AniSora V3 and V3.1: features, setup, and prompts

What AniSora V3 adds

The official README says V3 can generate a 5-second 360p shot within 8 seconds. Real speed depends on your GPU and settings.

• Arbitrary-frame inference: guide a clip with an image placed at the first, middle, or last frame.
• Character 3D video generation: turn a front-facing character illustration into a 360-degree rotation.
• Video style transfer using first and last frames combined with line-art guided generation.
• Multimodal guidance with pose, depth, line art, and audio.
• Ultra-low-resolution video super-resolution, from 90p up to 720p or 1080p.

What changed in V3.1

V3.1 improves motion range. The maintainers recommend a motion score between 2.0 and 4.0 for the best results. A version of V3.1 that fits in 12 GB of VRAM was published on ModelScope on 25 September 2025.

Run AniSora V3.1 locally

Create a Python 3.10 environment, install the dependencies from the anisoraV3 folder, and download the V3.1 checkpoints from Hugging Face or ModelScope. Then run single-GPU inference:

cd anisoraV3
conda create -n wan_gpu python=3.10
conda activate wan_gpu
pip install -r req-fastvideo.txt
pip install -r requirements.txt
pip install -e .

python generate-pi-i2v-any.py \
  --task i2v-14B --size 1280*720 \
  --ckpt_dir Wan2.1-I2V-14B-480P \
  --image output_videos_any \
  --prompt data/inference_any.txt \
  --base_seed 4096 --frame_num 81

Multi-GPU and frame count

• Use torchrun with --dit_fsdp, --t5_fsdp, and --ulysses_size; set --nproc_per_node and --ulysses_size to your GPU count.
• Frame count must follow 8x+1. 81 frames at 16 fps is about 5 seconds, which is also the recommended length for a full 360-degree character rotation.
• Each prompt-file line uses image_path@@prompt&&image_position, where the position is 0 (first frame), 0.5 (middle), or 1 (last frame).

Official prompt format

Write a detailed English description, then append an aesthetic score (5.0–7.0), a motion score (2.0–4.0), and the mandatory no-text clause:

A drone chase sequence through neon-lit city streets at night. aesthetic score: 5.5. motion score: 4.0. There is no text in the video.

Try anime image-to-video without a GPU

The hosted generator on AniSora Studio currently animates your image with Kling 2.1 image-to-video through fal. It does not run the open-source AniSora weights, and the model used is always disclosed before you spend credits.

AniSora Studio is an independent product. The official Index-anisora GitHub repository states that it is the only official homepage of the AniSora project.