Runway and Kling alternative: AI video on an RTX PRO 6000 Workstation Edition

A self-hosted AI video generator instead of Runway or Kling

A company that pays for Runway or Kling credits can make the same kinds of marketing clips on its own workstation with open video models: Wan 2.2 for text-to-video and image-to-video, LTX-2.5 for clips with sound and for 4K output, and InfiniteTalk for a talking presenter. One NVIDIA RTX PRO 6000 Blackwell Workstation Edition with 96 GB runs the large models at 720p. What the company gives up is access to the newest proprietary models and the convenience of a hosted service; what it gains is a fixed machine with no credits and no uploads.

This comparison is written for the marketing lead, agency owner or managing director who weighs a subscription against a machine in the office. A video model is the trained network that turns a text prompt (text-to-video) or a still picture (image-to-video) into a short clip; "open weights" means its maker publishes the model files, so it runs on your own computer. GPU memory, or VRAM, is the memory on the graphics card and decides which model fits. The model runs inside ComfyUI, a free program in which each kind of clip is a saved workflow with a prompt box. Memory per model is in our guide to video generation GPU requirements, the tiers in more depth in AI video workstation for marketing.

What the cloud providers' terms say

Two points on the providers' own pages matter for this comparison: how usage is counted, and what the provider may do with uploaded material. We describe only what the providers' own pages said when we read them on 30 October 2026, and we leave prices out.

Runway counts usage in credits. Its help centre states that "Gen-4.5 uses 12 credits per second of generated video" and that on the Standard, Pro and Unlimited plans "Credits do not roll over"; they are reset within a day of the billing date. Its terms of use, last updated on 23 October 2026, state: "The Company does not claim ownership of any of your Inputs or Outputs." Runway does not restrict commercial use of outputs, and the user grants it a perpetual, worldwide licence under which inputs and outputs "may be used by the Company to train and improve its AI models, algorithms and related technology". The services are "controlled and offered by Company from its facilities in the United States of America". Organisations on Runway Enterprise Services fall under separate enterprise terms, which we did not review.

Kling's user policy, updated on 21 April 2026, is issued by Kling AI Pte. Ltd. and governed by the laws of Singapore. It says "We do not claim ownership of the Content", grants Kling a licence over inputs and outputs, and lets Kling use inputs to "create, test, improve, train, or otherwise develop" the models behind its services; section 4.7.4 lets users revoke that licence by email.

Both providers leave the ownership of outputs with the user. The difference with a workstation is where the material goes. When the model runs on your machine, the prompts, the reference images (product shots, a presenter's face), the client's footage and the finished clips stay on its drives. The machine needs the internet to download models and software updates; check each program's update and telemetry settings at setup, as our guide to telemetry in self-hosted AI tools explains.

Cloud features and their open counterparts

The table maps the jobs teams buy credits for to open models, with what each model's own page states. All were read on 30 October 2026.

GitHub repository of Wan-Video/Wan2.2 and Hugging Face cards of Wan-AI, Lightricks/LTX-2.5, MeiGen-AI/InfiniteTalk, meituan-longcat/LongCat-Video-Avatar-1.5 and Comfy's SeedVR2 guide, read 30 October 2026. Memory figures are Wan's single-GPU commands with offloading switched on.

Wan 2.2 was released with weights on 28 July 2025; by default its large models make 81 frames at 16 frames per second, about five seconds. LTX-2.5 generates "synchronized, high-fidelity video and audio from text, image, and video inputs"; its examples use 121 frames at 24 frames per second at 1920 × 1088, and for 4K its card gives --spatial-upscalings 2 --width 3840 --height 2176. InfiniteTalk "generates good results from a single image for up to 1 minute"; LongCat-Video-Avatar-1.5 was evaluated in Chinese and English only. Presenter videos are covered in our local alternative to HeyGen, clips from packshots in product videos from photos, and upscaling in AI video upscaling to 4K.

