# AI models and licences

> The 37 AI models the AI nodes can download, with licences and tiers.

The AI nodes run open models in a separate AI worker process on your GPU. Nothing is sent to a cloud service. Models are downloaded to your machine the first time a node needs them and checked against the sizes and checksums below. Oink never redistributes model weights.

**Tiers:**

- **bundled**: downloaded when a node first needs it; permissive licence
- **optional**: you must accept its licence in the app before it downloads
- **research**: the research pack: NON-COMMERCIAL use only; each user accepts its licence in the node panel before it downloads, and a studio policy can forbid it (see [Research pack](/docs/concepts/research-pack))

| Model | Task | Tier | Licence | GPU memory | Download |
|---|---|---|---|---|---|
| [BiRefNet HR matting (automatic, 2048 px)](#birefnet-hr-matting) | matting | bundled | [MIT](https://huggingface.co/ZhengPeng7/BiRefNet_HR-matting) |  | 0.44 GB |
| [CUDA 12.8 compiler (nvcc + runtime + CCCL headers) for SplatTrain's one-time gsplat build](#cuda-nvcc-12-8) | toolchain | bundled | [NVIDIA CUDA Toolkit EULA](https://docs.nvidia.com/cuda/eula/index.html) |  | 0.08 GB |
| [DINOv2 ViT-S/14 (image embeddings for Prep Show key shots)](#dinov2-small) | image_embedding | bundled | [Apache-2.0](https://huggingface.co/facebook/dinov2-small) | 0.5 GB | 0.09 GB |
| [EdgeTAM (fast preview)](#edgetam) | video_segmentation | bundled | [Apache-2.0](https://huggingface.co/yonigozlan/EdgeTAM-hf) |  | 0.06 GB |
| [EfficientTAM-S (light video tracker, ViT-S image encoder, 34M params)](#efficienttam-s) | video_segmentation | bundled | [Apache-2.0](https://github.com/yformer/EfficientTAM/blob/abcd061ebd3cc6e7527d152d75b890126aaa53f6/LICENSE) | 2 GB | 0.25 GB |
| [EfficientTAM-TI (light video tracker, ViT-Ti image encoder, 18M params)](#efficienttam-ti) | video_segmentation | bundled | [Apache-2.0](https://github.com/yformer/EfficientTAM/blob/abcd061ebd3cc6e7527d152d75b890126aaa53f6/LICENSE) | 2 GB | 0.19 GB |
| [FLUX.2 [klein] 4B (distilled, 4 steps, inpaint + reference)](#flux2-klein-4b) | generate | bundled | [Apache-2.0](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B) | 10 GB | 16.0 GB |
| [Grounding DINO Base](#grounding-dino-base) | text_detection | bundled | [Apache-2.0](https://huggingface.co/IDEA-Research/grounding-dino-base) |  | 0.93 GB |
| [MapAnything (Apache checkpoint): multi-view focal, camera poses, metric depth](#map-anything-apache) | camera_prior | bundled | [Apache-2.0](https://huggingface.co/facebook/map-anything-apache) | 12 GB | 4.9 GB |
| [MoGe-2 ViT-B (metric geometry, FOV, normals)](#moge-2-vitb-normal) | monocular_geometry | bundled | [MIT](https://huggingface.co/Ruicheng/moge-2-vitb-normal) |  | 0.42 GB |
| [Qwen-Image-Edit 20B (Aug 2025) Q8_0 GGUF + Lightning 8-step LoRA](#qwen-image-edit-q8) | generate | bundled | [Apache-2.0](https://huggingface.co/Qwen/Qwen-Image-Edit) | 8 GB | 40.3 GB |
| [Real-ESRGAN x2plus (fast, per frame)](#realesrgan-x2plus) | upscale | bundled | [BSD-3-Clause](https://github.com/xinntao/Real-ESRGAN/blob/master/LICENSE) |  | 0.07 GB |
| [Real-ESRGAN x4plus (fast, per frame)](#realesrgan-x4plus) | upscale | bundled | [BSD-3-Clause](https://github.com/xinntao/Real-ESRGAN/blob/master/LICENSE) |  | 0.07 GB |
| [RIFE 4.26 (frame interpolation)](#rife-4-26) | frame_interpolation | bundled | [MIT](https://github.com/hzwer/Practical-RIFE/blob/main/LICENSE) |  | 0.02 GB |
| [Robust Video Matting (humans, no prompt)](#rvm-mobilenetv3) | matting | bundled | [GPL-3.0](https://github.com/PeterL1n/RobustVideoMatting/blob/master/LICENSE) |  | 0.02 GB |
| [SAM 2.1 Hiera Large](#sam2-1-hiera-large) | video_segmentation | bundled | [Apache-2.0](https://huggingface.co/facebook/sam2.1-hiera-large) |  | 0.90 GB |
| [SEA-RAFT M (optical flow)](#sea-raft-m) | optical_flow | bundled | [BSD-3-Clause](https://github.com/princeton-vl/SEA-RAFT/blob/main/LICENSE) |  | 0.08 GB |
