A point release with no announcement, and the practical effect is larger than the version number suggests. The author swaps only the model loader, drops the acceleration LoRA entirely, and gets digital humans whose mouths finally move and image-to-video that holds its weapon and flame effects to the end of the clip.

Evidence note: three workflow JSON files ship with this episode and were parsed directly; checkpoint, LoRA names and strengths, samplers, sigma schedules and latent sizes below come from those files.

When this is useful

  • Upgrading an existing LTX 2.3 pipeline with a minimal edit: one loader swap.
  • Digital human where the mouth previously refused to move on non-human or stylised faces.
  • Image-to-video with weapon or effect elements that used to deform late in the clip.
  • Choosing a voice-clone model, with a three-way comparison in this episode.

How this differs from the previous build

The migration is the news here, and it is close to one node.

  • Drop the accelerate LoRA. With the 1.1 distilled model there is no need for the separate acceleration LoRA; the author removes both the LoRA and the Dev model.
  • One loader swap. Because the author uses a split model, only Load Diffusion Model changes — the rest of the graph is untouched.
  • Digital human now works on stylised heads. In the earlier 2.3 flow the pig-head test produced no mouth movement at all; in 1.1 it triggers properly.
  • VBVR stays in. The reasoning LoRA is still loaded on top of the new base.

Running it

  1. Swap the loader to the 1.1 distilled model, then delete the acceleration LoRA.
  2. Keep the VBVR LoRA connected — the author does not remove it in the upgrade.
  3. For digital human: load image, load audio, set the audio trim length, set frame rate, write a short prompt.
  4. Run and compare against your 1.0 output on the same seed to see the dynamics difference.
  5. For the voice track, normalise the reference audio before cloning. The author calls this node the difference between a good and a terrible result on one of the models.

Parameters, and where the numbers come from

Item Value in the shipped graphs
Checkpoint ltx-2.3-22b-distilled-1.1.safetensors
Text encoder gemma_3_12B_it.safetensors
Samplers euler_ancestral_cfg_pp, euler_cfg_pp
CFG 1 (CFGGuider)
Sigma schedule, main 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0
Sigma schedule, refine 0.85, 0.7250, 0.4219, 0.0
Video latent 320x288, length 97
Audio latent 25 fps, 97 frames
Resize for references 960x544
NAG scale 11, alpha 0.25, tau 2.5
Image conditioning strength 0.7 and 1
LoRA, VBVR Ltx2.3-Licon-VBVR-I2V-96000-R32.safetensors at 1.0
LoRA, accelerate ltx-2.3-22b-distilled-lora-384.safetensors at 0.5 (still present in these graphs)
LoRA, motion track ltx-2.3-22b-ic-lora-motion-track-control-ref0.5.safetensors at 1.0

The accelerate LoRA is still present in the shipped JSON at 0.5, which contradicts the narration's advice to remove it. If you follow the video, delete it; if you run the file as shipped, it stays.

Honest limits

  • The 1.1 model is still ~46GB in its full form; the LoRA route exists for people without the headroom.
  • The author reports the character distorts slightly at the tail of one 1.1 image-to-video result, and attributes it to chance rather than the model.
  • Voice cloning is judged by ear: LongCat won on timbre and emotion, VoxCPM2 sounded flatter, OmniVoice was compared only on cloning.
  • LongCat needs the audio-normalisation node enabled; bypassing it with a dynamic reference track produces very bad output.
  • The narration and the shipped JSON disagree about the accelerate LoRA — treat the graph as the literal artifact and the video as advice.

Episode resources

RunningHub workflows