One gap has no open counterpart. The cloud services run their own proprietary models, and those weights are not published. On the open side, the Wan-AI organisation on Hugging Face listed no model after Wan 2.2 when we read it on 30 October 2026; its newest release is Wan2.2-Animate-2-14B, whose weights came out on 7 August 2026.

Licences for client and marketing work

Client work and advertising are commercial use, so the licence of each model is the first filter, before the hardware.

Licence files and model cards on GitHub and Hugging Face, read 30 October 2026. General information, not legal advice.

Wan's repository states: "The models in this repository are licensed under the Apache 2.0 License." It claims no rights over generated content and says users are "fully accountable for your use of the models".

The LTX-2.x Community License, dated 11 August 2026, gives entities below the threshold a royalty-free licence to use the model "for any purpose", subject to the restrictions in its Attachment A. "Entities with annual revenues of at least $10,000,000" are "required to obtain a paid license for any use" beyond its non-commercial purposes, and "all subsidiaries, affiliates, and companies under common Control shall be considered collectively". Testing and evaluation by such a company remain allowed. Lightricks "claims no rights in the Output", and Attachment A forbids publishing generated content "without expressly and intelligibly disclaiming that the information and/or content is machine generated". LTX-2.5 is not part of the setup we deliver; a company that meets its licence terms installs it itself.

A company in the EU cannot use HunyuanVideo 1.5 or MiniMax H3 under their licences, which exclude the EU, so we do not install or recommend them. Your legal department confirms the treatment of any specific use.

You can order the workstation alone, the setup of ComfyUI and Wan 2.2 alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

What a local setup gives up

The newest models. A cloud service can switch you to its latest model the day it launches. Locally you work with the open models that exist, and their quality is judged by eye on your own material: run the same brief through both before you cancel a subscription.

Speed per clip. We found no published time per clip for Wan 2.2 or LTX-2.5 on an RTX PRO card from the model makers or NVIDIA. Our article on a Wan 2.2 video server puts the upper bound at about three five-second clips at 720p per hour on one RTX PRO 6000 Server Edition at Wan's default 40 steps, by our estimate from NVIDIA's peak compute; real runs are slower. Time grows with resolution, length and steps, and a card works on one clip at a time.

Upkeep. ComfyUI published five releases between 13 and 27 October 2026, and new models arrive with new workflows. Someone has to update the programs and keep the saved workflows working.

Hybrid use. A team can also keep both: drafts, variants and volume on the workstation, an occasional hero shot in the cloud. Whatever is uploaded for that shot then falls under the provider's terms quoted above.

Which workstation replaces which seat

GPU memory sets what each tier can run; the number of cards sets how many clips run at the same time. A clip occupies a whole card while it renders, so a card serves one person generating and others reviewing, with further jobs waiting in the queue.

Memory as stated by Wan and our arithmetic for LTX-2.5; card specifications from NVIDIA datasheets as listed on our GPU pages. WS = Workstation Edition, SE = Server Edition.

Entry. The NVIDIA RTX PRO 5000 Blackwell with 72 GB is below the 80 GB Wan states for its own 16-bit command for the large models. It runs the 5B model, which Wan states at 24 GB, and by our file arithmetic also the A14B models from Comfy's FP8 files: two experts of 14.3 GB, the text encoder (which reads the prompt) and the VAE (which turns the result into frames) come to about 35.6 GB, as our guide to product videos from photos shows. Lightricks publishes no memory figure for LTX-2.5. Its published files are 42 GB for the 22B transformer and 26.3 GB for the Gemma 4 12B text encoder in 16-bit; the INT8 files for ComfyUI take 21.5 GB and 15.4 GB, together 36.9 GB, or 34.4 GiB. That fits within the 61.5 GiB our sizing rule counts as usable on a 72 GB card (0.9 × the memory the driver reports, minus 3 GiB), with room for the frames. The NVFP4 transformer, a 4-bit format, is an 18.7 GB file that Lightricks lists for Blackwell GPUs, which include the RTX PRO cards. NVIDIA's blog of 10 March 2026 reports NVFP4 in ComfyUI as 2.5 times faster with 60 per cent less VRAM on GeForce RTX 50 Series cards, with a chart for a GeForce RTX 5090, a consumer card we do not supply; it gives no figure for RTX PRO cards. A 32 GB RTX PRO 4500 suffices for upscaling and the 5B model.