| [SEA-RAFT S (optical flow, fast)](#sea-raft-s) | optical_flow | bundled | [BSD-3-Clause](https://github.com/princeton-vl/SEA-RAFT/blob/main/LICENSE) |  | 0.04 GB |
| [SeedVR2 3B (one-step video restoration, temporal)](#seedvr2-3b) | upscale | bundled | [Apache-2.0](https://huggingface.co/ByteDance-Seed/SeedVR2-3B) |  | 3.9 GB |
| [Shot-trained upscaler (fidelity, trained on your plate)](#shot-sr) | upscale | bundled | [GPL-3.0](https://www.gnu.org/licenses/gpl-3.0.html) | 2 GB |  |
| [TAPNext++ (point tracking through occlusion, re-detection)](#tapnextpp) | point_tracking | bundled | [Apache-2.0](https://github.com/google-deepmind/tapnet#license-and-disclaimer) | 4 GB | 2.5 GB |
| [Video Depth Anything Small (temporal relative depth)](#video-depth-anything-small) | video_depth | bundled | [Apache-2.0](https://huggingface.co/depth-anything/Video-Depth-Anything-Small) |  | 0.12 GB |
| [ViTMatte base (trimap, Composition-1k)](#vitmatte-base) | matting | bundled | [Apache-2.0](https://huggingface.co/hustvl/vitmatte-base-composition-1k) |  | 0.39 GB |
| [ViTMatte small (trimap, Composition-1k)](#vitmatte-small) | matting | bundled | [Apache-2.0](https://huggingface.co/hustvl/vitmatte-small-composition-1k) |  | 0.10 GB |
| [Wan 2.1 VACE 1.3B (video inpainting, ComfyUI split files)](#wan2-1-vace-1-3b) | video_inpaint | bundled | [Apache-2.0](https://huggingface.co/Wan-AI/Wan2.1-VACE-1.3B) | 13 GB | 11.3 GB |
| [Wan 2.2 I2V A14B (MoE high/low noise, fp8 scaled)](#wan2-2-i2v-a14b) | generate | bundled | [Apache-2.0](https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B) | 20 GB | 35.6 GB |
| [Wan 2.2 TI2V 5B (text/image to video, 720p 24 fps)](#wan2-2-ti2v-5b) | generate | bundled | [Apache-2.0](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) | 20 GB | 19.6 GB |
| [Z-Image Turbo (6B, 8-step text/img2img, ComfyUI split files)](#z-image-turbo) | generate | bundled | [Apache-2.0](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) | 15 GB | 20.7 GB |
| [Cosmos DiffusionRenderer 7B (forward renderer: G-buffers + HDR environment map -> relit video)](#cosmos-diffusion-renderer-forward) | relight | optional | [NVIDIA-Open-Model-License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/) | 18.7 GB | 29.2 GB |
| [Cosmos DiffusionRenderer 7B (inverse renderer: video -> base colour, normals, depth, roughness, metallic)](#cosmos-diffusion-renderer-inverse) | material_passes | optional | [NVIDIA-Open-Model-License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/) | 18.7 GB | 29.2 GB |
| [LTX-2.3 22B distilled (fp8, video + audio)](#ltx-2-3-22b-distilled) | generate | optional | [LTX-2 Community License](https://huggingface.co/Lightricks/LTX-2.3/blob/main/LICENSE) | 22 GB | 37.8 GB |
| [MoGe-2 ViT-L (metric geometry, FOV, normals; larger)](#moge-2-vitl-normal) | monocular_geometry | optional | [MIT](https://huggingface.co/Ruicheng/moge-2-vitl-normal) |  | 1.3 GB |
| [SAM 3 (text prompts + points / box / mask, per-object tracker)](#sam3) | text_video_segmentation | optional | [SAM-License](https://huggingface.co/facebook/sam3) | 8 GB | 3.5 GB |
| [SAM 3.1 (text prompts, Object Multiplex joint multi-object tracking)](#sam3-1) | text_video_segmentation | optional | [SAM-License](https://huggingface.co/facebook/sam3.1) | 8 GB | 3.6 GB |
| [FLUX.2 [dev] 32B fp8mixed (non-commercial)](#flux2-dev) | generate | research | [FLUX [dev] Non-Commercial License](https://huggingface.co/black-forest-labs/FLUX.2-dev/blob/main/LICENSE.md) | 22 GB | 53.8 GB |
| [MatAnyone 2 (video matting from a rough mask) — RESEARCH PACK, non-commercial](#matanyone2) | video_matting | research | [S-Lab-License-1.0 (non-commercial)](https://github.com/pq-yang/MatAnyone2/blob/main/LICENSE.txt) | 5 GB | 0.16 GB |
| [MiniMax-Remover (video object removal) — RESEARCH PACK, non-commercial](#minimax-remover) | video_inpainting | research | [CC-BY-NC-4.0 (weights); code: no licence](https://huggingface.co/zibojia/minimax-remover) | 16 GB | 2.8 GB |

