fix(ltx): dev transformer + distill LoRA so camera LoRAs bind

[mcarlson]

Why

The LTX "Camera" presets produced no camera motion in any direction. Root cause: the per-direction camera-control LoRAs are trained for LTX-2 19B, but the deployed transformer was ltx-2.3-22b-distilled. ComfyUI can't bind the mismatched keys — it logs lora key not loaded and silently emits base-model output while the job reports success. Full diagnosis: docs/plans/2026-06-19-ltx-2.3-camera-lora-not-applied.md (metarepo).

The fix (3 coordinated PRs)

Switch the LTX i2v path to the full "dev" 22B transformer (which the 19B camera LoRAs do bind to), and add the distill LoRA as a model-only loader ahead of the camera LoRA so few-step LCM sampling is preserved.

Repo	PR	Change
gpu-container-ltx	this/linked	Bake the dev transformer + ltx-2-19b-distilled-lora-384; test_input.json updated to the dev chain
worker	this/linked	createLtxI2VWorkflow: node 1 = dev UNET, node 7 = LoraLoaderModelOnly distill @0.6, node 6 Power Lora Loader (camera LoRA) chains off node 7; both CFGGuider passes consume node 6
frontend	this/linked	Enable all 7 LTX Camera preset directions (static + dolly in/out/left/right + jib up/down)

Model chain: UNETLoader(dev) → LoraLoaderModelOnly(distill @0.6) → Power Lora Loader(camera @0.4) → both CFGGuiders.

Deploy order (important)

The worker references the dev model + distill LoRA by filename, so the image must ship them first:

1. Merge gpu-container-ltx → GHCR build (Actions) → update the RunPod LTX endpoint to the new image.
2. Deploy worker to Heroku (stage).
3. Merge frontend (just enables the UI presets — harmless on its own, but moves only work once 1+2 are live).

Verification runbook (this is "done")

Stage A — container-direct (isolates the model/LoRA chain from the app):

1. After the GHCR build, point the RunPod LTX endpoint at the new image and cold-start a worker.
2. Submit test_input.json (runs static @0.4). Confirm it completes and the output is coherent — this validates the dev model + distill LoRA produce good few-step output.
3. Swap node 6 lora_01.lora to ltx-2-19b-lora-camera-control-dolly-in.safetensors; resubmit; download the video.
4. Check the ComfyUI/RunPod log for node 6: confirm there is no lora key not loaded / N keys not loaded warning — this is the definitive bind check. Also confirm no VRAM OOM (dev fp8 ≈ distilled footprint, but verify on the endpoint's GPU).
5. Visually: static holds still; dolly-in shows a forward push. If motion is weak, raise camera strength (0.4 → 0.6/0.8). If quality degrades / looks under-denoised, adjust distill strength (0.6).

Stage B — full Studio E2E:
6. With the worker deployed and frontend merged to stage, open Studio → "LTX Camera" preset → run a batch across one image.
7. Confirm each column shows the correct motion: static still; dolly in/out push/pull; dolly left/right slide; jib up/down rise/descend.

Tuning & risks

• Starting strengths: distill 0.6, camera 0.4. Both are starting points — tune per Stage A/B results.
• dolly-left / dolly-right are the least-validated directions (community reports were mixed on left/right even on the dev model). Verify them specifically in Stage B; if they show no motion, bump their strength or drop those two preset actions.
• Stale reference workflow: gpu-container-ltx/workflows/jef-ltx-2.3-i2v.json still points at the distilled model. The worker builds its own workflow so production is unaffected, but local docker-compose testing would use the old model. Update it if you test locally.
• Image size / cold start: baking the dev model + distill LoRA + 7 camera LoRAs grows the image → slower cold starts.

🤖 Generated with Claude Code