DEV Community

AI-KSK
AI-KSK

Posted on

Eight LTX 2.3 workflows in one image: face swap, ID LoRA, digital human, repaint, multi-image keys

#comfyui #ai #tutorial

LTX 2.3's ecosystem filled in fast, and this episode is the consolidation: one image, eight graphs, all reachable from ComfyUI's left panel. The interesting part is not the count. It is that three of the eight — BFS video face swap, ID LoRA voice consistency, and audio-as-timbre-reference — are things people were stitching together by hand two weeks earlier.

Evidence note: this episode ships no workflow JSON (工作流链接：无). Everything below is from 字幕.srt plus the verified mirror page, which contributes the image size, runtime versions, launch command and deployment count. The 0.1-0.15 megapixel guidance and the mask-colour convention are the author's stated practice.

Where this fits

You want LTX 2.3's recent capabilities without assembling eight environments. Deploy one image, open one ComfyUI, pick a workflow. The author's own description of the situation: the community turned out a lot of interesting applications in a short time, and this packages them.

The image

  • Compshare: LTX2.3 创作增强旗舰版
  • https://www.compshare.cn/images/jjPe40PfI3Ff?referral_code=A7fTQUYY9lCDFJwA6MMvoq
  • Page facts confirmed by GET: 150 GB image, PyTorch 12.8 / CUDA 12.8, JupyterLab on 8888, 144 deployments, and the launch command cd /workspace/ComfyUI && source myenv/bin/activate && python main.py --listen 0.0.0.0 --port 8188 --lowvram.
  • It is built around LTX 2.3 and bundles ID-LoRA, face swap, local repaint and multi-image first/last-frame control.

Deploy

  1. Open the image, scroll the description, copy the launch command, click deploy.
  2. In JupyterLab, open a terminal anywhere, paste, enter.
  3. When the address appears, click ComfyUI, then click workflows in the left panel.

The eight graphs

1. Text-to-video

Positive prompt, fps, frame count — about 10 seconds of video in the demo. Size is set in the megapixel node at 0.1-0.15, deliberately small because the pipeline does a 4x upscale afterwards. Demo: pouring milk, described as passable.

2. Image-to-video

Identical controls plus one image. The one thing to change: describe what is in the image, and what should be said. This is an audio-visual model, so the prompt can carry speech — the demo has the character speaking a line. Omitting it also works.

3. Three-stage sampling

The pipeline the author has been refining for several episodes. His verdict: it drops the similarity penalty very low, and he considers it a mature, usable flow.

4. Video local repaint

A community LoRA turned into a workflow; compared with earlier model configurations, you only add one LoRA.

  • Load the video, set width/height and the duration you want.
  • You must build a mask video. Mark the region to keep with green points and the region to discard with red: Shift + left click = green, Shift + right click = red.
  • Demo: a dancing woman becomes a person in a Batman-style mask. The author's read is that this is close to real video editing, and far easier than the SkyReels-era approach.

5. Multi-image first/last frame

  • Four images in the shipped demo; set size, again at 0.1-0.15 MP.
  • Left column: where each image appears inside the 240-frame window. Right column: its weight.
  • A 12-image version is also included, which the author calls extremely fiddly.

6. Digital human (image + audio)

Load image, load audio, set how much audio to take (that sets the video length), size at 0.1-0.15 MP, fps 24, short prompt, run. The demo has the character delivering dialogue.

7. ID LoRA

The trick here is audio as a timbre reference, not as lip-sync. Load an image plus an audio clip, and the generated speech follows the reference voice — effectively skipping a separate voice-cloning step, which the author calls out explicitly in the demo. Want a different line? Write it in the prompt. The same audio-loading trick applies to the local-repaint graph (toggle with Ctrl+B).
To run it as text-to-video, flip the boolean and the image reference is skipped; then set fps.

8. BFS video face swap

First appearance of this flow on the channel. The model author trained a head-swap LoRA and wrote a node to drive it.

  • Load a video, load a face reference image, set the duration and the longest-edge size.
  • You can load your own audio, but you must reconnect the disconnected input on the right — otherwise the graph takes audio from the source video. The same pattern appears in the local-repaint graph.

Limits

  • No workflow JSON shipped, so no node-level parameters can be verified — the numbers above are spoken guidance.
  • The 0.1-0.15 MP first-pass rule assumes the 4x upscale; raising it without changing the pipeline will cost VRAM and time.
  • Mask colour is a convention you have to follow, not a preference (green keeps, red discards).
  • The face-swap and repaint graphs are community LoRAs wrapped into workflows; their behaviour tracks those LoRAs, which change independently of the image.
  • The 12-image variant is described as fiddly by the author himself.
  • The author's own three-stage flow is the one non-community part of the set; the rest is close to the authors' or official workflows.

Episode resources

Top comments (0)