license: other base_model: ltx-video-2b # or ltx-video-22b depending on your base tags:
- text-to-video
- image-to-video
- ltx-video
- ltx2.3
- cinematic
- motion-control
- lora
- physics-fix model-index:
- name: LTX2.3_Singularity_OmniCine_Motion_Preview results: []
> The 10% Miracle.
> This LoRA is a comprehensive optimizer for LTX2.3 Image-to-Video (I2V) and First/Last Frame workflows. Although only 11,500 steps (10% of the target) have been completed, the results in physical coherence and cinematic expression have already exceeded all expectations.
Significant optimization for fingers and toes, reducing common AI distortions and maintaining structural integrity during movement.
- Video Preview:
<video controls autoplay muted loop src=https://huggingface.co/WarmBloodAban/Singularity/resolve/main/assets/VX-paolaoshiAICG__00016_p84-audio_bccco_1778276150.mp4 style="width: 100%; border-radius: 10px;">
- Prompt Window: > `[Macro realism style, soft indoor daylight, calm mood. 0-1秒, [Thumb with 3D floral nail art displayed, calm emotion, static extreme close-up]. 1-4秒, [Hands cut white and grey plastic parts from runners using a hobby knife, focused precision, top-down macro close-up, jump cuts]. 4-8秒, [Hands assemble blue, yellow, and red torso components and shield, static macro close-up]. 8-15秒, [Fully assembled Gundam model stands on desk, hands hold and adjust the figure, medium close-up, slight pan]. Indoor desk scattered with plastic sprues, soft overhead daylight, low contrast, natural warm skin tones, vibrant primary model colors. Soft relaxing piano BGM, SFX of plastic clicking and precise cuts.macro lens, shallow depth of field, stop-motion editing style，无字幕]`
The first implementation of precise shot-switching logic via prompt control. Use the `0-X seconds` format to dictate exactly when the camera cuts or moves.
- Video Preview: <video controls autoplay muted loop src=https://huggingface.co/WarmBloodAban/Singularity/resolve/main/assets/VX-paolaoshiAICG__00013_p81-audio_epjdi_1778329533%20(1).mp4 style="width: 100%; border-radius: 10px;">
Eliminates the "robotic" look of AI characters. Features natural facial micro-expressions and optimized lip-syncing for Chinese speech while removing hardcoded AI subtitles.
- Video Preview: <video controls autoplay muted loop src=https://huggingface.co/WarmBloodAban/Singularity/resolve/main/assets/VX-paolaoshiAICG__00006_p84-audio_zqcpl_1778262776%20(1).mp4 style="width: 100%; border-radius: 10px;">
Improved limb consistency under high-dynamic movement. Reduces "erratic" or "physics-defying" glitches, ensuring movements follow natural logic.
Despite no specific anime training, the model demonstrates incredible capability in generating and maintaining stylized/anime aesthetics.
- Video Preview: <video controls autoplay muted loop src=https://huggingface.co/WarmBloodAban/Singularity/resolve/main/assets/VX-paolaoshiAICG__00012_p84-audio_azumf_1778274653.mp4 style="width: 100%; border-radius: 10px;">
- Base Model: `ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors`
- Online Demo: https://www.runninghub.ai/post/2053064469290463233/?inviteCode=sdhs0trb
To achieve the best results, please strictly follow this bilingual logic: English for Visuals, Chinese for Time & Dialogue.
- Scene & Style: Core visual theme (e.g., Cinematic Wuxia style).
- Visual & Action Timeline: `0-X seconds, [English action description]`.
- Camera Timeline: `0-X seconds, [Camera movement/parameters]`.
- Environment & Lighting: Ambient details (e.g., High contrast, Rim light).
- Dialogue Timeline: `0-X seconds, [Character]说：[Chinese Content]`.
- Audio & SFX: Background music and sound descriptions.
- Ending Tag: `, 无字幕` (No subtitles).
> Cinematic wuxia style, indoor dim lighting, mysterious mood. 0-10 seconds, young man in ancient white robes looks down with a confused expression. 0-10 seconds, tight close-up, static camera with slight handheld movement. Dark stone background, warm candlelight bokeh. 0-10 seconds，男子说：这都什么呀？都没听说过呀。Voice: low and confused, Pace: slow. Precise lip-sync, film grain, cinematic bokeh，无字幕。
I believe LTX2.3 requires a more "logical" approach to motion training rather than brute-force data stacking. This fine-tune focuses on "understandable" motion sets, bridging the gap between static frames and complex acting.
This is just the beginning. If this preview helps your workflow, please leave a Star or feedback. I will continue to optimize the dataset and push toward the 100% official release based on community input!
Download them in the Files & versions tab.
If you have questions, feedback, or want to collaborate on AI video workflows, feel free to reach out:
WeChat (微信): aigctyd
QQ Group (QQ社群): 1058747239
Email: a592991299@gmail.com
I'm actively looking for community feedback to refine the full version of this LoRA. Let's push the boundaries of LTX2.3 together!
