Dev.to AI 🤖 Ai 👁 0 📖 3 min read

VBVR: A LoRA That Improves Instruction Following, Not Image Quality

Most video LoRAs are sold on sharpness. This one is different — VBVR targets the model's ability to understand and execute complex instructions, which is a different axis entirely and harder to evaluate by looking at a s

Most video LoRAs are sold on sharpness. This one is different — VBVR targets the model's ability to understand and execute complex instructions, which is a different axis entirely and harder to evaluate by looking at a single frame.

The three graphs

Workflow Nodes LoRA
VBVR multi-image first/last frame 78 Ltx2.3-Licon-VBVR-I2V-96000-R32 at 1.0
VBVR image-to-video 15 ltx-2.3-22b-distilled-lora-384 at 0.5
VBVR digital human 71 Ltx2.3-Licon-VBVR-I2V-96000-R32 at 1.0

Note the checkpoint: ltx-2.3-22b-distilled.safetensors — the plain distilled build, without the 1.1 revision that later releases in this family use.

That is expected: this release predates the 1.1 update. If you are mixing workflows from both, the checkpoint name is where the difference lives.

What "reasoning enhancement" actually means here

Look at the two LoRA families in play:

Ltx2.3-Licon-VBVR-I2V-96000-R32 — the VBVR likeness / reasoning LoRA, at full weight 1.0 in the two large graphs. The 96000 is a step count, lower than the 240K and 390K variants elsewhere. It appears at full weight when the graph is reference-driven.

ltx-2.3-22b-distilled-lora-384 at 0.5 — the distillation LoRA, backed off, in the 15-node prompt-only graph.

The claim of VBVR is that it improves complex instruction comprehension, logical execution and frame stability — not sharpness. That means the way to evaluate it is not a still frame. It is:

  • Give the model a multi-step instruction ("pick up the cup, turn it over, set it down") and see whether all three actions appear in order.
  • Give it a spatial instruction and see whether the relationship holds across frames.
  • Give it a duration-sensitive instruction and see whether the timing is right.

Those are checkable, and they are exactly the things a sharpness-oriented LoRA would not improve.

Why instruction following is the harder problem

A model that produces beautiful frames but ignores half the prompt is less useful than a model that produces adequate frames and does what you asked. In production, the failure that costs time is not softness — it is re-rolling twenty times because the subject keeps doing the wrong thing.

VBVR targets that failure. Whether it succeeds is something you have to test on your own content, because "complex instruction" is content-dependent.

The reference-driven versus prompt-driven split, again

The same pattern appears here as in adjacent releases:

  • 78 and 71 nodes → likeness LoRA at 1.0 → reference-driven
  • 15 nodes → distillation LoRA at 0.5 → prompt-driven

Once you have seen this pattern three or four times, node count becomes a reliable indicator of which mode a graph is in. That is worth internalising — it saves reading the graph every time.

The persistent audio VAE mismatch

Both large graphs load:

  • CheckpointLoaderSimple → ltx-2.3-22b-distilled.safetensors
  • LTXVAudioVAELoader → ltx-2.3-22b-dev.safetensors

Distilled video, dev audio. Same mismatch as everywhere else in this family. One filename change to align them, and it is worth trying once to hear whether it matters for your content.

Where this release is weak

  1. The claim is not quantifiable from the files. "Better instruction following" needs a test protocol, and the workflows do not provide one.
  2. Audio VAE mismatch in both large graphs.
  3. The 15-node graph has VBVR in its name but does not load the VBVR LoRA. It loads the distillation LoRA. That is a naming inconsistency that will confuse anyone reading the graph expecting reasoning enhancement.
  4. 96000 versus 240K/390K — three step-count variants of the same LoRA family across releases, with no guidance on when to use which.

How to actually test it

Pick one instruction with three distinct required actions. Render it with and without the VBVR LoRA at matched seed and prompt. Score on: how many of the three actions appear, in order, without extra invented actions. That is a measurable result, and it is a better use of an afternoon than comparing stills.

Workflows

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

  • LTX 2.3 + VBVR multi-image first/last frame
  • VBVR image-to-video
  • VBVR digital human

Files

https://huggingface.co/AI-KSK/ksk-video-221-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 values 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.