Paints-UNDO (Easy Install Fork)

This is a fork of lllyasviel/Paints-UNDO with a setup script that handles installation, dependency fixes, and GPU detection automatically.

What's different from the original

• setup.py — one-command install that auto-detects your GPU and installs the correct PyTorch CUDA version
• Fixes a gradio_client bug that causes UI buttons to be greyed out (APIInfoParseError: Cannot parse schema True)
• Fixes a localhost binding issue that prevents the Gradio UI from loading
• Fixes Step 3 video generation on RTX 5000 series (and other GPUs where xformers lacks CUDA support) by patching diffusers_vdm/vae.py and diffusers_vdm/attention.py to fall back to PyTorch native attention
• Fixes Step 3 video saving which fails on newer torchvision versions (write_video missing) by switching to imageio
• Pins dependency versions known to work together
• Verifies PyTorch can see your GPU after install
• Warns if your VRAM is too low before you start
• --update flag to re-apply patches without reinstalling

Added features (beyond the original)

• Multi-model WD tagger with a Step 1 picker — choose one or several taggers in the UI; multiple are run as an ensemble and merged by max confidence. Newer, more accurate models than the original (EVA02-Large / ViT-Large v3) are the default.
• GPU tagging — the WD tagger runs on CUDA (via onnxruntime-gpu, CUDA-matched at install) with automatic CPU fallback.
• Cancel buttons for Step 2 (key frames) and Step 3 (video) so you can stop a long run without killing the app.
• Input validation — friendly messages instead of tracebacks when no image is uploaded, no operation steps are selected, or there are too few key frames.
• VRAM-capped key frames — Step 2 generates in sub-batches (output unchanged) so large step selections don't OOM.
• High-VRAM mode — on GPUs with >= 20 GB, models stay resident for faster runs (auto-enabled by the installer).

Requirements

• Python 3.7+ (just to run the setup script — any Python will do)
• Miniconda or Anaconda installed and in PATH
• Git installed and in PATH
• NVIDIA GPU with up-to-date drivers (RTX 2000 series or newer recommended, 10GB+ VRAM for best results)

Install

Windows (easiest) — double-click setup.bat. It finds conda automatically (even if it is not on your PATH), then runs the installer. You will be prompted for an install location.

From a terminal (any platform):

python setup.py

You will be prompted for an install location. Everything else is automatic.

On Windows, run setup.py from an Anaconda Prompt (or just use setup.bat). Double-clicking setup.py directly will fail because conda is not on PATH.

Or specify the path directly:

setup.bat --install-dir C:\AI\PaintsUndo
python setup.py --install-dir C:\AI\PaintsUndo
python setup.py --install-dir ~/PaintsUndo

First launch downloads the models automatically (~8-10 GB):

• lllyasviel/paints_undo_single_frame — keyframe generation (SD1.5-based)
• lllyasviel/paints_undo_multi_frame — video interpolation model

Launch

After setup, a launch script is written to your install directory. Run it from there:

Windows — run start.bat from your install folder, or double-click it in Explorer.

Linux — run bash start.sh from your install folder.

Then open http://127.0.0.1:7860 in your browser.

Note: the start.bat and start.sh in this repo are templates. The actual launch scripts are written to your install directory by setup.py.

Usage

Step 1 — Upload your image, optionally choose which tagger model(s) to use in the dropdown, then click Generate Prompt. Selecting multiple runs them as an ensemble and merges the tags. Models download on first use.

Step 2 — Click Generate Key Frames. Default operation steps (400, 600, 800, 900, 950, 999) work well. Steps 900 and 950 produce the most useful loose gesture sketches.

Step 3 — Video interpolation between keyframes. Works on all supported GPUs. May still OOM on cards with less than 12 GB VRAM — if so, reduce Image Width and Height in Step 2 before generating.

Steps 2 and 3 each have a Cancel button to stop a running generation.

Advanced configuration (environment variables)

Set these before launching (or add them to your start script) to change behavior without editing code:

Variable	Default	Effect
PAINTS_UNDO_TAGGERS	eva02-large-v3,vit-large-v3	Comma-separated taggers for Step 1. Options: moat-v2, vit-v3, vit-large-v3, swinv2-v3, convnext-v3, eva02-large-v3. The Step 1 dropdown overrides this.
PAINTS_UNDO_HIGH_VRAM	0	Set to 1 to keep all models resident on the GPU (faster, needs more VRAM). Auto-set by the installer on >= 20 GB cards.
PAINTS_UNDO_KEYFRAME_BATCH	6	Max key frames generated per sub-batch in Step 2. Lower it if Step 2 runs out of memory.

Updating patches

If you upgrade any packages and the UI breaks, re-apply the patches without reinstalling:

python setup.py --update --install-dir C:\AI\PaintsUndo

Troubleshooting

Buttons are greyed out / APIInfoParseError Re-run python setup.py --update to re-apply the gradio_client patch.

When localhost is not accessible error On Windows, reset your network stack as administrator:

netsh int ipv4 reset
netsh int ipv6 reset

Then reboot.

conda not found error You need to run the script from an Anaconda Prompt on Windows, or a terminal where conda init has been run on Linux.

OOM on Step 2 Reduce Image Width and Height in the UI. Try 448x448 instead of 512x512.

Step 3 OOM error Reduce Image Width and Height in the UI in Step 2 before running Step 3. Try 384x448 instead of 512x640.

WSL (Windows Subsystem for Linux) Not recommended. Run natively on Windows instead for best GPU support.

Notes

• Models are saved to <install_dir>/hf_download/ on first launch
• The conda environment (paints_undo) is fully isolated from all other Python environments on your system
• AMD and Intel GPUs are not supported — CUDA is NVIDIA-only

Credits

Original project by lllyasviel. This fork adds installation tooling plus the added features listed above.