DEV Community

AI-KSK
AI-KSK

Posted on

Pinning down LTX 2.3 with the VBVR reasoning LoRA: same seed, one variable

#comfyui #ai #tutorial

The VBVR LoRA for LTX 2.3 — listed on Civitai as Video Reasoning LoRA — is the cleanest kind of upgrade to talk about, because the author runs it as a controlled experiment: lock the seed, change only the LoRA, show both results. What it buys is prompt adherence: the model does what the prompt says, in the right order, and the audio lands where the prompt says it should.

Evidence note: this episode does ship a workflow JSON, but the file is exported as a node group (subgraph), so the top level only exposes the loader and control nodes — UNET, checkpoint, LoRAs, a 2 GB VRAM reserve — while the sampler internals sit inside the collapsed group. No sampler step count or CFG is claimed for this graph. All LoRA weights, model filenames and test outcomes below are read from that JSON and from 字幕.srt; quality comparisons are the author's reported results.

Where this fits

You have an LTX 2.3 graph that mostly works but feels disobedient: the subject does the wrong action, the motion drifts, or the generated audio ignores the prompt. The VBVR LoRA is a one-node insertion (Load LoRA + the model) that the author describes as safe enough to drop into existing pipelines without hurting image texture.

What the LoRA actually does

Per the author, its first and biggest advantage is making generated video follow the prompt more closely. Beyond that: more stable temporal ordering and more accurate action. It is the same family as the Wan 2.2 VBVR LoRA he covered earlier, so if you already know that one, the mental model carries over.

The graph

  • Layers: a three-pass image-to-video render, wrapped into node groups.
  • Quick start: load the image, set the size, set frames and fps, write positive and negative prompts — then run. That is the whole quick start, because the author converted the complex part into groups.
  • Added for this version: an LTX 2 NAG node, so the negative prompt actually takes effect, and a VRAM reserve of 2 GB to avoid out-of-memory on the online service. He explicitly notes that omitting the reserve can still OOM.
  • Variant weights in the JSON: LTX2.3_Reasoning_V1.safetensors appears at 1.0, 1.5 and 0.75 across the test subgraphs, and ltx-2.3-22b-distilled-lora-384.safetensors at 1.0 and 0.5 — these are the A/B variants of the same test, not simultaneous loads.
  • Checkpoints: ltx-2.3-22b-dev.safetensors and ltx-2.3-22b-distilled.safetensors.
  • Size constraint: the first pass is small because the pipeline finishes with a 4x upscale; the author's advice is to size the first pass "within your means".

The controlled tests

Every test below locks the seed, so the LoRA is the only variable.

  1. Ears detached from the head. Without the LoRA, the head's ears detach from the figure. With it, no clear degradation appears. This is the clearest single demonstration in the episode.
  2. Weight sweep, 0.5 to 1.0. On this test the author reports no obvious difference between 0.5 and 1.0 — 0.5 is already enough to have an effect.
  3. Dancing in water. With the weight at 1.0 the character performs the dancing turn the prompt asks for; without the LoRA, or at a very low weight, the turn does not appear.
  4. Chess. At 0.75 the pacing is noticeably better.
  5. Fan dance, with audio. Without the LoRA the background music does not match the prompt description. Raising the weight produces the most harmonious result and the best match to the prompt — the audio is part of what the LoRA fixes, which is easy to miss in a video model.

Prompt rules

The LoRA only helps if the prompt is written the way it expects. The author's five rules, which he says you can hand to an LLM:

  1. Be concrete, not vague. "A ball bouncing, first moving right, hitting the wall, rebounding to the centre" — not a mood word.
  2. Write in order.
  3. Separate three stages: starting state, action, ending state.
  4. Name the subject's action explicitly.
  5. Avoid generic description (this overlaps rule 1, which is the point).

Inserting it into your own graph

One node: Load LoRA, load the model, connect. The author's recommendation is a weight of 1.0 — across most of his results there was no obvious degradation at 1.0, and unlike some other LoRAs in this line it does not visibly damage image texture.

Limits

  • The workflow JSON is a subgraph export: no sampler steps, CFG, sampler or scheduler values are readable or claimed here.
  • Weight behaviour is test-dependent: 0.5 sufficed in one test, 0.75 was better in another, and 1.0 was fine in most. There is no single correct value.
  • The prompt must follow the rules; a vague prompt will not be rescued by the LoRA.
  • Video model quality remains probabilistic; locking the seed makes comparisons honest but does not make results deterministic across hardware.
  • The 2 GB VRAM reserve is a mitigation, not a guarantee.

Episode resources

Top comments (0)