demobook

ComfyUI: Comparing LTX 2.3 Vanilla vs. TensNodes for Facial Expressions

ComfyUITry it →Watch full video →SOTAI ·

Demo summary

A side-by-side demonstration shows the TensNodes workflow maintaining a consistent facial structure during a smile and wave, whereas the vanilla LTX model drifts from the reference image.

Step-by-step

  1. Upload a reference photo of the subject to both workflows
  2. Enter a prompt describing the desired action or facial expression
  3. Set identical core parameters and random noise seeds for both the TensNodes and vanilla LTX workflows
  4. Run the split test to generate the video-to-video comparison
  5. Inspect the results frame-by-frame to check for identity drift

Watch out for

  • Vanilla LTX workflows may cause identity drift, such as elongating jawlines or changing ethnic features, during complex facial expressions

Tips

  • Keep every core parameter and the random noise seed identical when performing side-by-side workflow comparisons for objective results
  • Pixel peep frame-by-frame against the original reference to identify subtle morphing in facial structure
  • Use the TensNodes workflow for scenarios requiring complex emotions to prevent facial warping

Highlights

“you guys can see the night and day difference, right? Her face is anchored flawlessly to that input image across the entire sequence of frames.”

All demos from “Consistent Character in LTX 2.3 ComfyUI: STOP Losing Your Character's Face Between Scenes”

  1. 4:261:28Overview of the TensNodes Consistent Character Workflow in ComfyUIThe creator walks through a four-stage ComfyUI workflow utilizing TensNodes to process reference images and text for stable LTX video generation.ComfyUIAI Animation Generator
  2. 5:541:04Configuring the Latent Anchor Aware NodeThe demonstration shows how to adjust the strength, cache at step, and similarity threshold parameters of the Latent Anchor Aware node to prevent background and detail morphing.ComfyUIAI Image Generator
  3. 6:581:02Setting up Likeness Guide and Likeness Anchor NodesThe video shows the Likeness Guide node detecting a face from a reference photo and the Likeness Anchor node locking the facial region to maintain identity during expressions.ComfyUIAI Avatar Generator
  4. 8:392:29Comparing LTX 2.3 Vanilla vs. TensNodes for Facial ExpressionsCurrentA side-by-side demonstration shows the TensNodes workflow maintaining a consistent facial structure during a smile and wave, whereas the vanilla LTX model drifts from the reference image.ComfyUIVideo to Video
  5. 11:082:10Testing Long-Duration Character Consistency (20 Seconds)The creator demonstrates a 20-second video generation where TensNodes successfully preserves the character's face after a full body turn, compared to the base model which fails.ComfyUIAI Animation Generator
  6. Watch “Consistent Character in LTX 2.3 ComfyUI: STOP Losing Your Character's Face Between Scenes” →

Video to Video

  1. 0:000:54Mocha character replacement showcaseThe video demonstrates Mocha's ability to replace characters in existing videos with new characters from reference images while maintaining original movements and lighting.AI Search
  2. 2:280:44Swap human character for 3D Pixar model in MochaThe creator demonstrates swapping a real person in a video with a 3D Pixar-style character using Mocha, showing the transfer of lip movements and hand gestures.AI Search
  3. 20:110:27Generating final character swap videoThe video shows the K-Sampler process and the final rendered output of a character replacement task within ComfyUI.AI Search
  4. 15:344:59Character replacement with Wan AnimateThe creator demonstrates using Wan Animate and SAM 3 to mask a character in a reference video and replace them with a new character from a static image while maintaining the original motion.pixaroma
  5. 21:072:43Cartoon animation with Wan SCAILThe user shows how to animate a cartoon ballerina using a real-person video as a motion reference via the Wan SCAIL workflow in ComfyUI.pixaroma
  6. 8:150:26Using Keyframes to guide video generationThe demo shows how to slide an image to a later point in the timeline to use it as a target keyframe, allowing the model to generate the action leading up to that specific visual.What Dreams Cost
  7. 15:352:08Final video generation with Wan 2.1The demonstration shows how to use the Wan 2.1 V2V model to combine the generated 3D motion and a reference character image into a final, high-quality AI video animation.Benji’s AI Playground
  8. 1:003:01Generate character-consistent video with Wan 2.6 R2VThe creator demonstrates a ComfyUI workflow using the Wan 2.6 API node to transform a 5-second reference video of himself into a new scene while maintaining character consistency and native audio lip-sync.Sebastian Kamph
  9. 4:031:00Generating Guide Videos with BFS NodeThe demo shows how to use the BFS Node extension to combine a reference face image and a driving video into a green-screen guide video required for the LTX sampling process.Veteran AI
  10. 5:030:47Injecting Latent Constraints with Add Guide MultiThe video demonstrates using the 'Add Guide Multi' node from the KJ Node extension to inject the guide video constraints and VAE-encoded latents into the sampling pipeline.Veteran AI
  11. 15:031:04Apply camera movement from reference videoThe creator demonstrates copying camera motion and pose dynamics from a reference video (like a dance clip) and applying it to a still character image.Yaroflasher
  12. 1:551:00Adding reference images for character consistencyThe video shows how to use the MSR custom node to input up to four reference images and a background image to maintain character facial consistency across video frames.bigboss97
demobook

Show, don’t tell.

Categories

Video & AnimationImage Generation & EditingMusic & AudioSales TechEmail Marketing

About

SubmitAboutPrivacyTermsContact

© demobook.co