Matt Wolfe tests image and video tools in this 25-minute weekly roundup, with practical differences between a model’s advertised capabilities and its hosted demo.
Image generation and editing
- ChatGPT’s Images 2.6 produces a roller-skating-monkey animation, then accepts edits to its hat and surroundings.
- Wolfe times one still-image request at 20.4 seconds; this is a single demonstration, not a controlled speed comparison.
- Pruna’s PH3 Image offers downloadable weights and a hosted trial, but its skyline example loses to ChatGPT on geographic fidelity.
Open weights still need hardware
- Wolfe shows PH3 memory estimates of roughly 56GB at 16-bit precision and 21GB at 4-bit for an eight-megapixel image.
- Pika H3 Flex’s advertised generation speed uses datacenter hardware; his shared Hugging Face trials take about 30 seconds.
- The hosted Pika demo rejects a longer request because of its GPU-time allowance, illustrating a separate limit from model capability.
- Download size, working memory, license terms, and shared-demo quotas each need their own check.
Video controls and service limits
- Gemini Omni 1.2 Flash is presented as supporting 15-second clips and better reference-image adherence and editing.
- Wolfe generates an avatar clip through Google Vids using a free allowance described as three videos per month.
- Optimizely sponsors a separate demonstration of its SEO and answer-engine-visibility assistant.
Access and adoption deserve the same scrutiny
- Wolfe describes a read-only Plaid connection for Claude and says he has yet to connect his own bank account.
- His rapid-fire items include shared ChatGPT/Codex usage allowances and newly released Kimi K3.1 Pro weights.
- He cites low delivery-order use of agents, but mixes different reporting periods and service denominators.
- Treat the roundup’s launch, pricing, adoption, and benchmark claims as source reports that need checking before a purchasing decision.
“It blocked us because of the GPU duration.” — Matt Wolfe, 11:24