## birefnet-hr-matting

**BiRefNet HR matting (automatic, 2048 px)**, licence `MIT`, tier bundled.

Provenance: Zheng Peng et al., BiRefNet (CAAI AIR 2024), Swin-L backbone, HR matting fine-tune at 2048x2048. Trained on academic matting datasets (P3M-10k, AM-2k, Distinctions-646, TR-humans etc.) — dataset provenance flagged for legal review. Needs the repo's birefnet.py (MIT, trust_remote_code; timm.layers shimmed by opencomp_ai/ops/matte.py).

Weights: Hugging Face `ZhengPeng7/BiRefNet_HR-matting` at `5d6b6f8adcb5`. Registry file `opencomp_ai/models/birefnet-hr-matting.json`.

## cuda-nvcc-12.8

**CUDA 12.8 compiler (nvcc + runtime + CCCL headers) for SplatTrain's one-time gsplat build**, licence `NVIDIA CUDA Toolkit EULA`, tier bundled.

Provenance: NVIDIA CUDA 12.8.1 redistributable archives (developer.download.nvidia.com/compute/cuda/redist, redistrib_12.8.1.json). Downloaded on first SplatTrain run, never shipped with Oink. Matches the torch cu128 build.

## dinov2-small

**DINOv2 ViT-S/14 (image embeddings for Prep Show key shots)**, licence `Apache-2.0`, tier bundled.

Provenance: Meta AI DINOv2 (Oquab et al. 2023), self-supervised ViT-S/14 trained on LVD-142M (curated web images, no labels).

Weights: Hugging Face `facebook/dinov2-small` at `ed25f3a31f01`. Registry file `opencomp_ai/models/dinov2-small.json`.

## edgetam

**EdgeTAM (fast preview)**, licence `Apache-2.0`, tier bundled.

Provenance: Meta Reality Labs EdgeTAM (SAM 2 distilled to a RepViT encoder), HF-transformers conversion by the transformers team (yonigozlan). Trained on SA-1B + SA-V.

Weights: Hugging Face `yonigozlan/EdgeTAM-hf` at `c266ce53b3fc`. Registry file `opencomp_ai/models/edgetam.json`.

## efficienttam-s

**EfficientTAM-S (light video tracker, ViT-S image encoder, 34M params)**, licence `Apache-2.0`, tier bundled.

Apache-2.0 for the code and the checkpoints (README: 'Efficient track anything checkpoints and codebase are licensed under Apache 2.0'). Commercial use allowed; keep the licence + NOTICE with redistributions.

Provenance: Meta / yformer EfficientTAM (Efficient Track Anything, arXiv:2411.18933): SAM 2 architecture with a plain non-hierarchical ViT image encoder and efficient memory cross-attention, trained on SA-1B + SA-V like SAM 2. Weights: HF yunyangx/efficient-track-anything. Code: the official repo at a pinned commit, downloaded as a tarball next to the weights, never vendored.

Weights: Hugging Face `yunyangx/efficient-track-anything` at `9bdd8ab585b1`. Registry file `opencomp_ai/models/efficienttam-s.json`.

## efficienttam-ti

**EfficientTAM-TI (light video tracker, ViT-Ti image encoder, 18M params)**, licence `Apache-2.0`, tier bundled.

Apache-2.0 for the code and the checkpoints (README: 'Efficient track anything checkpoints and codebase are licensed under Apache 2.0'). Commercial use allowed; keep the licence + NOTICE with redistributions.

Provenance: Meta / yformer EfficientTAM (Efficient Track Anything, arXiv:2411.18933): SAM 2 architecture with a plain non-hierarchical ViT image encoder and efficient memory cross-attention, trained on SA-1B + SA-V like SAM 2. Weights: HF yunyangx/efficient-track-anything. Code: the official repo at a pinned commit, downloaded as a tarball next to the weights, never vendored.

Weights: Hugging Face `yunyangx/efficient-track-anything` at `9bdd8ab585b1`. Registry file `opencomp_ai/models/efficienttam-ti.json`.

## flux2-klein-4b

**FLUX.2 [klein] 4B (distilled, 4 steps, inpaint + reference)**, licence `Apache-2.0`, tier bundled.

Provenance: black-forest-labs/FLUX.2-klein-4B @e7b7dc2 (Apache-2.0), diffusers transformer + VAE downloaded once to the shared cache; its Qwen3-4B text encoder is byte-identical to the ComfyUI qwen_3_4b.safetensors (checked tensor by tensor) and is reused from there.

