开源AI视频模型领域近日迎来重大更新！本次的主角是 <strongLTX 2.3 </strong——Lightricks团队打造的新一代开源音视频生成模型。相较于前代版本，该模型在视频画质、动态表现以及提示词理解能力上均有显著提升，更令人惊喜的是，它支持在本地环境中同步生成视频与音频，且官方已公布完整开放权重，用户可自由进行本地部署。
<img src="https://www.foxnan.com/wp-content/uploads/2026/08/6f5c2d21a45dabe.webp" alt="" width="2560" height="1440" class="alignnone size-full wp-image-25082" decoding="async"</p
<div class="code-block code-block-center code-block-1 ai-track" data-ai="WzEsMCwiQmxvY2sgMSIsIiIsMV0="
对于本地AI爱好者而言，这意味着你可以摆脱对在线平台的依赖，视频生成能完全在自己的设备上完成，模型、素材及生成流程尽在掌控。更值得关注的是，LTX 2.3 已在Hugging Face平台上开放了基础模型、FP8、NVFP4版本，以及丰富的社区量化和扩展模型，玩法极为多样。
那么LTX 2.3 的实际生成效果如何？8GB显存是否真能流畅运行？RTX 3060、3070、4090等不同显卡应选用哪个版本？又该如何在ComfyUI中完成安装配置？本文将带领大家按照<strong模型下载 → 环境搭建 → ComfyUI配置 → 量化版本选择 → 实际生成测试</strong的完整流程，全面体验这款最新开源AI视频模型。
一、安装最新版ComfyUI
要支持新版LTX 2.3 越狱模型，必须使用最新版本的ComfyUI客户端！若已安装过旧版，请重新下载并覆盖安装。
网盘下载：【🔒 下载链接仅对登录用户可见，请 登录 或 注册 后查看。
二、下载模型文件
1、越狱主模型
PinkCherry_FineTune_bf16_v1_8_LTX23（满血版）
PinkCherry_FineTune_Q5_K_M_v18_LTX23 （8G显存版）
2、LoRA
LTX-2.3ltx-2.3-22b-distilled-lora-384-1.1
【点击下载
3、VAE
共需3个文件
LTX23_video_vae_bf16
【点击下载
taeltx2_3
【点击下载
LTX23_audio_vae_bf16
【点击下载
4、文本编码器
共需2个文件
gemma-3-12b-it-heretic-v2
<div class="code-block code-block-center code-block-1 ai-track" data-ai="WzEsMCwiQmxvY2sgMSIsIiIsMl0="
【点击下载
ltx-2.3_text_projection_bf16
【点击下载
模型存放路径：C:Users你的用户名AppDataLocalComfy-DesktopComfyUI-Sharedmodels 对应的文件夹中
对应关系：
文件
放置路径（相对 ...ComfyUI）
ltx-2.3-22b-distilled-lora-384-1.1.safetensorsmodelslorasLTX-2.3 ← 注意要新建一个 LTX-2.3 子文件夹
LTX23_video_vae_bf16.safetensorsmodelsvae
taeltx2_3.safetensorsmodelsvae
LTX23_audio_vae_bf16.safetensorsmodelsvae
gemma-3-12b-it-heretic-v2.safetensorsmodelstext_encoders
ltx-2.3_text_projection_bf16.safetensorsmodelstext_encoders
重点提示 ：LoRA模型文件需放置于modelslorasLTX-2.3文件夹内（LTX-2.3需手动创建）
三、越狱模型工作流 下载完成后，直接将其拖入ComfyUI的空白工作区即可。首次拖入时系统会提示缺少节点文件，点击详情中的下载按钮便可自动完成安装！
若你选用的是（8G显存版）PinkCherry_FineTune_Q5_K_M_v18_LTX23 .gguf格式的越狱模型，则需在管理扩展功能中心安装ComfyUI-GGUF节点包，否则无法识别GGUF格式模型。视频
<img src="https://www.foxnan.com/wp-content/uploads/2026/08/be6d6101fd88a69.webp" alt="" width="364" height="573" class="wp-image-25085 aligncenter" decoding="async"</p
删掉当前这个Checkpoint加载器，将新节点的MODEL 输出端连接到原来那根紫线上
<img src="https://www.foxnan.com/wp-content/uploads/2026/08/93bfc5c765ee712.webp" alt="" width="2560" height="869" class="alignnone size-full wp-image-25086" decoding="async"</p
<img src="https://www.foxnan.com/wp-content/uploads/2026/08/11bcf5911dabb3e.webp" alt="" width="301" height="433" class="alignnone size-full wp-image-25092" decoding="async"</p
0-3秒：
3-7秒：
7-11秒：
11-15秒：
说完轻轻一笑，目光停留在镜头上。最后一片花瓣从镜头前缓缓飘过，画面停留在女子温柔的眼神上。
整体风格：
镜头运动平稳缓慢，人物动作自然连续，避免突然运动。
2、生成AI短剧（图生视频）
Cinematic realistic image-to-video scene, 10 seconds.
