Skip to content

Releases: patraxo/ltx2-vidgen-skill

v1.1.0 — performance research, batch overlap, OOM fix

11 Jun 07:03
@patraxo patraxo
v1.1.0
7e8e202

Choose a tag to compare

View all tags
v1.1.0 — performance research, batch overlap, OOM fix Latest
Latest

What's new

  • Batch decode/encode overlap: 3-clip batch wall down 20.4% (92.5s → 73.7s). Finalize of clip N overlaps denoise of clip N+1.
  • Warm-container OOM fix: first new-prompt request no longer dies at 93.7GB — text encoder streams layer-wise when VRAM is tight (LTX_EMB_STREAM_FREE_GB, identical embeddings).
  • Reference-exact VAE decode: LTX_VAE_TILE_PX=0 disables tiling entirely (50-51dB vs tiled, latency-neutral).
  • Performance research published: three findings write-ups in docs/ (SageAttention +5% slower under torch.compile, NVENC vs memory-snapshot containers, the text-encoder OOM benchmarks miss) + full measured record in references/.
  • Re-measured performance table: 5s clip 23s warm ($0.02), 10s 45s ($0.04).
  • Filed upstream: Lightricks/LTX-2#232 (per-component offload).

Full details: README Performance research section.

Assets 2
Loading

v1.0.0 — ltx2-vidgen-skill

08 Jun 11:14
@patraxo patraxo
v1.0.0
c2f7600

Choose a tag to compare

View all tags
v1.0.0 — ltx2-vidgen-skill

A Claude Code skill that deploys your own LTX-2.3 (22B) video backend to your Modal account and drives it — drop a photo in Claude Code, get a video. You own the GPU; no SaaS, no per-clip meter.

Modes

text-to-video · image-to-video · keyframe interpolation · video-to-video (retake/restyle) · IC-LoRA canny/depth/pose control — with synced audio.

Highlights

  • Batch: --variations N (one prompt, N takes) + --prompts-file (many prompts), one warm container.
  • Formats: reel/TikTok/Shorts (9:16), YouTube (16:9), square — native, no cropping.
  • Optimized: resident pre-fused pipeline + activation-aware cap + cross-resolution purge (no OOM on mode/resolution switching), first-block cache, CPU-pinned weights, embedding cache, persisted torch.compile, tunable VAE-decode tiling → ~1.96× faster, bf16 throughout, output bit-identical. fp8/SageAttention/flash-attn tested + rejected.
  • Cost: Modal RTX PRO 6000 @ $0.000842/s — ~2.6¢ per 4 s clip; new accounts get $30/mo free (~1,000 clips). Idle = $0.

Install

npx skills add patraxo/ltx2-vidgen-skill
pip install modal && modal token new
git clone https://github.com/patraxo/ltx2-vidgen-skill && cd ltx2-vidgen-skill && ./deploy.sh

Powered by Lightricks LTX-Video on Modal.

Loading