Weights: Hugging Face `black-forest-labs/FLUX.2-klein-4B` at `e7b7dc27f91d`. Registry file `opencomp_ai/models/flux2-klein-4b.json`.

## grounding-dino-base

**Grounding DINO Base**, licence `Apache-2.0`, tier bundled.

Provenance: IDEA Research. Trained on O365, GoldG, Cap4M, COCO, RefCOCO (academic/web datasets; see paper).

Weights: Hugging Face `IDEA-Research/grounding-dino-base` at `12bdfa3120f3`. Registry file `opencomp_ai/models/grounding-dino-base.json`.

## map-anything-apache

**MapAnything (Apache checkpoint): multi-view focal, camera poses, metric depth**, licence `Apache-2.0`, tier bundled.

Checkpoint Apache-2.0 (the 'apache' variant, trained only on data whose licences allow it; the default facebook/map-anything checkpoint is CC-BY-NC and NOT used). Code: map-anything Apache-2.0, uniception BSD-3-Clause, DINOv2 code Apache-2.0.

Provenance: Meta Reality Labs / CMU MapAnything (Keetha et al. 2025, arXiv:2509.13414). DINOv2-giant encoder + alternating-attention transformer, trained on the Apache-compatible subset of its 13 datasets. Code pinned: facebookresearch/map-anything@3d10cf7, uniception 0.1.7 (PyPI wheel), facebookresearch/dinov2@7764ea0 (hub code only, no DINOv2 weights downloaded).

Weights: Hugging Face `facebook/map-anything-apache` at `00f9c245bbcb`. Registry file `opencomp_ai/models/map-anything-apache.json`.

## moge-2-vitb-normal

**MoGe-2 ViT-B (metric geometry, FOV, normals)**, licence `MIT`, tier bundled.

MIT (model card metadata; code github.com/microsoft/MoGe MIT). DINOv2 backbone code Apache-2.0.

Provenance: Microsoft Research MoGe-2 (Wang et al. 2025), DINOv2 ViT-B/14 encoder fine-tuned; trained on a mix of synthetic and real metric datasets. Code vendored in opencomp_ai/vendor/moge (MIT).

Weights: Hugging Face `Ruicheng/moge-2-vitb-normal` at `ca5f0e07ff01`. Registry file `opencomp_ai/models/moge-2-vitb-normal.json`.

## qwen-image-edit-q8

**Qwen-Image-Edit 20B (Aug 2025) Q8_0 GGUF + Lightning 8-step LoRA**, licence `Apache-2.0`, tier bundled.

Provenance: Qwen/Qwen-Image-Edit (Aug 2025) MMDiT, Q8_0 GGUF by QuantStack (Qwen-Image-Edit-GGUF @acab6f9); text encoder Qwen2.5-VL-7B bf16 and VAE as repackaged by Comfy-Org; optional lightx2v Qwen-Image-Edit-Lightning 8-step LoRA (Apache-2.0). Big files reused from the ComfyUI models dir (external). Configs pinned from Qwen/Qwen-Image-Edit.

Weights: Hugging Face `Qwen/Qwen-Image-Edit` at `ac7f9318f633`. Registry file `opencomp_ai/models/qwen-image-edit-q8.json`.

## realesrgan-x2plus

**Real-ESRGAN x2plus (fast, per frame)**, licence `BSD-3-Clause`, tier bundled.

Provenance: Real-ESRGAN (Wang et al., Tencent ARC, 2021), RRDBNet generator trained on DIV2K + Flickr2K + OutdoorSceneTraining with synthetic degradations. Weights from the official GitHub release.

## realesrgan-x4plus

**Real-ESRGAN x4plus (fast, per frame)**, licence `BSD-3-Clause`, tier bundled.

Provenance: Real-ESRGAN (Wang et al., Tencent ARC, 2021), RRDBNet generator trained on DIV2K + Flickr2K + OutdoorSceneTraining with synthetic degradations. Weights from the official GitHub release.

## rife-4.26

**RIFE 4.26 (frame interpolation)**, licence `MIT`, tier bundled.

Provenance: hzwer Practical-RIFE v4.26 (2024-09-21). Official release is on Google Drive; this HF mirror (lopi999/rife_v4.26) is byte-identical to it (sha256 of flownet.pkl checked against the official zip, 2026-09-25). Trained on Vimeo90K + in-house video (per the authors).

Weights: Hugging Face `lopi999/rife_v4.26` at `93b1dd7b5447`. Registry file `opencomp_ai/models/rife-4.26.json`.

