An IC-LoRA (In-Context LoRA) for LTX-Video 2.3 (22B) that conditions video generation on a pose video (DWPose 133-keypoint skeleton render) so the generated subject's pose follows a user-supplied reference. Trained on the same 78-episode Chinese historical drama corpus as the sibling character LoRA, so it preserves the live-action Han dynasty visual style while you control the motion.

Model details

Field	Value
Base model	`Lightricks/LTX-2.3-22B`
Adapter type	IC-LoRA (with reference-video conditioning)
Conditioning input	Pose video (DWPose 133 keypoints rendered to MP4)
Rank	128
Alpha	128
Target modules	`to_k`, `to_q`, `to_v`, `to_out.0`
Training steps	6000
Optimizer	AdamW
Learning rate	1e-4, linear schedule
Mixed precision	bf16
Reference channel	concatenated to the latent via `--video-conditioning` (strength 1.0)

Training data

Identical corpus + caption format to the character LoRA, plus per-clip pose videos extracted via DWPose 133-keypoint model (whole-body keypoints, COCO-WholeBody schema). The pose video is rendered at the same resolution as the source clip and serves as the structural reference during training.

Usage

IC-LoRA inference requires a pose reference video at the target resolution + an inline-weave prompt. The reference can be extracted from any source video with the matching DWPose pipeline.

LTX `ltx_pipelines.ic_lora` invocation

bash

python -m ltx_pipelines.ic_lora \
    --prompt "char_0_person, char_1_person. Framed in a static eye level medium shot, on a 35mm normal lens. Set in a Han dynasty courtyard, the subjects face each other. Live-action photorealistic, cinematic Chinese drama." \
    --negative-prompt "no CGI, no animation, no illustration, no painterly style, no anime" \
    --lora <path_to>/lora_weights_step_06000.safetensors 1.0 \
    --video-conditioning <pose_reference>.mp4 1.0 \
    --width 1280 --height 544 --num-frames 89 \
    --guidance-scale 4.0 --num-inference-steps 20 \
    --skip-stage-2

Recommended strengths

Component	Value
LoRA strength	1.0 (validated default)
Video conditioning strength	1.0 (faithful pose following)
Lower video conditioning (0.5–0.7)	softer pose adherence, more creative interpretation

When to use this vs the character LoRA

Use case	Reach for
"Generate a Chinese drama scene with these characters"	Character LoRA
"Generate a Chinese drama scene where the subject follows THIS pose video"	Pose IC-LoRA (this one)
"Generate a scene that exactly mirrors the camera motion of THIS reference"	Depth IC-LoRA (different repo)
"Generate a scene whose composition matches THIS line-art reference"	Canny IC-LoRA (different repo)

The IC-LoRAs are stackable with the character LoRA — load both, set adapter weights, and you get character identity + pose control simultaneously. Validated stack: char 0.9 + pose 1.0.

What this LoRA does well

Faithful skeletal pose transfer from reference to generated subject.
Preserves the Chinese drama style anchor (palace lighting, period costume) regardless of which pose you feed in.
Multi-character pose tracking works when 2-person reference videos are used.

What it does NOT do

No identity — IC-LoRA controls motion, not who the subject is. Use it stacked with the character LoRA or with character trigger tokens in the prompt to fix identity.
No camera motion control — it follows pose, not camera trajectory. For camera control see the v3 adapter stack (separate project).
Pose reference must be at the same resolution as the target output.

Related models

SyFeee/ltx2.3-chinese-drama-charlora — character / style LoRA, stackable with this IC-LoRA.
SyFeee/ltx2.3-chinese-drama-iclora-depth — depth-controlled IC-LoRA.
SyFeee/ltx2.3-chinese-drama-iclora-canny — canny-edge-controlled IC-LoRA.

License

Apache 2.0. See LICENSE for terms.

Attribution: SyFe.

SyFeee/ltx2.3-chinese-drama-iclora-pose

LTX-2.3 Chinese Drama IC-LoRA — Pose Control