What we tested building TallyReel
Real runs, real numbers, real outputs
TallyReel runs open AI models for voices, pictures, talking hosts and stories. Before we ship one, we test it against the others on our own machines and on cloud GPUs. These are the write-ups: what we ran, what it cost, what came out, and what we chose. Follow along with the RSS feed.
What it really costs to render a faceless video on a cloud GPUWe rendered the same three faceless videos on a cloud L4 GPU and timed every step. About $0.04 of GPU per Short, about $0.08 all in, and where it adds up.
Running FLUX image models locally on a Mac for video picturesWe swapped SDXL for FLUX.2 klein 4B to draw story pictures on an M1 Pro Mac. Speeds, memory, side-by-side pictures, and the bugs we only found by looking.
LTX vs Wan for image-to-video: why we parked moving scenesWe tested Wan 2.2, FastWan and LTX image-to-video for story scenes, on a cloud GPU and a Mac. Times, costs, the clips, and why stills stayed the default.
We tested open-source lip-sync: MuseTalk vs EchoMimic for talking AI hostsWe ran MuseTalk 1.5 and EchoMimic V3 Flash on the same line and the same face on an M1 Pro Mac. Render times, the clips, and why we kept MuseTalk.How we write a story in one AI call (and why five passes lost)Our story writer used five to seven AI calls per story. One call scored higher. What we measured, what still went wrong, and the fix that won 7 of 8 blind.Pick one. Make it tonight.
Type the idea into TallyReel, read the script, and it makes the video for you: in the browser, on your phone, or in Studio on your own computer. Start on the free plan.