## rvm-mobilenetv3

**Robust Video Matting (humans, no prompt)**, licence `GPL-3.0`, tier bundled.

Provenance: Lin et al., Robust High-Resolution Video Matting with Temporal Guidance (WACV 2022), MobileNetV3 variant. Official weights are on GitHub releases (v1.0.0); this HF mirror serves the same file (sha256 pinned). Model code vendored in opencomp_ai/vendor/rvm (GPL-3). Trained on VideoMatte240K, Distinctions-646, AM-2k, COCO/YouTubeVIS segmentation.

Weights: Hugging Face `SJTU-TES/RobustVideoMatting` at `3c88b194e7e4`. Registry file `opencomp_ai/models/rvm-mobilenetv3.json`.

## sam2.1-hiera-large

**SAM 2.1 Hiera Large**, licence `Apache-2.0`, tier bundled.

Provenance: Meta FAIR. Trained on SA-1B + SA-V (Meta, licensed/collected video) per the SAM 2 paper.

Weights: Hugging Face `facebook/sam2.1-hiera-large` at `665f8e2ad61c`. Registry file `opencomp_ai/models/sam2.1-hiera-large.json`.

## sea-raft-m

**SEA-RAFT M (optical flow)**, licence `BSD-3-Clause`, tier bundled.

Provenance: Princeton Vision & Learning Lab, SEA-RAFT (ECCV 2024) medium, checkpoint Tartan-C-T-TSKH-spring540x960-M. Trained on TartanAir, FlyingChairs, FlyingThings3D, Sintel, KITTI, HD1K, Spring (synthetic + academic benchmark data).

Weights: Hugging Face `MemorySlices/Tartan-C-T-TSKH-spring540x960-M` at `eb97ef34ba5d`. Registry file `opencomp_ai/models/sea-raft-m.json`.

## sea-raft-s

**SEA-RAFT S (optical flow, fast)**, licence `BSD-3-Clause`, tier bundled.

Provenance: Princeton Vision & Learning Lab, SEA-RAFT (ECCV 2024) small, checkpoint Tartan-C-T-TSKH-spring540x960-S. Same training data as sea-raft-m.

Weights: Hugging Face `MemorySlices/Tartan-C-T-TSKH-spring540x960-S` at `31b9b4b711bd`. Registry file `opencomp_ai/models/sea-raft-s.json`.

## seedvr2-3b

**SeedVR2 3B (one-step video restoration, temporal)**, licence `Apache-2.0`, tier bundled.

Provenance: ByteDance Seed SeedVR2-3B (one-step diffusion video restoration, adversarial post-training). FP8 e4m3fn DiT + fp16 VAE repack by numz (ComfyUI-SeedVR2_VideoUpscaler, Apache-2.0); inference code pinned from that repo. Training data: undisclosed large-scale image/video corpus (see paper arXiv:2506.05301).

Weights: Hugging Face `numz/SeedVR2_comfyUI` at `09ced7102363`. Registry file `opencomp_ai/models/seedvr2-3b.json`.

## shot-sr

**Shot-trained upscaler (fidelity, trained on your plate)**, licence `GPL-3.0`, tier bundled.

Oink's own code (opencomp_ai/ops/upscale.py). No downloaded weights: a small residual CNN is trained on the shot itself, on this machine.

Provenance: Zero-shot super-resolution (Shocher et al. 2018, method only): trained on the plate's own frame (downscaled -> Lanczos -> the frame). No external training data; the weights stay in the local cache (shot-sr/).

## tapnextpp

**TAPNext++ (point tracking through occlusion, re-detection)**, licence `Apache-2.0`, tier bundled.

Code and all released checkpoints (TAP-Net, TAPIR, BootsTAPIR, TAPNext, TAPNext++) Apache-2.0 per the tapnet README.

Provenance: Google DeepMind TAPNext++ (BootsTAPNext fine-tuned on PointOdyssey v1.2 + Kubric-1024 synthetic sequences; BootsTAP self-training used real YouTube video without labels). Torch checkpoint from storage.googleapis.com/dm-tapnet; code pinned google-deepmind/tapnet@730cda1.

## video-depth-anything-small

**Video Depth Anything Small (temporal relative depth)**, licence `Apache-2.0`, tier bundled.

Apache-2.0 (model card metadata + upstream README: 'Video-Depth-Anything-Small model is under the Apache-2.0 license'; Base/Large are CC-BY-NC and are not listed).

Provenance: ByteDance Video Depth Anything (Chen et al. 2025), DINOv2 ViT-S encoder + temporal DPT head; trained on synthetic video depth + Depth Anything V2 pseudo-labelled images. Code vendored in opencomp_ai/vendor/vda (Apache-2.0).

Weights: Hugging Face `depth-anything/Video-Depth-Anything-Small` at `256875362cff`. Registry file `opencomp_ai/models/video-depth-anything-small.json`.

