Same scene · Same seed

Singularity vs Standard H3

Two models, one red sailboat. Compare detail, motion, reflections, and how each follows the prompt.

1344 × 768 each24 fps · 5.17 secondsSeed 42Both: Turbo 4 steps

Top and bottom · generation stats

Singularity above · Standard H3 below. Generation times and settings are included in the video. Silent, synchronized playback. Download vertical comparison

Watch together

Singularity on the left · Standard H3 on the right. This synchronized player is silent. Use fullscreen for a closer look. Download comparison

The shared prompt

A cinematic close-up of a small red toy sailboat floating on a calm garden pond. A gentle breeze moves the boat slowly from left to right and creates soft ripples. Sunlight sparkles on the water. The camera remains steady. Natural water sounds and distant birdsong. No dialogue, no music, no text.

Generation settings and models

Only the diffusion checkpoint changes. Both use pruned INT8 weights, the same text encoder and VAEs, seed 42, Euler/simple sampling, the Ref2VA four-step Turbo adapter, video/audio shifts 12/3, and Comfy Kitchen HIP INT8 attention.

This is a comparison with Turbo enabled, not a test of maximum non-Turbo quality. The adapter was trained at 544p; both clips use the same 768p output settings. One scene and seed cannot establish general model quality.

Rendered locally on AMD Radeon 8060S: Singularity 13m38s; standard H3 12m23s. Standard H3 reused cached conditioning, so these timings are not a controlled speed benchmark.

Singularity weights · Standard H3 weights · Turbo adapter settings