A field-tested image-to-video character-consistency LoRA for Lightricks/LTX-2.3 (22B distilled), tuned for two-character dialogue scenes and multi-shot cinematic video generation.

⚠️ Naming note (corrected 2026-05-21): The original filename and ModelScope repo include the string "IC-LORA", but this is NOT an IC-LoRA in the strict technical sense (parallel-canvas / video_conditioning mechanism). An A/B/C test (same prompt + seed, three reference-channel variants) confirmed that the LoRA's actual conditioning mechanism is first-frame pixel pinning (the regular i2v path), not parallel-canvas attention. Earlier copy on this card incorrectly described it as IC-LoRA — that has been removed. Credit to ZKong for raising the discrepancy in the discussions tab.

Example renders

Episode is an 8-shot Chinese palace drama (《玉佩定情》 + 《暗夜阴谋》) with three characters: 沈月华 (Shen Yuehua, heroine), 萧云霄 (Xiao Yunxiao, prince), 慕容静 (Murong Jing, antagonist). Render config: 1280×704, 121 frames @ 24 fps, ambient audio.

Single-character identity — Shen Yuehua walking in the garden, picks up a jade pendant

Dual-character dialogue — Shen + Xiao meet (the LoRA's signature use case)

Cross-scene identity — Murong Jing in a different location (palace night chamber)

Three-character composition — the LoRA's upper limit

What this LoRA does

Fine-tuned on Lightricks/LTX-2.3 (22B distilled), specifically for:

Two-character dialogue scenes — significantly reduces character drift when two people appear in the same frame
Cinematic shot composition — reinforced for dialogue-driven framing (close-up ↔ medium ↔ wide)
Multi-shot narrative continuity — better understanding of multi-segment prompts (storyboard-style descriptions)
Style compatibility — works well across 古风仙侠 (ancient Chinese fantasy), 现代都市 (modern urban), and 3D 动漫 styles

The reference image is consumed via first-frame pixel pin (standard i2v conditioning), not via the parallel-canvas / video_conditioning channel.

How to use (correct pattern)

Single-character shot

python

# Upstream LTX-2.3 distilled pipeline — single reference as first-frame pin
from ltx_pipelines.distilled import DistilledPipeline
from ltx_pipelines.utils.args import ImageConditioningInput
from ltx_core.loader import LoraPathStrengthAndSDOps, sd_ops as _sd_ops_mod
import torch

lora = LoraPathStrengthAndSDOps(
    "LTX2.3-IC-LORA-Dual-Character.safetensors",
    0.8,                                       # strength (standalone)
    _sd_ops_mod.LTXV_LORA_COMFY_RENAMING_MAP,
)

pipe = DistilledPipeline(
    distilled_checkpoint_path="ltx-2.3-22b-distilled-1.1.safetensors",
    spatial_upsampler_path="ltx-2.3-spatial-upscaler-x2-1.1.safetensors",
    gemma_root="google/gemma-3-12b-it-qat-q4_0-unquantized",
    loras=[lora],
    device=torch.device("cuda:0"),
)

video, audio = pipe(
    prompt="...",
    seed=42,
    height=704, width=1280,
    num_frames=121,                              # 5 s @ 24 fps, satisfies 8k+1
    frame_rate=24,
    images=[ImageConditioningInput(             # first-frame pin = THE reference mechanism
        path="character_ref.png",
        frame_idx=0,
        strength=0.9,
    )],
    enhance_prompt=False,
)

Dual-character shot

LTX's i2v pin rejects two pins at the same frame_idx, so two refs can't both be pinned at frame 0. Two workable patterns:

Pattern A (recommended): composite reference image. Build one image with character A on the left and character B on the right (e.g., via PIL Image.paste or any image editor), pin THAT at frame_idx=0. Both identities transfer in one pin.

Pattern B: stagger the pins. Pin character A at frame 0, character B at a later latent boundary (e.g., frame 64 — must be a multiple of 8 per the VAE's temporal compression). Only works if B doesn't need to be visible from the very first frame.

Recommended parameters

Setting	Value
Resolution	1280 × 704 (16:9, native LTX-2.3 distilled training resolution)
Faster preview	960 × 544 (~40% faster, slightly less detail)
Frames	satisfy 8k+1 — e.g. 121 (5 s), 193 (8 s), 241 (10 s), 361 (15 s) at 24 fps
Strength	Standalone 0.7-0.9 · stacked with style LoRAs 0.3-0.5
Pin strength	0.85-0.95 for tight identity, 0.7 for looser "inspired-by"
Trigger word	None

Field-tested production tips

Quirks of this LoRA + the LTX-2.3 distilled backbone that aren't in the original card but matter in practice.

1. Repeat color tokens for dark-clothed characters

This LoRA has a light-wuxia-robe bias. Dark outfits drift toward white at low pin strength. Repeat the color token glued to each clothing noun:

text

BAD:  black fedora and black suit
GOOD: BLACK fedora, white shirt, BLACK suit jacket, BLACK trousers,
      ... BLACK suit, BLACK trousers throughout

Also bump pin strength to ~0.95 for color fidelity on dark outfits.