## vitmatte-base

**ViTMatte base (trimap, Composition-1k)**, licence `Apache-2.0`, tier bundled.

Provenance: HUST VL ViTMatte (Yao et al. 2023), ViT-B DINO backbone, trained on Adobe Composition-1k (academic dataset; flag for legal review). Only a pickle .bin is published (loaded with torch weights_only).

Weights: Hugging Face `hustvl/vitmatte-base-composition-1k` at `bf486d01a7d9`. Registry file `opencomp_ai/models/vitmatte-base.json`.

## vitmatte-small

**ViTMatte small (trimap, Composition-1k)**, licence `Apache-2.0`, tier bundled.

Provenance: HUST VL ViTMatte (Yao et al. 2023), ViT-S DINO backbone, trained on Adobe Composition-1k (academic dataset; flag for legal review). HF transformers VitMatteForImageMatting.

Weights: Hugging Face `hustvl/vitmatte-small-composition-1k` at `6a58ad764640`. Registry file `opencomp_ai/models/vitmatte-small.json`.

## wan2.1-vace-1.3b

**Wan 2.1 VACE 1.3B (video inpainting, ComfyUI split files)**, licence `Apache-2.0`, tier bundled.

Provenance: Wan 2.1 VACE 1.3B (Alibaba Tongyi Wan team, 2025): all-in-one video creation/editing DiT (masked video-to-video). umT5-XXL text encoder (fp8 scaled), Wan 2.1 VAE. Weights as repackaged by Comfy-Org (fp16 DiT), configs/tokenizer from Wan-AI/Wan2.1-VACE-1.3B-diffusers; reused from the user's ComfyUI models dir when present. Training data: undisclosed large-scale image/video corpus (Wan tech report arXiv:2503.20314).

Weights: Hugging Face `Comfy-Org/Wan_2.1_ComfyUI_repackaged` at `123acf1cc74b`. Registry file `opencomp_ai/models/wan2.1-vace-1.3b.json`.

## wan2.2-i2v-a14b

**Wan 2.2 I2V A14B (MoE high/low noise, fp8 scaled)**, licence `Apache-2.0`, tier bundled.

Provenance: Wan-AI/Wan2.2-I2V-A14B (Apache-2.0): two 14B experts (high / low noise) in ComfyUI fp8 e4m3 scaled form (weights x per-tensor scale, upcast per layer), Wan 2.1 VAE and umt5-xxl fp8-scaled as repackaged by Comfy-Org (reused from the ComfyUI models dir); diffusers configs pinned from Wan-AI/Wan2.2-I2V-A14B-Diffusers.

Weights: Hugging Face `Wan-AI/Wan2.2-I2V-A14B-Diffusers` at `596658fd9ca6`. Registry file `opencomp_ai/models/wan2.2-i2v-a14b.json`.

## wan2.2-ti2v-5b

**Wan 2.2 TI2V 5B (text/image to video, 720p 24 fps)**, licence `Apache-2.0`, tier bundled.

Provenance: Wan-AI/Wan2.2-TI2V-5B (Apache-2.0): fp16 DiT, Wan 2.2 VAE (16x16x4) and umt5-xxl fp8-scaled text encoder as repackaged by Comfy-Org (reused from the ComfyUI models dir); diffusers configs pinned from Wan-AI/Wan2.2-TI2V-5B-Diffusers. The Wan2.2 VAE comes from the diffusers repo (diffusers 0.39 cannot convert the ComfyUI wan2.2_vae single file: residual decoder layout).

Weights: Hugging Face `Wan-AI/Wan2.2-TI2V-5B-Diffusers` at `b8fff7315c76`. Registry file `opencomp_ai/models/wan2.2-ti2v-5b.json`.

## z-image-turbo

**Z-Image Turbo (6B, 8-step text/img2img, ComfyUI split files)**, licence `Apache-2.0`, tier bundled.

Provenance: Z-Image Turbo (Tongyi-MAI / Alibaba, 2025), distilled 6B single-stream DiT; Qwen3-4B text encoder; FLUX.1 autoencoder (ae.safetensors, Apache-2.0). Files as repackaged by Comfy-Org; reused from the user's ComfyUI models dir when present (external), hashes = Hugging Face LFS sha256.

Weights: Hugging Face `Comfy-Org/z_image_turbo` at `6fc90a3b1b65`. Registry file `opencomp_ai/models/z-image-turbo.json`.

## cosmos-diffusion-renderer-forward

**Cosmos DiffusionRenderer 7B (forward renderer: G-buffers + HDR environment map -> relit video)**, licence `NVIDIA-Open-Model-License`, tier optional.

NVIDIA Open Model License: commercial use permitted; you may create and distribute derivative models; NVIDIA claims no ownership of outputs. Conditions: keep the notice 'Licensed by NVIDIA Corporation under the NVIDIA Open Model License'; rights terminate if you bypass or disable the model's safety guardrails without a substantially similar replacement, or if you start patent/copyright litigation over the model. Weights are not redistributed by Oink. Read the full text at the licence URL before accepting.

