Video generation studio

Short videos from typed descriptions, on a single 8 GB laptop GPU.

Language
Python
Core
~2 weeks
Models
Two open-weight
Budget
One 8 GB GPU
repo

01The studio

Everything runs from one terminal UI. A typed shot description is translated by a resident 4-bit vision-language model (Qwen3-VL) into concrete generation settings, while readout meters estimate VRAM use, quality, drift, and render time before a render begins. Two open-weight video backends share the 8 GB card, and short shots can be chained into longer clips. Each run archives its settings and telemetry for later comparison or re-analysis. I built the orchestration, UI, and measurement layer around models trained by others.

NEW RUN screen: backend schematic on a fast-to-nicer axis, contextual info panel, and readout meters predicting VRAM, quality, drift, and render time

Setting up a run: the backend schematic (fast → nicer), a contextual INFO panel, and the READOUT meters.

LIVE view mid-render: last-frame preview of a bioluminescent forest, phase timings, and job controls

The LIVE tab mid-render. More screenshots and sample clips in the repo README.

02Output

The measurement layer records seam and drift telemetry and includes a blind A/B harness, so pipeline changes can be compared without relying only on the appearance of a generated clip. Drift control between shots uses cited AdaIN latent anchoring plus a custom color-matching pass.

Stills from two generated clips. Left: bioluminescent forest at night, 35 mm. Right: fireflies over a glowing stream, 64 fps after interpolation. Moving versions in the repo.

03Two points in development

The first clip the project produced (June 22, LTX smoke test) beside a day-fourteen run of the same dog-on-a-beach prompt (July 6, Wan backend), on the same 8 GB card. This shows the project's trajectory, not a controlled comparison: the backend changed between runs.

The studio grew out of an earlier AnimateDiff pipeline adapted to the same unsupported 8 GB card through a PyTorch-SDPA attention backport, a from-scratch LCM scheduler, and a convolution-aware LoRA merge. Modern open-weight models produced the later clips shown here; the earlier implementation work supplied the model-level foundation for the current orchestrator.

Full detail — architecture, the marker protocol, running it — in the README: github.com/harrisonwolf/ltx-studio

← Back to the portfolio