File size: 2,633 Bytes
a50f08e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | ---
license: mit
pipeline_tag: image-to-video
tags:
- animate-anyone
- sprite-sheet
- game-character
- pose-guided
- stable-diffusion
- sd1.5
---
# Sprite-Sheet-Diffusion weights (community re-host)
Re-host of the released model weights for **Sprite Sheet Diffusion:
Generate Game Character for Animation** (Hsieh, Zhang, Yan;
[arXiv:2412.03685](https://arxiv.org/abs/2412.03685);
[upstream repo](https://github.com/chenganhsieh/Sprite-Sheet-Diffusion)).
The authors released these only via a Google Drive folder whose readme
links had gone stale; this repo preserves the recovered files.
An AnimateAnyone-style checkpoint set, finetuned (30000 steps) on game
sprite data:
| file | role | note |
|---|---|---|
| `denoising_unet.pth` | denoising UNet (SD1.5 family) | finetuned, differs from the Moore/patrolli AnimateAnyone baseline |
| `reference_unet.pth` | ReferenceNet (headless SD1.5 UNet) | finetuned, differs from the baseline |
Not included, by upstream release status:
- **pose guider: never released** (upstream issue #3 is open and
unanswered; no mirror is known to exist). Use the Moore baseline
`pose_guider.pth` from
[patrolli/AnimateAnyone](https://huggingface.co/patrolli/AnimateAnyone).
- **motion module**: the Drive copy is byte-identical to the baseline
`motion_module.pth` from patrolli/AnimateAnyone, so it is not
duplicated here.
Companion files needed for inference: VAE
[stabilityai/sd-vae-ft-mse](https://huggingface.co/stabilityai/sd-vae-ft-mse)
and the CLIP vision encoder from
[lambdalabs/sd-image-variations-diffusers](https://huggingface.co/lambdalabs/sd-image-variations-diffusers)
(`image_encoder/`).
## Provenance
Recovered 2026-08-28 from the paper's public Google Drive folder
`1VxbOv5PE441NsNStQlmqbIw0iyY9Mn9L` (files
`denoising_unet-30000.pth`, `reference_unet-30000.pth`). Verified against
the Moore baseline: identical tensor key sets and shapes, weight values
differ (max sampled abs diff ~0.010). sha256:
```
341cca53cfaa4e0c05098e511b8a3dc1a0db90c6ec68f345ba14115e0d3e43ac denoising_unet.pth
84194364a42b5ae8b2a93a60b02a36ed0f989c230bb4c3ee33d7956bba5d0dcc reference_unet.pth
```
## Usage
Runs in PyTorch with the upstream repo, or natively in C++ via the
AnimateAnyone port of stable-diffusion.cpp at
[fszontagh/stable-diffusion.cpp](https://github.com/fszontagh/stable-diffusion.cpp)
(branch `feat/sprite-sheet-diffusion`, see `docs/animate_anyone.md`):
pass these as `--diffusion-model` and `--reference-net`, with the Moore
baseline pose guider and motion module.
License: MIT (upstream project license); training data CC0 per the
upstream readme.
|