Provenance: NVIDIA Toronto AI Lab, Cosmos-Transfer1-DiffusionRenderer (Liang et al., CVPR 2025, arXiv:2501.18590). 7B Cosmos-Predict1 DiT fine-tuned; trained on 150k synthetic videos (rendered G-buffers) + 150k auto-labelled real videos, 57 frames at 1280x704. Latent tokenizer: nvidia/Cosmos-Tokenize1-CV8x8x8-720p (same licence). Inference code (Apache-2.0) re-implemented in pure PyTorch in opencomp_ai/vendor/cosmos_dr (no transformer-engine / megatron).

Weights: Hugging Face `nvidia/Diffusion_Renderer_Forward_Cosmos_7B` at `7d5f9b81affa`. Registry file `opencomp_ai/models/cosmos-diffusion-renderer-forward.json`.

## cosmos-diffusion-renderer-inverse

**Cosmos DiffusionRenderer 7B (inverse renderer: video -> base colour, normals, depth, roughness, metallic)**, licence `NVIDIA-Open-Model-License`, tier optional.

NVIDIA Open Model License: commercial use permitted; you may create and distribute derivative models; NVIDIA claims no ownership of outputs. Conditions: keep the notice 'Licensed by NVIDIA Corporation under the NVIDIA Open Model License'; rights terminate if you bypass or disable the model's safety guardrails without a substantially similar replacement, or if you start patent/copyright litigation over the model. Weights are not redistributed by Oink. Read the full text at the licence URL before accepting.

Provenance: NVIDIA Toronto AI Lab, Cosmos-Transfer1-DiffusionRenderer (Liang et al., CVPR 2025, arXiv:2501.18590). 7B Cosmos-Predict1 DiT fine-tuned; trained on 150k synthetic videos (rendered G-buffers) + 150k auto-labelled real videos, 57 frames at 1280x704. Latent tokenizer: nvidia/Cosmos-Tokenize1-CV8x8x8-720p (same licence). Inference code (Apache-2.0) re-implemented in pure PyTorch in opencomp_ai/vendor/cosmos_dr (no transformer-engine / megatron).

Weights: Hugging Face `nvidia/Diffusion_Renderer_Inverse_Cosmos_7B` at `41f63abccfb7`. Registry file `opencomp_ai/models/cosmos-diffusion-renderer-inverse.json`.

## ltx-2.3-22b-distilled

**LTX-2.3 22B distilled (fp8, video + audio)**, licence `LTX-2 Community License`, tier optional.

Free for entities with < $10M annual revenue; above that a commercial licence from Lightricks is required. Acceptable-use restrictions apply. You must accept before the weights are used.

Provenance: Lightricks/LTX-2.3 @5948be4 22B distilled DiT (fp8 v2 as used by ComfyUI) + LTX-2.3 video VAE, reused from the ComfyUI models dir; text encoder Gemma-3-12B-it (Google Gemma terms) from Comfy-Org/ltx-2.

Weights: Hugging Face `Lightricks/LTX-2.3` at `5948be4ced3a`. Registry file `opencomp_ai/models/ltx-2.3-22b-distilled.json`.

## moge-2-vitl-normal

**MoGe-2 ViT-L (metric geometry, FOV, normals; larger)**, licence `MIT`, tier optional.

MIT (model card metadata; code github.com/microsoft/MoGe MIT). DINOv2 backbone code Apache-2.0.

Provenance: Microsoft Research MoGe-2 (Wang et al. 2025), DINOv2 ViT-L/14 encoder fine-tuned; the larger sibling of moge-2-vitb-normal. Code vendored in opencomp_ai/vendor/moge (MIT).

Weights: Hugging Face `Ruicheng/moge-2-vitl-normal` at `cb0e8bbd6b1e`. Registry file `opencomp_ai/models/moge-2-vitl-normal.json`.

## sam3

**SAM 3 (text prompts + points / box / mask, per-object tracker)**, licence `SAM-License`, tier optional.

Meta SAM License (custom, not OSI). Allows: use, copy, modify, distribute, derivative works, commercial use (royalty-free, no user/revenue cap). Conditions: redistribution only under the same agreement with a copy of it; acknowledge SAM in published research; comply with laws, trade controls and privacy law; no ITAR / military / warfare / nuclear / espionage / weapons use; do not reverse engineer the SAM Materials; indemnify Meta; licence ends if you sue Meta over SAM IP; Meta may amend the terms. Weights and code are downloaded to your machine after you accept; Oink never redistributes them. Read the full text at the licence URL before accepting.

