, and deployment workflows.

The goal of AnimeGen I2V is to make anime-style video generation more accessible for creators, researchers, and developers. It is intended for tasks such as visual prototyping, animation previsualization, motion concept exploration, and short-form anime-style video generation.

• Developed by: AIdeaLab, Inc.
• Funded by: METI, NEDO, and AIdeaLab, Inc.
• Base model: Wan-AI/Wan2.2-I2V-A14B
• Model type: Diffusion Transformer
• Training objective: Flow Matching
• Primary modality: Image-to-video
• Primary style domain: Japanese anime-style video
• Language: English prompts are recommended
• License: Apache-2.0

Intended Uses

Direct Use

AnimeGen I2V can be used for:

• anime-style text-to-video generation
• animation previsualization
• movie storyboard and video storyboard prototyping
• visual concept exploration
• character motion studies
• background, composition, and atmosphere exploration
• short-form anime-style content creation
• research and experimentation in video generation

The model is especially suitable for creators and researchers who want to quickly explore anime-style moving images from text prompts.

Research and Development Use

AnimeGen I2V may also be useful for:

• studying anime-style video generation
• evaluating temporal consistency in generated video
• testing text-to-video workflows
• building animation production support tools
• developing controllable video generation pipelines

Out-of-Scope Uses

The model must not be used for:

• creating deepfakes of real people
• impersonation or identity deception
• generating misleading political, journalistic, or documentary-style content
• harassment, defamation, or non-consensual depictions
• creating sexualized content involving minors or minor-looking characters
• generating illegal, harmful, or abusive content
• falsely presenting generated content as real footage

Users are responsible for ensuring that their use of the model complies with applicable laws, platform policies, and ethical guidelines.

Bias, Risks, and Limitations

AnimeGen I2V is optimized for anime-style visual expression. It is not intended to generate photorealistic video.

Known limitations include:

• limited suitability for photorealistic or live-action video generation
• possible instability in complex character motion
• possible inconsistencies in hands, fingers, eyes, accessories, and fine details
• possible identity drift across frames
• possible temporal flickering
• difficulty with long or highly detailed action sequences
• limited understanding of highly complex physical interactions
• reduced performance outside anime-style prompts
• limited suitability for NSFW generation

Because this is a generative model, outputs may contain unexpected artifacts or unintended visual content. Users should review generated outputs before publication or commercial use.

Training and Development Background

AnimeGen I2V was developed by AIdeaLab as part of a broader research and development effort in Japanese text-to-video and anime-style generative AI.

AIdeaLab has worked on video generation models and related generative AI systems, including Japanese text-to-video model development and animation-oriented generation workflows. AnimeGen I2V builds on this experience and focuses specifically on improving anime-style video generation using a strong open text-to-video base model.

The model was trained using GPU compute resources provided through a KDDI GPU cluster located in Kagawa, Japan.

Technical Specifications

Architecture

• Architecture: Diffusion Transformer
• Base model: Wan-AI/Wan2.2-I2V-A14B
• Objective: Flow Matching
• Pipeline: WanPipeline / Diffusers-compatible workflow
• Recommended inference dtype: bfloat16
• Optional memory optimization: CPU offload and layerwise casting

Compute Infrastructure

AnimeGen I2V was developed using a GPU cluster provided through KDDI.

Hardware

• 24 × NVIDIA H200 GPUs

Software

The development workflow was based in part on:

• DiffSynth-Studio
• PyTorch
• Diffusers
• Transformers
• Accelerate

License

This model is licensed under the Apache License, Version 2.0. See the LICENSE file for details.

Copyright 2026 AIdeaLab, Inc.

Citation

If you use AnimeGen I2V in your research, product, article, or creative project, please cite it as:

@article{AIdeaLab2026AnimeGenI2V,
  title   = {AnimeGen I2V},
  author  = {AIdeaLab Development Team},
  year    = {2026},
  note    = {Anime-style text-to-video generation model}
}

Model Card Contact

For inquiries, please contact AIdeaLab:

https://aidealab.com/contact