Use case: "Wanderer", a 17-shot short film made locally with Pika H3 Flex, YuE2 and Qwen Image 2.1 on one RTX 4090
ReAiLity Labs (@CreaitiveLabs) made "Wanderer", about 96 seconds of cinematic footage in 17 shots, entirely inside ComfyUI on a single RTX 4090, with 50 minutes of video render time. Their write-up shares the benchmark of Pika H3 Flex's 6 steps against 30-step MiniMax H3, the settings for each stage and every prompt.
What it does
Four stages, all ComfyUI templates: the score came from YuE2 Music Cover (1:46.7 long, measured at 103.4 BPM with a big hit every 2 bars, about 4.6 seconds, which became the cut rhythm); a storyboard split the track into six sections and planned 22 shots; first frames came from Qwen Image 2.1 at 1920×1088, 30 steps, CFG 4, with a character sheet as reference for shots showing the woman's face or hands; and every clip came from the unchanged Pika H3 Flex FL2VA template at 1344×768, 24 fps, with H3's structured prompt format. They muted H3's own audio and cut to the score's drum hits.
Results and limits
In their benchmark (same still, prompt and seed, 9.4 s at 1344×768) a 30-step H3 render took 27.4 minutes and the Pika 6-step template 5.7 minutes, and the two looked almost identical; both runs loaded the Pika adapter, so the test compares the schedules. Stills took about 32 seconds each (42 with the reference), and 5-second clips about 2.1 minutes, so the 17 final clips took about 50 minutes against an estimated 3½ hours at 30 steps. In very wide shots a detailed costume description made the model draw the woman twice, and a simple description fixed it.
Worth knowing
It fits on a 24 GB card because it runs the int8 H3 FL2VA model with the int8 video VAE. The creator used Claude to analyse the track, write the storyboard and prompts and queue the stages, and made the creative choices and final cut. The idea credits a MiniMax H3 4K video posted on r/StableDiffusion. MiniMax H3's community licence excludes the EU, UK, South Korea and US unless you get a licence from MiniMax.
Try it yourself
Character reference sheet on a plain mid-grey studio backdrop, three full-length views side by side of the same slender woman in her late twenties with dark hair pulled into a tight low bun, wearing a long fitted pale silver-blue gown and an enormous floor-length cape of thin translucent silk: on the left a full back view, in the center a three-quarter front view with her face visible, calm and serious, high cheekbones, on the right a side profile. Soft even studio light, the silk cape draping to the floor and pooling around her feet, identical costume and hairstyle in all three views, live-action photograph, soft 35mm film grain, photorealistic, 16:9