The Intel B70 GPU with 32 gigabytes of video RAM remains one of the least expensive cards on the market (affiliate link)for its capacity, and it has served as the centerpiece for my recent explorations into local AI. My current setup utilizes this card alongside an Oculink dock (affiliate link) and a Mini PC running Ubuntu 26.04.
My latest video takes a look at running Comfy UI on the Intel B70 to generate video and images – the results are staggering for local hardware!
In one test using the Minimax H3 model, a five-second video clip took approximately two minutes and 36 seconds to generate. This particular output was rendered at 480p resolution. While the hardware can handle up to 1080p, performance decreases as the resolution increases.
In my practical applications, the model has shown varying degrees of success with complex prompts. I experimented with making a spoof of a McDonald’s commercial and several Star Trek-themed clips. The system effectively handles motion and visual consistency, but it faces challenges with dialogue. In clips involving two characters, the model often struggles to keep voices distinct, sometimes mixing them together. It also requires precise prompting to get exactly what you’re looking for. But that’s no different than the frontier video generators.
Beyond narrative clips, the hardware handles utility-focused tasks like creating looping backgrounds for motion graphics. It also supports image-to-video generation, where a single still image is animated, or two images are used as “before and after” anchors for the model to bridge with video. The node-based interface of Comfy UI provides a level of control that allows for complex orders of operation, which can help refine these outputs.
Static image generation is notably faster. Using the Flux 2 Klein 9B model, I generated a 1024×1024 image in 35 seconds. I also tested a more complex request: a three-panel storyboard featuring a consistent character across different frames. The model maintained the character’s appearance and clothing throughout the sequence, which is a significant requirement for practical design work. This specific model requires an API key for access (likely for watermarking output), though it runs locally.
Configuring these tools on Linux requires some technical effort. I have relied on frontier models like ChatGPT’s Codex command line to assist with the configurations necessary to bridge the gap between the software and the Intel hardware.
One of the benefits for doing image and video generation locally is that it provides a sandbox for iterating on prompts without consuming paid cloud tokens. A user can refine their ideas on local hardware and then take those perfected prompts to a high-end cloud model for a final, high-resolution render.
The current state of local AI feels increasingly more accessible compared to just a year ago. The quality of these open-weight models is approaching what industry-leading frontier models were producing in the recent past. As these models become more efficient, the entry barrier will likely continue to drop. My next objective is to borrow an AMD-based GPU rig to observe how it handles these same workloads and whether it offers any specific performance advantages over the Intel setup as the AMD AI Pro 9700 is now closer in price (affiliate link) to the Intel board.