LTX 2.3 understands physics poorly, and no amount of prompt polish fixes that. This LoRA targets exactly that gap: it improves complex instruction understanding, logical action execution and frame stability. In the domino test the base model topples the wrong end; in the weight-on-box test the base model cracks the pedestal before the weight even lands.

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

When this is useful

  • Prompts describing a sequence of actions that the base model ignores or reorders.
  • Object interactions where contact and physics have to look plausible.
  • Hands and grip: the author's mirror and helmet tests both improve with the LoRA.
  • Inserting the LoRA into first/last-frame and digital-human graphs, not only plain image-to-video.

How this differs from the previous build

This is the third VBVR iteration the author has covered, and the first from this particular studio.

  • It is an I2V LoRA by the trainer's own description, so most demos are image-to-video.
  • It also slots into first/last-frame and digital-human graphs, because LTX 2.3 digital human still starts from an image.
  • The published graphs hide redundancy in node groups. Everything you would normally adjust is left outside the groups, and professional users can ungroup to reach the rest. Nothing is password-protected or hidden.
  • Prompt rules come with the LoRA. Concrete, ordered, stepwise descriptions; vague mood words actively hurt.

Running it

  1. Load the LoRA with a standard Load LoRA node on top of the distilled checkpoint.
  2. Write the prompt in three parts: start state, action sequence, end state. Keep each step concrete and in order.
  3. Drop abstract adjectives — the author specifically calls out vague dynamic or dreamy phrasing as counterproductive.
  4. Run image-to-video first to feel the difference, holding the seed fixed against a no-LoRA run.
  5. For first/last-frame or digital-human work, drop the same LoRA into those graphs.
  6. If you need finer control, right-click the node groups and ungroup them.

Parameters, and where the numbers come from

Item Value in the shipped graphs
Checkpoint ltx-2.3-22b-distilled.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

Note the checkpoint here is the plain distilled build, not 1.1 — this episode predates the 1.1 release covered in episode 223.

Honest limits

  • The LoRA improves compliance, not physical realism in absolute terms; the author says LTX 2.3's physics adherence is poor to begin with.
  • Mirror reflections remain wrong even when the hands improve — the author concedes this while still calling the result better.
  • Every comparison is a side-by-side visual with the seed held; nothing is scored.
  • The studio behind the LoRA is still training, so this is an interim checkpoint.
  • This episode uses the 96000 build; later episodes ship 240K and 390K versions that are not interchangeable.

Episode resources

RunningHub workflows