AniSora V3.2 vs V3.1
- Base model: V3.2 uses Wan2.2 (task i2v-A14B); V3.1 uses Wan2.1 image-to-video 14B.
- Speed: V3.2 is documented with 8 sampling steps.
- Checkpoints: V3.2 loads separate high-noise and low-noise models.
- Features: both support arbitrary-frame inference and character 3D video generation.
- Memory: a 12 GB VRAM build has been published for V3.1, not for V3.2.
Run AniSora V3.2 locally
Set up the environment from the anisoraV3.2 folder and download the V3.2 checkpoints from Hugging Face. Single-GPU inference:
cd anisoraV3.2
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_txt_new.py \
--task i2v-A14B --size 1280*720 \
--ckpt_dir Index-anisora-V3.2 \
--prompt_list data/inference_any.txt \
--save_dir test_any_outputs \
--sample_steps 8 --sample_shift 5 --sample_guide_scale 1 \
--ckpt_dir_lowname low_noise_model \
--ckpt_dir_highname high_noise_model \
--base_seed 4096Character 360-degree rotation
Use data/inference_360.txt as the prompt list and keep --frame_num 81. The maintainers recommend about 5 seconds so the character completes a full turn. For multiple GPUs, launch with torchrun and set --ulysses_size to the GPU count.
Prompt tips for V3.2
V3.2 was trained on long, detailed prompts. The maintainers suggest expanding short ideas with an LLM first, then appending the aesthetic score, motion score, and no-text clause:
[Detailed video description]. 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.