LTX-2 Camera Control LoRAs: Seven Ways to Direct the Move, and What NVFP4 Actually Costs
Camera movement in generated video is usually a suggestion. You write "the camera pushes in", and you get something vaguely going forward. LTX-2's camera-control LoRA family changes that by making the move a conditioning input rather than a prompt hope.
This release ships six workflows across two precisions — bf16 and NVFP4 — which makes it the cleanest comparison of quantisation cost available in this model family.
The two LoRA families
Every graph loads one of two:
| LoRA | What it does |
|---|---|
ltx-2-19b-lora-camera-control-dolly-out.safetensors |
Explicit camera move (dolly out in the shipped config) |
ltx-2-19b-ic-lora-depth-control.safetensors |
Depth-based control, feeding structure rather than a move |
The camera-control LoRA is the one the release title means by "seven ways". The filename names one move (dolly-out); the family covers seven, and swapping the file is the whole change. That is the actual insight: the move is a file, not a prompt.
The depth-control LoRA is a different mechanism — it takes a depth map as structure. Useful when you have a reference and want the model to respect its geometry while moving.
bf16 versus NVFP4: what changes and what does not
The same graphs exist twice:
| Precision | Checkpoint | VAEs |
|---|---|---|
| bf16 | ltx-2-19b-dev.safetensors |
same file for audio VAE |
| NVFP4 | ltx-2-19b-dev-fp4.safetensors |
same file |
NVFP4 is a 4-bit float format. Compared with bf16 that is a quarter of the weight memory, which is the entire reason it exists — a 19B model at bf16 does not fit comfortably on a single consumer card alongside the VAEs and the sampling buffers.
The important structural detail: the graphs are otherwise identical. Same node counts (74 vs 75, 62 vs 62, 67 vs 62), same latent geometry per pair, same sampler, same LoRA. So a difference in output is attributable to the precision, not to a different pipeline.
What 4-bit costs in practice:
- Fine texture. Skin pores, fabric weave, foliage. The first thing to go.
- Small text. Unreadable before bf16 becomes unreadable.
- Gradient smoothness. Banding in skies and soft shadows.
What it does not cost:
- Composition
- Motion
- Identity
- Camera move fidelity
That split is why NVFP4 is usable: the things it degrades are things a moving image hides, and the things it preserves are the ones the viewer tracks.
The latent geometries tell you the job
Two distinct configurations:
768 × 512, 97 frames — the 2-image / 3-image first-last-frame graphs.
960 × 544, 121 frames — the camera-control text-to-video and image-to-video graphs.
Higher base resolution and 121 frames for camera control makes sense: a camera move needs room to happen. 97 frames catches roughly a move and a half; a dolly needs more.
The first-last-frame graphs use the smaller geometry because the constraint is already strong — two frames define the motion, so the sampling budget goes elsewhere.
EmptyImage as resolution declaration
All graphs load EmptyImage at either 1280×720 or 1920×1080 with colour 0 (black).
That is not a placeholder to fill — it is how the graph declares its working resolution to the latent encoder. The 2/3-image graphs declare 1280×720 while generating at 768×512; the camera-control graphs declare 1920×1080 while generating at 960×544. In both cases the declaration is double the generated size.
If you change one without the other you get an implicit rescale, and the camera move scales with it — which makes the move look faster or slower than intended.
euler rather than ancestral
Every graph uses plain euler.
For camera control that is the right call: an ancestral sampler injects noise per step, and noise perturbs trajectory. A camera move is a trajectory. Deterministic sampling keeps it on the intended path.
If your dolly looks like it wobbles, the sampler is the first thing to check.
Where this release is weak
- One LoRA file per graph. The seven moves exist as seven files; the workflows ship one each. You have to source the others.
- No comparison output. The bf16 and NVFP4 pairs exist but the release does not show them side by side; the comparison is left to the viewer.
- Depth control graph is 67 nodes and the release does not explain where the depth map comes from.
-
dolly-outin the filename while the title claims seven — easy to assume you have full coverage when you have one move.
Practical order
Run the bf16 camera-control graph first if memory allows — it is the reference. Then the NVFP4 of the same graph on the same prompt. The difference you see is the quantisation cost, and that is a number worth knowing for your own hardware.
For first-last-frame work, the geometry is more constrained and precision matters less.
Workflows
Six graphs, published on RunningHub with invite code rh-v1111:
- Non-distilled high-quality LTX-2 camera-control text-to-video
- Non-distilled high-quality LTX-2 camera-control image-to-video
- LTX-2 two-image / three-image first-last frame
- NVFP4 non-distilled LTX-2 camera-control text-to-video
- NVFP4 non-distilled LTX-2 camera-control image-to-video
- LTX-2 two-image / three-image NVFP4 first-last frame
Files
https://huggingface.co/AI-KSK/ksk-video-155-comfyui-workflows
Video walkthrough (Chinese): https://youtu.be/DJRhC6IdZVM
Bilibili: https://www.bilibili.com/video/BV1vfrNBkEHc/
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- Node counts, geometries and loader filenames quoted above are read from the published JSON files.
- Independent AI-KSK release. Not affiliated with Lightricks or RunningHub.
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)