Keep the woman's face, hairstyle, clothing, body proportions and the original room completely consistent with the reference image. She is sitting indoors, speaking naturally to someone off camera.
0-3 seconds:
She continues speaking softly, looking slightly toward the right side of the frame. Her lips move naturally and accurately synchronized with the Chinese dialogue. She has a calm and natural expression.
Dialogue in natural Mandarin Chinese, female voice:
"你终于来了，我等你很久了。"
3-6 seconds:
She suddenly notices something unusual behind the camera. Her speech stops briefly. Her eyes shift slightly, her expression becomes confused and cautious. Natural breathing and subtle facial movement.
6-8 seconds:
She slowly turns her head toward the camera and looks directly into the lens. Her expression becomes slightly nervous and serious.
8-10 seconds:
She quietly whispers in Mandarin Chinese:
"但是……你身后有人。"
Her lips clearly articulate the Chinese words with accurate lip synchronization. Her voice is a natural young Chinese female voice, soft, realistic, slightly nervous, conversational, not exaggerated.
Audio:
Natural Mandarin Chinese female dialogue, clear human voice, accurate lip synchronization, subtle breathing, quiet indoor room ambience, very subtle background noise, cinematic realistic sound design. No music.
Camera:
Slow cinematic push-in toward her face, very subtle handheld movement, shallow depth of field, realistic focus, natural lighting.
Photorealistic cinematic image quality, realistic skin texture, natural eye movement, subtle facial expressions, realistic hair movement, smooth motion, consistent identity.
No subtitles, no text on screen, no extra people, no voice-over, no robotic voice, no exaggerated facial expressions, no face distortion, no lip deformation, no sudden camera movement.
<img src="https://www.foxnan.com/wp-content/uploads/2026/08/c33bd84d626c459.webp" alt="" width="759" height="427" class="wp-image-25095 aligncenter" decoding="async"</p
<div class="donation-modal" id="donationModal"
感谢您的支持
<img src="https://www.foxnan.com/wp-content/uploads/2026/08/11c254db930eb75.png" alt="alipay" class="payment-icon" decoding="async"</div
<span支付宝</span
</div
<div class="donation-option" data-type="wechat"
<div class="payment-icon-wrapper"<img src="https://www.freedidi.com/wp-content/plugins/donation-button/images/wechat.png" alt="wechat" class="payment-icon" decoding="async"</div
<span微信支付</span
</div
<div class="donation-option" data-type="bitcoin"
<div class="payment-icon-wrapper"<img src="https://www.freedidi.com/wp-content/plugins/donation-button/images/bitcoin.png" alt="bitcoin" class="payment-icon" decoding="async"</div
<spanBitcoin</span
</div
<div class="donation-option" data-type="usdt"
<div class="payment-icon-wrapper"<img src="https://www.freedidi.com/wp-content/plugins/donation-button/images/usdt.png" alt="usdt" class="payment-icon" decoding="async"</div
<spanUSDT</span
</div
</div
<div class="qrcode-display" id="qrcodeDisplay"
</div
<div id="donationAddressArea" style="display:none;"
<button class="back-btn" id="backBtn"返回选择</button
</div
<div class="clear"</div
安卿辰博客