Provenance: Meta FAIR SAM 3 (Segment Anything with Concepts, arXiv:2511.16719), 848M params. Gated on Hugging Face (manual approval). Trained on SA-Co (Meta) + SA-1B + SA-V. Inference code: official facebookresearch/sam3 at a pinned commit, fetched as a tarball (same SAM License), never vendored.

Weights: Hugging Face `facebook/sam3` at `3c879f39826c`. Registry file `opencomp_ai/models/sam3.json`.

## sam3.1

**SAM 3.1 (text prompts, Object Multiplex joint multi-object tracking)**, licence `SAM-License`, tier optional.

Meta SAM License (custom, not OSI). Allows: use, copy, modify, distribute, derivative works, commercial use (royalty-free, no user/revenue cap). Conditions: redistribution only under the same agreement with a copy of it; acknowledge SAM in published research; comply with laws, trade controls and privacy law; no ITAR / military / warfare / nuclear / espionage / weapons use; do not reverse engineer the SAM Materials; indemnify Meta; licence ends if you sue Meta over SAM IP; Meta may amend the terms. Weights and code are downloaded to your machine after you accept; Oink never redistributes them. Read the full text at the licence URL before accepting.

Provenance: Meta FAIR SAM 3.1 (SAM 3 + Object Multiplex: up to 16 objects per shared-memory bucket). Gated on Hugging Face (manual approval). Inference code: official facebookresearch/sam3 at a pinned commit, fetched as a tarball (same SAM License), never vendored. Checkpoint only on HF (no transformers integration).

Weights: Hugging Face `facebook/sam3.1` at `daa63191845a`. Registry file `opencomp_ai/models/sam3.1.json`.

## flux2-dev

**FLUX.2 [dev] 32B fp8mixed (non-commercial)**, licence `FLUX [dev] Non-Commercial License`, tier research.

Non-commercial use only (research / evaluation). Outputs may not be used commercially without a paid licence from Black Forest Labs. Hidden unless the studio enables the research tier.

Provenance: black-forest-labs/FLUX.2-dev as repackaged by Comfy-Org (fp8mixed transformer, Mistral-3-Small fp8 text encoder, VAE); reused from the ComfyUI models dir (external).

Weights: Hugging Face `Comfy-Org/flux2-dev` at `ed33133cd564`. Registry file `opencomp_ai/models/flux2-dev.json`.

## matanyone2

**MatAnyone 2 (video matting from a rough mask) — RESEARCH PACK, non-commercial**, licence `S-Lab-License-1.0 (non-commercial)`, tier research.

NON-COMMERCIAL USE ONLY. NTU S-Lab License 1.0 covers the code and the weights: redistribution and use are permitted for non-commercial purposes only, keeping the copyright notice; commercial use needs written permission from the authors (Peiqing Yang, Shangchen Zhou, Chen Change Loy, NTU S-Lab). Not usable on paid / client work. Code and weights download to your machine after you accept; Oink never bundles or redistributes them.

Provenance: NTU S-Lab MatAnyone 2 (Scaling Video Matting via a Learned Quality Evaluator, arXiv:2512.11782). Weights: Hugging Face PeiqingYang/MatAnyone2 (model.safetensors 141 MB; the repo's LICENSE file is empty, the licence is the GitHub LICENSE.txt). Inference code: pq-yang/MatAnyone2 at a pinned commit, fetched as a tarball after acceptance, never vendored.

Weights: Hugging Face `PeiqingYang/MatAnyone2` at `40c894a6f68d`. Registry file `opencomp_ai/models/matanyone2.json`.

## minimax-remover

**MiniMax-Remover (video object removal) — RESEARCH PACK, non-commercial**, licence `CC-BY-NC-4.0 (weights); code: no licence`, tier research.

NON-COMMERCIAL USE ONLY. The weights (Hugging Face zibojia/minimax-remover) are Creative Commons Attribution-NonCommercial 4.0: share and adapt for non-commercial purposes, with attribution (Bojia Zi, Weixuan Peng et al., arXiv:2505.24873). The inference code (github.com/zibojia/MiniMax-Remover) has NO licence file, so its authors reserve all rights; it is fetched to your machine for your own non-commercial use only. Built on the Wan 2.1 architecture/VAE (Apache-2.0). Not usable on paid / client work.

Provenance: MiniMax-Remover (Zi, Peng, Qi, Wang, Zhao, Xiao, Wong; arXiv:2505.24873): simplified Wan 2.1 1.3B DiT without text conditioning, minimax-distilled for 6-12 steps without CFG. Training data per the paper: public video datasets with synthetic masks (not audited). Code: two files at a pinned commit, fetched after acceptance, never vendored.

Weights: Hugging Face `zibojia/minimax-remover` at `889e41651d90`. Registry file `opencomp_ai/models/minimax-remover.json`.


---
Sources: opencomp_ai/models