跳到正文
DEV Community· AI-KSK·· 13 天前AI 評分59

LTX Director Node 以單一時間軸整合五套 LTX2.3 工作流

LTX Director Node: One Timeline Instead of Five Separate LTX2.3 Workflows

AI 導讀

LTX Director Node 把文字轉影片、圖像轉影片、數碼人、多圖及首尾幀五套 LTX2.3 工作流改為單一時間軸,讓用戶按時間加入圖像、文字和音訊。音訊只有在自訂音訊開關啟用時才會參與生成;作者把每次生成 20-25 秒列為舒適範圍,並指出三階段渲染的第二、三階段音訊輸入仍需完善。作者比較 HiDream-O1 與 2511,認為 2511 在編輯表現上較佳。

正文

AI-KSK

The LTX director node replaces the pile of separate LTX2.3 graphs - text-to-video, image-to-video, digital human, multi-image, first/last frame - with a single timeline where you drop images, text and audio at times you choose. This episode walks the node, adds the OmniNFT audio-video LoRA, and ends with a quick HiDream-O1 versus 2511 comparison.

Evidence note: one workflow JSON ships with the episode; the sampler/scheduler values and model files below are read from it. Behaviour and the comparison verdict are the author's, as reported in the video.

When to use it

  • You are building a clip where narration, dialogue and shot changes must land on specific seconds.
  • You want multi-image and digital-human work on one timeline instead of swapping graphs.
  • You need 20-25 seconds of output per run without visible degradation.

What is different in this workflow

  • Timeline as the interface: the first field is total duration in seconds (switchable to frames); fps defaults to 24.
  • Three insert types: Add Image, Add Text, Add Audio. Audio only drives the render if the custom-audio toggle is on.
  • Transition sharpness is one number: the author explains 0.001 as very rigid separation between text segments, larger values blend them.
  • Size alignment to 32 px, plus an image compression strength value folded into the node.
  • Global prompt is optional but recommended for story-driven clips; per-image guide strength (1.0 or 0.5) controls the shot's binding strength, not identity similarity.

Step by step

  1. Set total duration (seconds or frames) and keep fps 24.
  2. Set width/height; the author says 2K is fine because a 2x upscale follows.
  3. Pick a scale mode: keep ratio, stretch, crop, or pad (pad gives black bars). For mixed-size inputs the author recommends crop.
  4. Click Add Image / Add Text / Add Audio. Turn on the custom-audio switch or timeline audio is ignored.
  5. Optionally enable the global-prompt checkbox and write the through-line.
  6. Tune the 0.001 transition value: smaller = harder cuts between segments.
  7. Set per-image guide strength (1.0, or 0.5 for softer binding); the author reports 3-4 images inside a 10-15 s piece strengthens guidance with minimal similarity decay.
  8. Run; 20-25 s per generation is the author's comfortable range.

Parameter basis

  • Graph: ltx导演节点.json, 35 nodes.
  • Checkpoint: CheckpointLoaderSimple -> ltx-2.3-22b-dev-dare-ties-distilled-1.1.safetensors.
  • Text encoders: DualCLIPLoader -> gemma_3_12B_it_fp8_e4m3fn.safetensors + ltx-2.3_text_projection_bf16.safetensors, type ltxv.
  • LoRA: LoraLoaderModelOnly -> OmniNFT-comfyui.safetensors weight 1.0.
  • Schedulers: BasicScheduler linear_quadratic, 8 steps, denoise 1.0 (pass 1) and linear_quadratic, 4 steps, denoise 0.42 (pass 2).
  • Sampler: KSamplerSelect -> euler, CFGGuider CFG 1, RandomNoise seed 12 fixed, SamplerCustomAdvanced x2 with LTXVConcatAVLatent / LTXVSeparateAVLatent.
  • Video out: CreateVideo + VHS_VideoCombine; LTXVLatentUpsampler with LatentUpscaleModelLoader for the 2x.

Limits

  • The author calls this a 3.0-version director update, not a finished pipeline: he plans to polish it before moving on.
  • A finished three-stage render still needs work on the director path: audio input during second/third stage needs the same set-latent-noise-mask dance as the digital-human graphs.
  • HiDream-O1 has no text encoder and no VAE (pixel-space): the author reports graying, detail loss and weaker instruction following in both Base and Demo, and still ranks 2511 ahead for editing.
  • OmniNFT is described as a drop-in gain, not a new capability.

Episode resources

來源:DEV Community · dev.to