Video models — companion files
Text encoders and a VAE for the Off Grid video GGUF generators with stable-diffusion.cpp. These files do not generate video by themselves. Use them with one generator from the repositories below.
Files
| File | Role | sd-cli flag | Size | Used by |
|---|---|---|---|---|
umt5-xxl-encoder-Q4_K_M.gguf |
UMT5-XXL text encoder | --t5xxl |
3.655 GB | Wan2.1-T2V-1.3B-GGUF, Wan2.2-TI2V-5B-GGUF |
wan_2.1_vae.safetensors |
Wan 2.1 VAE | --vae |
0.254 GB | Wan2.1-T2V-1.3B-GGUF, LingBot-Video-Dense-1.3B-GGUF |
Qwen3VL-4B-Instruct-Q4_K_M.gguf |
Qwen3-VL 4B Instruct text encoder | --llm |
2.497 GB | LingBot-Video-Dense-1.3B-GGUF |
Sizes use decimal GB. Wan 2.2 also needs wan2.2_vae.safetensors, which stays in Wan2.2-TI2V-5B-GGUF because no other model uses it.
All files are unchanged copies of the source files. Checksums are in SHA256SUMS.
umt5-xxl-encoder-Q4_K_M.ggufis from city96/umt5-xxl-encoder-gguf at revisionb535255bee98c2b0a59ea7c0ae2dcd0c6657b3b7.wan_2.1_vae.safetensorsis from Comfy-Org/Wan_2.1_ComfyUI_repackaged at revision123acf1cc74bccbb9bfff8ac1ee72edc08c2341d.Qwen3VL-4B-Instruct-Q4_K_M.ggufis from Qwen/Qwen3-VL-4B-Instruct-GGUF at revision1cd86afb9a95c410a6038ab3b40d8b578c892266.
The Hugging Face GGUF panel shows one GGUF file per repository. Here it describes one of the two text encoders. Neither file is a video generator.
These files were in the generator repositories before 2026-10-21. Links pinned to an earlier commit of those repositories still work.
Use with stable-diffusion.cpp
Download the files that your generator needs into the same folder as the generator. For example, for Wan 2.1:
hf download off-grid-ai/Video-Companions-GGUF umt5-xxl-encoder-Q4_K_M.gguf wan_2.1_vae.safetensors --local-dir .
The generator repositories have complete commands.
Credit & license
UMT5-XXL: google/umt5-xxl by Google; GGUF conversion by city96.
Wan 2.1 VAE: Wan-AI/Wan2.1-T2V-1.3B by the Wan team (Alibaba), repackaged by Comfy-Org.
Qwen3-VL-4B-Instruct: Qwen/Qwen3-VL-4B-Instruct by the Qwen team. The upstream model card is kept in provenance/.
License: apache-2.0 for all three files (carried over from the sources). See NOTICE for attribution. This repository redistributes the files unchanged; all credit belongs to the original creators.