| --- |
| 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. |
|
|
