# Kollektive First Look: LTX-Video 2B four-case batch

Gallery: https://kollektive-video-first-look.pages.dev/

Completed: 4; failed: 0; skipped: 0.
Rental duration: unknown until provider cleanup is confirmed.
Intent recorded: 2026-10-18T20:15:04.084000+00:00; cleanup confirmed: unknown.
Provider: Vast; instance: 56612364; region: Belgium, BE; GPU: A40.
Estimated total rental compute: unknown.
Provider reconciled charge: unknown until the provider billing record is checked.
Shared preparation, idle, upload and cleanup: unknown until cleanup.
Cost per usable output second: unknown until human usable intervals are reviewed.

## Experiments

| ID | Outcome | Actual video | First completed step | Loading | Inference | Encoding | R2 upload | Case total | Sampled VRAM max | Estimated marginal compute | Error |
|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
 | t1 | completed | 4.05s | 0.72s | 42.5s | 3.9s | 0.5s | 4.1s | 50.8s | 24355 MiB | $0.00055 |  | 
 | i1 | completed | 4.05s | 0.32s | unknown | 3.6s | 0.7s | 4.3s | 5.5s | 24357 MiB | $0.00065 |  | 
 | t2 | completed | 12.05s | 0.42s | unknown | 10.0s | 1.1s | 4.0s | 11.6s | 41223 MiB | $0.00136 |  | 
 | i2 | completed | 12.05s | 0.47s | unknown | 10.5s | 1.6s | 4.3s | 13.5s | 41223 MiB | $0.00158 |  | 

## Technical limitations

- Child phase and allocator timings are self reported; case totals and GPU samples are host observations.
- Sampled device memory is not a precise VRAM peak.
- Creative quality and usable seconds require human review.
- This was research infrastructure; outputs are not published Kollektive Works.

## Rental and preparation observations

- GPU instance absent from Vast inventory by 2026-10-18 20:36:53 UTC, at most 1308.9 seconds after rental intent. At the recorded $0.422222/hour quote, the time-based estimate is at most $0.1535; provider-reconciled charges remain unknown.
- The four verified files total 32.20 generated seconds and 644 frames. The full rental quote estimate is at most $0.00477 per generated second or $0.000238 per frame. Usable-second cost awaits review.
- SSH/container readiness took 85.7 seconds. The 3,996,143,615-byte runtime transfer and hash verification took 157.0 seconds (25.5 MB/s). Runtime extraction took 94.6 seconds; model loading within T1 took 42.5 seconds.
- The first parallel model pass failed the SHA-256 check for one PixArt shard after 202.2 seconds; the exact shard passed on retry. The failed first batch was caused by a mode-700 ancestor directory blocking the restricted inference process. It produced no media; its record is retained.
- The temporary SSH key remains in Vast account inventory (ID 1603604). The API key lacks the `api.ssh.key` deletion permission (HTTP 401). The GPU instance is destroyed. The temporary local private key and signed R2 GET grants were removed; the orphaned public key still needs account-level deletion.
- Raw one-second host samples, the full four-case manifest, initial setup failures, upload readback receipt, staging report and rental journal are preserved as separate research evidence.

## Technical findings

- The corrected pinned runtime generated and decoded every requested 81-frame and 241-frame clip at 768 × 448, 20 fps. The 12-second cases were technically practical on this A40 at the tested settings; sampled VRAM reached 41,223 MiB of 46,068 MiB, leaving about 4,845 MiB measured headroom.
- Warm inference took 3.64–3.94 seconds for 4.05-second clips and 9.99–10.49 seconds for 12.05-second clips. These figures exclude shared downloads and the first 42.53-second model load. They measure this checkpoint and these prompts, not expected performance across other models or settings.
- The videos are published without a quality verdict. Whether the joke is understandable, the source image survives motion, or the checkpoint belongs in Kollektive requires the owner's visual review. A technically complete run alone does not justify model admission.