Mid. One 96 GB card, 83.0 GiB of it usable by our rule, holds Wan2.2-T2V-A14B or I2V-A14B at 720p, whose weights take about 64.4 GiB each in BF16 by our Wan server article's arithmetic, or LTX-2.5 in BF16, whose published files take 68.3 GB. Two Workstation Edition cards, at 600 W each, are the realistic maximum for that edition; by the estimate on our RTX PRO 6000 Workstation Edition page, one card draws about 1.1 kW at the wall and two about 1.8 kW. Three or four cards are the Max-Q, the 300 W edition, and only in a chassis built for it with strong directed airflow, as our four-card Max-Q workstation guide explains.

Large. A rack server with four NVIDIA RTX PRO 6000 Blackwell Server Edition cards runs one clip per card for a department, draws about 3.4 kW by our estimate and needs a server room; our four-card Server Edition server covers it, and our AI server for creative agencies adds image models and rendering on the same machine.

Tell us which clips your team makes in a typical month, and we suggest the tier that fits your tasks and budget.

The interface staff work in

ComfyUI is under GPL-3.0 and runs on Windows, Linux and macOS. Its App Mode, which NVIDIA's blog of 10 March 2026 calls App View, exposes a workflow as a simple form in which "Users only need to enter a prompt, adjust simple parameters and hit generate", while the full node editor stays available. Wan 2.2 "has been integrated into ComfyUI", "Official LTX-2.5 workflow templates ship in ComfyUI", and SeedVR2 and NVIDIA's RTX Video Super Resolution are available as nodes for the step to 4K. ComfyUI queues jobs and offers a local API, so a marketing team can submit clips from a browser while one person maintains the workflows.

WanGP, a one-window interface for the same models, has its own licence (WanGP Community License 2.0) and is not part of the software we install.

Real people, logos and the AI Act

A clip that shows a real person needs that person's consent, and a presenter's face and voice are personal data; your data-protection officer confirms what the consent must cover. Logos and products of other companies raise trademark questions that your legal counsel answers.

The European Commission's page on Article 50 of the AI Act, read on 30 October 2026, states that "Article 50 of the AI Act applies as from 2 August 2026". Providers of systems that generate synthetic content must mark it in a machine-readable format, and "Deployers must disclose deepfake content to a natural person upon first exposure at the latest". A deepfake is content that resembles "existing persons, objects, places, entities or events" and would falsely appear authentic; for evidently artistic, creative, satirical or fictional works the disclosure is made "in an appropriate manner". For systems placed on the market before 2 August 2026, the Commission's page gives 2 December 2026 as the date from which the marking obligation applies; the amending act is Regulation (EU) 2026/1744 (Digital Omnibus on AI) on EUR-Lex. Our article on the EU AI Act for companies gives general information; your legal counsel confirms how it applies to your clips.

Notes for the IT team

• System memory. Wan's offloading moves the inactive expert and the text encoder into system RAM, about 69 GB per A14B job by our Wan server article's arithmetic. Our workstation pages follow NVIDIA's rule of at least twice the GPU memory: 192 GB for one 96 GB card, 384 GB for two; our guide to how much RAM a GPU server needs explains it.
• Storage. The Wan2.2-T2V-A14B repository is 126 GB and LTX-2.5's downloads total about 66 GiB, as our video guide records. By our estimate several video models take several hundred GB of NVMe; keep prompt, seed and model revision with every clip.
• Software. LTX-2.5 recommends Python 3.12 or later, CUDA 12.7 or later and PyTorch 2.7; the RTX PRO Blackwell cards run FP8 and NVFP4 in hardware, as our guide to FP8, NVFP4 and other formats explains.
• Network and updates. Keep the ComfyUI port on the internal network, update in planned windows and test the saved workflows afterwards.

FAQ