ltx2-vidgen-skill

A Claude Code skill that owns your AI video pipeline — drop a photo in Claude Code → get a video. It deploys your own optimized LTX-2.3 (22B) backend to your Modal serverless GPU and drives it: text-to-video, image-to-video, keyframe interpolation, video-to-video, and IC-LoRA (canny/depth/pose) control, with synced audio — a few cents a clip, your GPU, no per-clip API meter, no rate limit.

Side-by-side image-to-video — two stills brought to life (slow push-in + wind-blown hair). LTX-2.3 (22B) on RTX PRO 6000 (Blackwell, 96 GB, bf16). ▶ full-quality mp4

What this is

A Claude Code skill that owns the whole loop — it deploys your own LTX-2.3 (22B) video backend to your Modal account on first run, then drives it from Claude Code: drop a photo (or a prompt), ask for a video, get the .mp4. You never leave Claude Code; the GPU is yours; there's no SaaS in the middle.

• Modes: text-to-video, image-to-video, keyframe interpolation, video-to-video (retake/restyle), and IC-LoRA canny/depth/pose control — with synced audio.
• Batch: --variations N (one prompt, N takes) and --prompts-file (many prompts) in one warm container.
• Formats: reel / TikTok / Shorts (9:16), YouTube (16:9), square — native, no cropping.
• Optimized: resident pipeline + caches + torch.compile → ~1.96× faster, output bit-identical, bf16 throughout. Clips up to ~10 s; any resolution with sides ÷32.

Install

1. Install the skill (Claude Code, or any agent that reads SKILL.md):

npx skills add patraxo/ltx2-vidgen-skill        # or: cp -R skills/ltx2-video ~/.claude/skills/
pip install modal && modal token new            # Modal SDK + your account (free — $30/mo credits)

…or as a Claude Code plugin marketplace:

/plugin marketplace add patraxo/ltx2-vidgen-skill
/plugin install ltx2-vidgen@ltx2-vidgen-skill

2. Deploy your backend once — into your Modal account (downloads the LTX-2.3 weights + Gemma text encoder; public components only, no HuggingFace token):

git clone https://github.com/patraxo/ltx2-vidgen-skill && cd ltx2-vidgen-skill && ./deploy.sh

That's it. Now in Claude Code, drop a photo (or just describe a scene) and ask:

• "turn this photo into a video, subtle natural motion" — image → i2v
• "interpolate between these two frames" — two images → keyframe
• "restyle the middle of this clip" — a video → v2v
• "generate a neon city street, vertical reel" — prompt only → t2v
• "follow the edges of this clip" — a control render → IC-LoRA control

The skill validates the input, confirms cost (offers a cheap smoke first), calls your deployed app via modal.Cls.from_name (no endpoint, no auth, no secrets), and saves the .mp4 + a preview frame to ./video_out/. First clip cold-starts ~90 s; warm ~31 s.

PYTHONPATH=. uv run modal run deploy/ltx2_model.py::smoke_real --image-path pic.jpg   # i2v
PYTHONPATH=. uv run modal run deploy/ltx2_model.py::run_modes                          # t2v + i2v + keyframe
PYTHONPATH=. uv run modal run deploy/ltx2_model.py::run_retake                         # v2v
PYTHONPATH=. uv run modal run deploy/ltx2_model.py::kf_real --image-a a.jpg --image-b b.jpg
PYTHONPATH=. uv run modal run tests/ship_verify.py                                     # full verify

Modes

All four are exercised by run_modes / run_retake / kf_real and verified working (frame-inspected). The opt stack applies across every mode.

Performance

Runs on Modal serverless NVIDIA RTX PRO 6000 (Blackwell, 96 GB), bf16, billed per-second at $0.000842/s (~$3.03/hr) — so cost ≈ latency:

(5 s/10 s rows re-measured 2026-06-10, bench config with the first-block cache off — the production path with it on is faster still. First call at a NEW frame count pays a one-time shape compile, e.g. ~86 s at 241 f, then steady.)

Cold start (first clip on a fresh container) ~90–200 s; idle scales to $0. Both 22B stage transformers stay resident — peak ~75 GB / 96 GB, verified zero OOM over a 10-generation run (<1 GB drift).

Optimization stack — bf16 throughout, cache paths bit-identical to the unoptimized run, ~1.96× net faster:

• Resident pre-fused pipeline + activation-aware resident cap + cross-resolution purge → no per-clip rebuild and no OOM when switching mode/resolution.
• First-block feature cache (~17%), CPU-pinned weight cache, text-embedding cache, torch.compile (persisted Inductor cache), tunable VAE-decode tiling.
• fp8 / SageAttention / flash-attn all tested and rejected — fp8: quality rule; flash-attn: no sm_120 kernel; SageAttention 2.2: builds + runs on sm_120 but measured +5% slower here (compile graph breaks — see Performance research below). Speed comes from architecture, not from cheapening the math.

The pricing problem, solved

Per-clip APIs charge you for every take — including the failed ones. Iteration is exactly how you get good AI video, so the meter punishes the workflow that works. Self-hosting flips the economics:

Third-party prices are approximate (mid-2026) and change often — verify before quoting. The point isn't the decimals; it's the shape: a failed take here costs cents, twenty variations cost a coffee, and the quality knob is yours (full bf16, no silent quantization).

