Dev.to AI 🤖 Ai 👁 0 📖 3 min read

Why Your Video Writes But Will Not Move: Distillation LoRA Weight and the VBVR Fix

A generated clip where the subject is posed correctly but does not actually move is one of the most frustrating failure modes in video generation. The output looks fine as a still and dead as a video. This release is spe

A generated clip where the subject is posed correctly but does not actually move is one of the most frustrating failure modes in video generation. The output looks fine as a still and dead as a video. This release is specifically about cracking that, and the fix turns out to be a LoRA weight.

The three graphs

Workflow Nodes Likeness LoRA
Image-to-video (optimal) 14 ltx-2.3-22b-distilled-lora-384 at 0.5
High-probability digital human (optimal) 71 ltx-2.3-22b-distilled-lora-384 at 0.5
Four-grid multi-image first/last frame 89 Ltx2.3-Licon-VBVR-I2V-96000-R32 at 1.0

The weight that fixes motion

ltx-2.3-22b-distilled-lora-384.safetensors is the distillation LoRA — the component that lets the model run in few steps. In most LTX 2.3 workflows it is loaded at 1.0, because you want full distillation.

Here it is at 0.5.

That is the interesting choice. A distillation LoRA at full weight makes the model very efficient at producing a plausible frame. It also makes it conservative — the student learned to reproduce the distribution it was trained on, and the safest thing in that distribution is a small amount of movement.

Backing the distillation LoRA off to 0.5 restores some of the base model's willingness to move, at a cost in step efficiency. That is the trade this release is making: accept more sampling work in exchange for actual motion.

If your image-to-video output is "correct but static", the distillation LoRA weight is the first thing to check — before the prompt, before the sampler.

The VBVR LoRA at full weight

The multi-image graph uses Ltx2.3-Licon-VBVR-I2V-96000-R32 at 1.0.

Note the step count: 96000, not the 240K or 390K variants that appear in other releases. VBVR here is the image-to-video likeness carrier, trained for fewer steps than the larger variants. At full weight it holds reference adherence strongly.

The pairing is deliberate: image-to-video graphs get full-weight likeness because there is a reference to adhere to; prompt-only graphs get 0.5 distillation because there is no reference and the model must supply its own motion.

Checkpoint consistency

All three graphs load ltx-2.3-22b-distilled-1.1.safetensors — the 1.1 distilled build, not the older plain distilled and not the dev-dare-ties variants.

But check the audio VAE separately: in the digital human graph it loads ltx-2.3-22b-dev.safetensors. That is the non-distilled dev build, while the checkpoint is distilled. Two different weight sets in one graph.

This mismatch appears across many LTX 2.3 workflows in this family. It runs, it produces audio, and the audio is subtly wrong in a way that is hard to attribute. If you are tuning audio quality, this is the first thing to align.

The four-grid graph

89 nodes — the largest in the set — for a four-image first/last frame workflow. Four images means four reference slots plus the first/last frame pairing, and that is where the node count goes.

Worth noting: the grid structure means you are not just supplying four references, you are defining a spatial or temporal arrangement for them. Read the wiring before assuming it behaves like the three-image workflows.

Where this release is weak

  1. 0.5 distillation is a trade, not a free win. You pay in step count, and at very low step budgets the output regresses.
  2. The audio VAE mismatch in the digital human graph.
  3. 14 nodes is thin. The image-to-video graph has almost no internal validation; a bad LoRA weight or a missing model produces output that looks plausible rather than failing loudly.
  4. Four-grid is a different operation than multi-image reference; the naming invites confusion.

Practical approach

Start with the 14-node graph and try the distillation LoRA at 0.5 versus 1.0 on the same seed and prompt. The difference in motion is usually obvious within one comparison, and that comparison tells you whether this release's premise applies to your content.

Workflows

Three graphs, published on RunningHub with invite code rh-v1111:

  • Optimal LTX 2.3 Distill 1.1 + VBVR 240K image-to-video
  • Optimal LTX 2.3 Distill 1.1 + VBVR 240K high-probability digital human
  • Optimal four-grid multi-image first/last frame

Files

https://huggingface.co/AI-KSK/ksk-video-228-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • Node counts and LoRA weights quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with Lightricks or RunningHub.
📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.