A single rank-32 LoRA that drives a dolly-in camera move at seven selectable
speeds, for LTX-2.5 22B image-to-video. Speed is chosen by a trigger token in
the prompt, so one adapter covers the whole range — unlike the official LTX-2 19b
camera LoRAs, which are single-speed and 19b-only.
Usage
Prefix the prompt with the trigger, then describe the scene normally:
rbdollyin sp20, The camera pushes steadily into the room, the composition tightening.
token
Blender speed
measured zoom over 97 frames
sp05
0.5
1.108x
sp10
1.0
~1.22x
sp15
1.5
—
sp20
2.0
1.591x
sp30
3.0
—
sp40
4.0
—
sp50
5.0
2.285x
Trained for image-to-video with first-frame conditioning at probability 1.0 —
always start from a still.
Base: ltx-2.5-22b-dev-transformer-bf16
Text encoder: gemma4-12b-with-proj-ltx-2.5-bf16
Native resolution: 1024x576, 97 frames @ 24fps
Rank 32 / alpha 32, video attention only (attn1.*, attn2.*)
544 clips, 2000 steps on one H100; step 1250 is the best checkpoint
(3.4% mean absolute zoom error against ground truth across sp05/sp20/sp50)
Keys are already in ComfyUI layout (diffusion_model. prefix) — load with a
standard LoRA loader, no conversion.
Known limitation
Render sharpness degrades over the back third of the clip at sp20 and above.
The training data is sharp throughout and the base model degrades the same way
on identical prompts, so this is generator coherence at long travel, not the
adapter. Under investigation at production inference settings.
