# AniSora - Turn Table Test

For Sprute, we are exploring an offline alternative to Nano Banana Pro for generating directional character views. This test uses a single front-facing image to generate a full rotation, without a driving video.

**Model:** [Index-AniSora V3.2](https://github.com/bilibili/Index-anisora), based on Wan2.2 I2V A14B. We used [Kijai’s scaled FP8 high/low checkpoints](https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/I2V/AniSora).

**Configuration:**

```text
Resolution: 640 × 640
Frames: 81
Playback: 16 fps (~5.06 seconds)
Seed: 42
Steps: 8 total (3 high-noise + 5 low-noise)
CFG: 1
Sampler: uni_pc
Scheduler: simple
Model shift: 5
Batch size: 1
Additional LoRAs: none
Driving video: none
GPU: NVIDIA RTX PRO 6000 Blackwell 96GB
Measured Comfy graph execution: 84.6 seconds
```

Timing includes graph execution/loading and is not a standardized throughput benchmark. This is a native-Comfy adaptation using third-party FP8 weights, not an exact reproduction of the upstream F32 pipeline. Two sampler calls reset UniPC history at the expert switch.

**Input Image:**

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/cb790d6f-19b0-4b30-a49e-8910eec27bad.png align="center")

One full-body front view of Lily, padded to a 256×256 gray canvas.

**Prompt:**

```text
A 360-degree turning and circling video of a 2D RPG game character. This is a full-body long shot with a plain gray background. A girl with long brown hair and a large blue patterned hair bow wears a sleeveless blue patterned top, blue shorts, white socks and white shoes. She stands upright with her arms resting beside her and rotates in place through one full turn. Preserve her reference proportions, clothes and hairstyle. Clear readable silhouette, clean contours, flat local colors with restrained cel shading. Fixed camera, constant character size, feet on the same spot. aesthetic score: 5.5. motion score: 3.0. There is no text in the video.
```

**Negative Prompt (configured):**

```text
photorealistic, glossy plastic, sculpted figurine, dramatic lighting, gradients, walking, running, dancing, camera movement, zoom, cropped feet, outfit change, extra limbs, text, scene change
```

At CFG 1, standard classifier-free guidance does not apply a negative-prompt correction; listing this field does not mean it caused the result.

**Result:**

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/f8496bd2-1aa9-4a97-968f-21e4247ff09d.webp align="center")

**8-Directions Extracted**:

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/28cfdc7f-1b83-4605-87f5-53eddc967a1b.webp align="center")

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/024d7cec-2f8d-4c69-8410-1f6dff40fd25.png align="center")

**Seed Variations**:

![](https://cdn.hashnode.com/uploads/covers/682665f051e3d254b7cd5062/f881c26f-9e45-4591-a93c-db0e28addc29.webp align="center")

**Discussion:**

AniSora seems to do the work pretty well. It is a good competitor to the SCAIL2 and Seedance 2 Pro.

It does have that 3D animation look which I'm not super happy about but at least it works.

**Next Steps:**

*   We got few option source options to take out Nano Banana Pro out of Sprute.
    
    *   SCAIL 2 3x3 Grid
        
    *   SCAIL 2 Turntable
        
    *   AniSora Turntable
        
*   Let's figure out prompt engineering and post-processing for AniSora approach to make it more sprite animation like.
    
    *   We could add "pixel art" to the prompts.
        
    *   We could feed result into an I2I which consistently turns it into pixel art.
        

— Sprited Dev 🐛