🎥 LTX See Motion — Image-to-Video Workflow (LTX-2.3)

Two-stage ComfyUI workflow for the custom LTX See Motion LTX-2.3 checkpoint. Turns a single input image into a video with natural motion, synchronized audio and dialogue — identity, style and composition of the source image survive the whole clip.

How it works: 🎬 Stage 1 renders the base clip (8 steps, CFG 1.0) → 🆙 Stage 2 upscales the latent 2x and refines it (4 steps, denoise 0.42) → ⚡ RTX Video SR sharpens another 1.5x → 💾 the final MP4 (H.264 + audio) saves automatically. Live low-VRAM preview included. Everything is grouped, color-coded and documented with README notes inside the workflow.

• Ships in FAST mode: distilled LoRA at strength 0.8, 8 + 4 steps, CFG 1.0

• DEV mode: disable the LoRA, set Stage 1 to 40 steps / CFG 4.0 (inside the 🎬/🆙 stage subgraphs)

• Strong: image-to-video (SFW + NSFW), dancing/singing, camera moves. Weak: plain text-to-video — always feed an image

• Performance reference: RTX 5060 Ti 16 GB ≈ 300–400 s for a 15 s clip at ~1 MP

📥 Required Models

• LTX See Motion.safetensors → ComfyUI/models/checkpoints/
  Download: https://civarchive.com/models/2768763/ltx-see-motion?modelVersionId=3116891

• Gemma 3 12B IT FP8 text encoder (Heretic V2 used here) → ComfyUI/models/text_encoders/
  Download: https://huggingface.co/DreamFast/gemma-3-12b-it-heretic-v2/tree/main/comfyui

• taeltx2_3.safetensors (tiny preview VAE) → ComfyUI/models/vae/
  Download: https://huggingface.co/Kijai/LTX2.3_comfy/resolve/main/vae/taeltx2_3.safetensors?download=true

• ltx-2.3-spatial-upscaler-x2-1.1.safetensors → ComfyUI/models/latent_upscale_models/
  Download: https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-spatial-upscaler-x2-1.1.safetensors?download=true

• ltx-2.3-22b-distilled-lora-384-1.1.safetensors (required for FAST mode) → ComfyUI/models/loras/
  Download: https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-22b-distilled-lora-384-1.1.safetensors?download=true

⚠️ Use spatial upscaler 1.1 — version 1.0 is no longer used by this workflow.
Video VAE and audio VAE load directly from the checkpoint — no separate VAE or text-projection files needed.

Recommended:

• Ollama + Qwen 3.5 2B — powers the automatic reference-sheet / character analysis inside LTX Director CS. Install Ollama from https://ollama.com, then run:
  ollama pull huihui_ai/qwen3.5-abliterated:2b (1.9 GB)
  Without it the node still works — you just have to type the character descriptions manually.

Optional:

• LTX-2.3 Licon MSR V2 LoRA — better multi-subject references / character sheets (keep the distilled LoRA loaded alongside):
  https://huggingface.co/LiconStudio/LTX-2.3-Multiple-Subject-Reference

• GGUF text encoder (lower RAM; swap the loader for the GGUF-compatible LTXV Audio Text Encoder Loader):
  https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/tree/main

• Full official LTX-2.3 repo (base models, the distilled LoRAs and all spatial/temporal upscalers):

  https://huggingface.co/Lightricks/LTX-2.3

🧩 Required Custom Nodes

Install via ComfyUI-Manager (search the name) — except LTX Director CS, which must be cloned manually (see below):

• LTX Director CS (CGlide mod) — 🎛️ the timeline/director node this workflow is built on — see the red box below for installation

• ComfyUI-KJNodes — SageAttention patches, LTX2 NAG, preview override — https://github.com/kijai/ComfyUI-KJNodes

• ComfyUI-LTXVideo — tiled decode, latent-upscale loader — https://github.com/Lightricks/ComfyUI-LTXVideo

• ComfyUI-VideoHelperSuite — video combine (MP4 + audio) — https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite

• ComfyUI-Lora-Manager — LoRA loader + trigger words — https://github.com/willmiao/ComfyUI-Lora-Manager

• ComfyUI_NVIDIA_RTX_Nodes — RTX Video Super Resolution — https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI

• rgthree-comfy — seed control — https://github.com/rgthree/rgthree-comfy

• ComfyUI-Easy-Use — clean GPU VRAM — https://github.com/yolain/ComfyUI-Easy-Use

⚠️ LTX Director CS cannot be installed through the ComfyUI-Manager!
Close ComfyUI, open a terminal in ComfyUI/custom_nodes/ and run:

git clone -b main_cs https://github.com/CGlide/WhatDreamsCost-CSGlide.git

The -b main_cs part is required — the default branch does NOT contain the CS nodes (LTXDirectorCS, LTXDirectorGuideCS, …). If the Manager's "Install Missing Nodes" already created a WhatDreamsCost-ComfyUI folder from CGlide, delete that folder first — it's the wrong branch. The original WhatDreamsCost node is not needed for this workflow; if you have it installed for other workflows, the CS clone coexists with it.

📺 Video Guides / Walkthroughs

Step-by-step walkthroughs for every path this workflow can do. Each one is a short screen recording with on-screen chapters.

• Image to Video — the basics: turn a single image into a video with natural motion.
  https://civarchive.com/images/136919778

• Video to Video — extend any existing clip to any length you like.
  https://civarchive.com/images/136922039

• Audio to Video — feed a 30 s song (with vocals) + one image and get a 30 s music video, lip-synced to the track.
  https://civarchive.com/images/137063968

• IC-LoRA to Video — drive motion, poses or depth from a control video using an IC-LoRA (like the one shown in the clip).
  https://civarchive.com/images/137075598

• LoRA + Image to Video — multi-subject / character-sheet consistency. Needs the LTX-2.3 Licon MSR V2 LoRA.
  Guide: https://civarchive.com/images/137093666
  LoRA: LTX-2.3-Licon-MSR-V2.safetensors
  Multiple-Subject-Reference LoRA by LiconStudio — keeps multiple characters and the background consistent across frames. Load it alongside the distilled LoRA.

🛠️ Installation

1. Install 7 of the node packs with ComfyUI-Manager; install LTX Director CS manually with the git command above → restart ComfyUI.

2. Download the models into the folders listed above → restart or refresh the model list. (Optional: install Ollama + the Qwen model for automatic reference analysis.)

3. Open the workflow and check every red loader: checkpoint, text encoder, tiny VAE, spatial upscaler 1.1, distilled LoRA at 0.8.

4. Load a start image in the 🎛️ LTX Director CS, write a short motion prompt, hit Queue — the finished video lands in output/ via the 💾 FINAL SAVE node.

📝 Notes

• ⚡ RTX Video SR needs an NVIDIA RTX GPU + a recent driver. No RTX card: bypass group 8 and wire the final decode IMAGE straight into the last Video Combine.

• 🔮 The optional speed patches (SageAttention, Chunk FeedForward, FP16 accumulation) ship bypassed — only enable them if the sageattention python package is installed.

• ✍️ Prompting: the image defines who/what is in the scene — the prompt defines what moves. One flowing paragraph, present tense, one main action per clip. A full prompting guide is included as a note inside the workflow.

Adult content: consenting-adult content only, where platform and law allow.

Credits: Lightricks (LTX-2.3) · TenStrip (10Eros) · sulphur · WhatDreamsCost (Director) · CGlide (Director CS mod) · Kijai · checkpoint blend + workflow: SeeSee21