2. Never use quoted dialogue in prompts

This LoRA was trained on Chinese drama clips with burned-in Chinese subtitles. Any quoted dialogue (「…」 or "…") in the prompt causes the LoRA to hallucinate subtitle characters at the bottom of the frame. Single biggest gotcha.

text

BAD:  低声警告 「此茶不可饮！」    ← fake on-screen subtitles
GOOD: 低声急切警告她茶水有毒        ← clean output, indirect narration

If your app needs subtitles, burn them post-hoc via ffmpeg drawtext.

3. Avoid "object detaches" prompts during action

At high motion intensity, the model loses object tracking. A directive like "fedora flies off mid-spin and tumbles to the floor" produces broken output — the hat dematerialises. Either:

Keep the object attached and say so explicitly ("the fedora STAYS ON his head throughout the spin")
Or render attach + detach as two clips and concat

4. Cross-shot identity drift

For multi-shot dialogue scenes, character identity drifts across cuts. Workaround: re-pin the reference image at frame 0 of every shot. (Deterministic seed + same first-frame pin + same prompt scaffolding produces good repeatability.)

Render performance

Resolution: 1280 × 704, 121 frames @ 24 fps (~5 s output)
Hardware: NVIDIA A800 80 GB → ~70 s per shot
Output: mp4 with ambient audio track (no TTS)

On consumer hardware (RTX 4090 24 GB), expect ~3-4 minutes per shot.

Limitations

Subtitle hallucination with quoted dialogue (see tip #2)
Complex physical interactions (wrestling, hugging, intricate hand-on-hand) can deform
Tail-frame artifact of LTX-2.3 — last 6-8 frames may smear; trim post-hoc if needed
Action complexity ceiling — the 8-step distilled budget caps motion complexity at action peaks
Portrait orientation degrades identity (LoRA trained on landscape only)
Dual-character via two separate refs is awkward (see "How to use" above) — composite-image pin is the cleanest workaround

Original Chinese README (preserved)

The original Chinese model card from ModelScope is reproduced below for users who want the unmodified original documentation. (Note: the original card uses the "IC-LoRA" label — the term has been kept here for fidelity, even though the A/B/C test described above shows the conditioning mechanism is first-frame i2v pinning rather than parallel-canvas IC-LoRA.)

点击展开原版中文模型卡片 (click to expand original Chinese README)

LTX-Video (2.3) IC-LoRA: 双人分镜头对话增强模型

本模型是基于 Lightricks LTX-2.3 底模训练的 IC-LoRA，专为双人同框对话、角色互动及分镜头视频生成场景深度优化。

一、模型核心提升

角色参考稳定性：显著提升双人同框时的人物特征一致性，减少角色漂移。
分镜构图稳定性：针对影视化对话构图进行了加固，支持更精准的镜头控制。
叙事连贯性：增强了对多段描述的理解力，使分镜间的过渡衔接更自然。
风格兼容性：完美支持古风仙侠、现代都市、3D 动漫等主流视觉风格。

二、模型基本信息

基础模型：Lightricks/LTX-2.3
许可证：Apache-2.0
管道标签：image-to-video, text-to-video
模型用途：仅供学习交流使用
开发者：麻雀 AI

三、运行指南

推荐平台：ComfyUI
支持工作流：ComfyUI 官方 LTX 工作流、KJ-LTX 插件工作流
生成模式：文生视频 (T2V) 与图生视频 (I2V) 均支持
硬件参考：RTX 5090 显卡在 720P 分辨率下，单条视频生成耗时约 2 分钟

四、推荐参数配置

分辨率：建议使用 16:9 (如 1280x720)
时长与帧率：建议时长 ≥10 秒，帧率设定为 24 FPS
LoRA 权重设定：
- 独立使用建议：0.6 - 1.0
- 叠加其他 LoRA 使用时建议：0.3 - 0.5

五、Prompt 编写规范

编写逻辑：需包含完整的场景描述 + 角色设定 + 分镜设计 + 镜头语言，强化双人对话互动逻辑。
触发词说明：无需特定触发词。

六、效果说明与局限性

优势风格：在古风、现代、3D 动漫类双人对话场景中表现最佳。
已知限制：受限于 LTX-2.3 底模性能，极其复杂的双人肢体互动（如缠绕、打斗）可能出现形变。
运动幅度：建议以对话和微动作为主，大动态动作的连贯性仍有提升空间。

Hardware requirements

GPU	VRAM	Works?
A100 / A800 80 GB	80 GB	✅ ~70 s per 5 s shot
RTX 4090 / 3090	24 GB	✅ ~3-4 min per 5 s shot
RTX 4080 / 4070 Ti Super	16 GB	❌ won't fit 22B in bf16
anything < 24 GB	—	❌ no

Acknowledgements

麻雀 AI (Maque AI) — original author of this LoRA, original ModelScope repository
Lightricks — for the LTX-Video 2.3 base model
ZKong — for catching the IC-LoRA labeling discrepancy in the discussion thread; the empirical A/B/C test ran in response settled it

Source attribution

This is an English-language mirror of fxj1131's LTX2.3 Dual-Character LoRA on ModelScope. All credit for the model weights belongs to the original author, 麻雀 AI (Maque AI). This mirror exists to make the model + documentation accessible to HuggingFace users who cannot easily access ModelScope, and to share field-tested usage notes from a production deployment. The .safetensors weights file is unmodified and byte-identical to the ModelScope upload.

License

Apache License 2.0 — same as the original. See LICENSE and NOTICE.

SyFeee/LTX2.3-Dual-Character-en

LTX-Video 2.3 — Dual-Character LoRA (English mirror)