New Modal accounts get $30/month in free credits. At ~1.9¢ per 5-second clip that's ~1,500 free clips/month — explore a prompt 20 ways (--variations 20) and keep the one that lands. It's your Modal account and bill (visible in your dashboard), no per-clip meter, no rate limit, idle = $0.

How it works

LTX-2.3 is a 22B video diffusion transformer. Serving it naively has two costs: a slow cold start, and a per-clip cost where the pipeline re-assembles + re-fuses model internals every request. This repo attacks both:

1. Build once, stay resident. The fully-assembled, LoRA-fused transformer stays in GPU memory between clips (resolution-keyed, LRU-bounded) instead of rebuilt per request — the single biggest win when it fits (short/low-res: ~90 s → ~7 s).
2. Both stages stay resident — safely. Two ~35 GB stage transformers + the (tiled) forward working set peak ~75 GB / 96 GB, so both stay GPU-resident at every supported resolution (measured; zero OOM over a 10-gen run, <1 GB drift). An activation-aware cap + a cross-resolution purge (drop residents from a prior, different-resolution request before the next forward) keep it safe — back-to-back mode/resolution switching never OOMs. (The earlier 94 GB OOMs were that cross-resolution accumulation, not the resident footprint.)
3. CPU-pinned weight cache — weights pinned in host RAM, streamed to GPU, skipping disk reads. Bit-identical.
4. First-Block-Cache — skips recomputing early transformer blocks (~17%, near-lossless).
5. Embedding cache + streaming text-encoder guard — the text encoder isn't re-run for a repeated prompt. On a cache miss with both stages resident there isn't room for the full ~23 GB Gemma build (measured 93.7 GB → OOM), so the miss automatically switches to the upstream layer-streaming build (~5 GB peak, identical embeddings, ~+10 s once per prompt per container).
6. Batching + decode/encode overlap — 32 clips in one warm container ≈ 6.2× throughput; finalize (VAE decode + mp4 mux) of clip N overlaps the denoise of clip N+1 on a worker thread → −20% wall measured on a 3-clip batch (92.5 s → 73.7 s).
7. torch.compile + persisted Inductor cache — compiled once on the first cold container, restored from the volume on later cold starts (max-autotune tested + rejected, slower here).
8. Blackwell-native attention — flash-attn has no sm_120 kernel yet, so this runs PyTorch SDPA (exact). fp8 rejected (quality rule); SageAttention 2.2 builds + runs cleanly on sm_120 but measured +5% slower here (~770 torch.compile graph breaks per request — the kernel can't be traced). Speed comes from architecture, not from cheapening the math.

Helper modules live in utils/ and are mounted into the container.

Optimization flags (env vars, set in the image .env block)

Performance research (June 2026)

Three write-ups of the findings, written to be useful beyond this repo:

• Why SageAttention made our 22B video model slower — compile graph breaks vs kernel wins, and the trace-time SDPA gotcha
• NVENC fails in Modal memory-snapshot containers — the four-experiment evidence chain
• The text-encoder OOM your benchmarks never catch — why same-prompt load tests hide a multi-GB allocation

A systematic study of lossless latency levers for this stack — every lever measured A/B on the same warm container, gated on zero visual loss (blackdetect + PSNR/SSIM + frame inspection + audio). Full write-ups in references/:

Floor: a warm 5 s clip is ~19.5 s GPU-bound bf16 DiT + ~2.7 s CPU x264 encode + ~0.5 s overhead. The DiT term only moves with quantization or step cuts — both off the table by the quality rule. Full record: references/LATENCY_RESEARCH_2026_06.md, distilled gotchas in references/learnings.md.

Two transferable gotchas worth stealing:

1. Inductor captures the SDPA backend at trace time — benchmarking attention backends with runtime context managers on a compiled model measures nothing.
2. Same-prompt benchmarks never exercise the text-encoder memory path — always include a warm new-prompt arm, or you'll ship a latent OOM like the one found here.

Layout

pyproject.toml             # uv project (client-side dep: Modal SDK)
deploy.sh                  # one-command deploy
deploy/ltx2_model.py       # the Modal app: t2v / i2v / keyframe / v2v + opt stack
deploy/bench_*.py          # same-container A/B bench drivers (sage, cudnn, matrix, fixes)
deploy/probe_nvenc.py      # bare-container NVENC probe (no memory snapshot)
deploy/utils/              # helper package (weight registries, FBCache, guiders)
references/                # performance research notes (latency study, sage sm_120, learnings)
skills/ltx2-video/         # Claude Code skill (SKILL.md + scripts/submit_video.py)
tests/                     # smoke_test.py + ship_verify.py
assets/demo.gif

Acknowledgements & references

• LTX-Video (Lightricks) — the 22B model + upstream ltx-core/ltx-pipelines. Paper: HaCohen et al., LTX-Video: Realtime Video Latent Diffusion — arXiv:2501.00103.
• Modal — serverless GPU (RTX PRO 6000, per-second billing).
• Gemma-3 — text encoder.

Optimization techniques:

• Block-feature caching (the first-block cache here) — cf. BWCache: Accelerating Video Diffusion Transformers through Block-Wise Caching — arXiv:2509.13789; broader caching lineage e.g. SenCache (LTX-Video) — arXiv:2602.24208.
• VAE-decode spatial/temporal tiling, CPU-pinned weight cache, persisted torch.compile (Inductor) cache, bf16-exact throughout (no fp8 / quantization).

By @patraxo. LTX-2.3 weights are distributed under Lightricks' license.