| --- |
| tags: |
| - comfyui |
| - model-bundle |
| - image-generation |
| - video-generation |
| --- |
| |
| # Character Image Model Bundle |
|
|
| Deployment assets for the Character Image project, audited against commit `8f1d132` on 2026-09-07. This public repository contains the 25 referenced model files found on the source server, totaling 117,028,825,625 bytes (108.99 GiB). |
|
|
| ## Directory Layout |
|
|
| | Directory | Contents | Files | |
| | --- | --- | ---: | |
| | `models/checkpoints/` | WAI v170, Qwen Rapid AIO v23, Pornworks Illustrious | 3 | |
| | `models/diffusion_models/` | Wan Remix V30 High/Low, two ZImage variants | 4 | |
| | `models/text_encoders/` | Wan UMT5 and Qwen 3 4B | 2 | |
| | `models/clip_vision/` | CLIP Vision H | 1 | |
| | `models/ipadapter/` | AnimeIPA and SDXL Plus Face | 2 | |
| | `models/loras/` | Six enabled LoRA dependencies | 6 | |
| | `models/vae/` | AE and Wan 2.1 VAE | 2 | |
| | `models/upscale_models/` | 4x AnimeSharp | 1 | |
| | `models/onnx/` | RIFE 4.9 ONNX | 1 | |
| | `auxiliary/censor/` | EraX, NudeNet, anime censor detector | 3 | |
|
|
| `model-manifest.json` lists each path, SHA256, byte size, and the template keys that reference it. Shared weights appear once. The model bytes and filenames are preserved from the source deployment. |
|
|
| ## Download |
|
|
| Install `huggingface_hub`. This public repository can be downloaded without logging in. The following command downloads into the existing ComfyUI directory: |
|
|
| ```bash |
| python3 - <<'PY' |
| from huggingface_hub import snapshot_download |
| snapshot_download( |
| repo_id="ChatoyantAI/character-image-models", |
| local_dir="/root/ComfyUI", |
| max_workers=4, |
| token=False, |
| ) |
| PY |
| ``` |
|
|
| The existing `models/` layout is directly compatible with ComfyUI. Re-running the download supports resuming an interrupted transfer. |
|
|
| ## Restore Censor Dependencies |
|
|
| Two detectors are loaded through upstream HF cache paths, while NudeNet can load a local ONNX. After downloading: |
|
|
| ```bash |
| python3 /root/ComfyUI/restore_censor_models.py --repo-dir /root/ComfyUI |
| export CENSOR_MODELS_DIR=/root/ComfyUI/auxiliary/censor |
| ``` |
|
|
| Keep the `CENSOR_MODELS_DIR` export in the service startup command, before ComfyUI starts. Run restoration as the service account, or supply its HF cache directory with `--cache-dir`. The script verifies and restores the EraX and anime detector cache entries at the recorded upstream revisions; their `refs/main` entries are set to those revisions. `CENSOR_MODELS_DIR` covers EraX/NudeNet, while the anime detector continues to use its HF cache. |
|
|
| The original node packages and Python dependencies are still required, including `ComfyUI-GenitalCensor`, `nudenet`, `ultralytics`, and `imgutils`. |
|
|
| ## RIFE |
|
|
| `models/onnx/rife49_ensemble_True_scale_1_sim.onnx` is the input for `ComfyUI-Rife-Tensorrt`. Its node builds a TensorRT engine on the target machine when needed. This bundle contains no precompiled TensorRT engine. |
|
|
| ## Coverage |
|
|
| Model dependencies were found for WAI_AnimeIPA, WAI_Qwen, WAI + Wan22 (image/text variants), Illustrious, WanImage, ZImageTurbo, and image/video censoring. Dependencies shared with other configured templates are also included. |
|
|
| Eighteen model references were absent from the source server and are not included: the legacy Pony checkpoint and five associated LoRAs; Wan Remix V20 High/Low; the Qwen next-scene LoRA; Monica/Trixie LoRAs and their ZImage NSFW LoRA; and six legacy video action LoRAs. Those templates require additional assets. Legacy Pony templates also require missing Comfyroll nodes and a compatible CLIP discovery filename. |
|
|
| Only model files and deployment metadata are included. Application configuration, credentials, user media, and logs are excluded. See the manifest for the exact uploaded set. |
|
|
