LOCAL PRODUCTION · DEDICATED SERVER
Your own model space Open model options for image, video, text and audio work that can run on your team's own machines or on servers dedicated to you.
This page shows deployment CANDIDATES. Not every model listed here is ready in Velvet, installed, or tested end to end. A chosen model is put into service after it is verified with your devices, your data and your production flow.
Search a model or a capability Production typeAll types Image Video Text Audio Masking DeploymentAll deployments Local machine Dedicated server 6 model
Image Apache 2.0
FLUX.2 [klein] 4B Developing images from references, concept work and image generation.
Examples published by the maker · Click the image to enlarge. Deployment profile Local machine / dedicated server Hardware The maker reports about 13 GB of GPU memory. Precision and resolution change what is needed. Velvet integration: installation and flow still have to be verified
Model card, licence and examples ↗ Image Apache 2.0
Qwen-Image Posters, concept work and text rendered inside an image.
Examples published by the maker · Click the image to enlarge. Deployment profile A dedicated GPU server is recommended Hardware A 20B model; a strong GPU server is recommended. Real capacity is measured with the chosen optimisation. Velvet integration: installation and flow still have to be verified
Model card, licence and examples ↗ Video Apache 2.0
Wan2.2 TI2V-5B 720p, 24 fps video from text or from a reference image.
The Wan demo video from the model card. Which variant produced each shot is not stated. Deployment profile Local machine / dedicated server Hardware The maker's offload example needs at least 24 GB of VRAM. A workstation or a GPU server. Velvet integration: installation and flow still have to be verified
Model card, licence and examples ↗ Text Apache 2.0
Qwen3-8B A language model for script drafts, summaries and structured text.
Script draft → scene list → structured text
A diagram of the output shape, not something the model produced. Real examples are on the model card. Deployment profile Local machine / dedicated server Hardware A local machine or a server. Memory depends on quantisation and context length. Velvet integration: installation and flow still have to be verified
Model card, licence and examples ↗ Audio MIT
Whisper large-v3-turbo Speech to text; the basis for subtitle and interview transcript flows.
Audio recording → transcript → subtitle preparation
A diagram of the output shape, not something the model produced. Real examples are on the model card. Deployment profile Local machine / dedicated server Hardware Runs on CPU or GPU; a GPU is preferred for long recordings. Speed is measured on the device. Velvet integration: installation and flow still have to be verified
Model card, licence and examples ↗ Masking Apache 2.0
SAM 2.1 Hiera Large Segmentation masks for chosen objects in an image or a video.
Image / video + object selection → mask
A diagram of the output shape, not something the model produced. Real examples are on the model card. Deployment profile Local machine / dedicated server Hardware A GPU workstation or server. Video length and resolution affect capacity. Velvet integration: installation and flow still have to be verified
Model card, licence and examples ↗ No model matches these filters.
How does it work with Velvet? The target flow: choose the project and the assets in Velvet, send the job to your own machine or to your dedicated server, and judge the output inside the project. File access, the job queue and moving outputs are configured as part of the chosen deployment.
What do we decide during setup? Job type, GPU memory, how many people work at once and how data has to be kept are weighed together. A local setup aims to keep raw files on the device; a dedicated server aims to share capacity across the team. For offline use, model files and dependencies are prepared in advance.
The specific model versions listed here are published under Apache 2.0 or MIT. Licence notices are kept; other versions of the same model family may carry different terms. Open weights do not mean the training data is fully open. Hardware, hosting and setup costs are assessed separately.
Image and video examples are loaded from the makers' own sources; they are not Velvet output. They need an internet connection and depend on those services being reachable. The text, audio and masking cards describe the shape of the output; verified Velvet examples can be added after pilot productions.
![FLUX.2 [klein] 4B — output examples from the maker's model card](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B/resolve/main/realism.jpg)
