DEV Community

AI-KSK
AI-KSK

Posted on

LTX 2.3 with VBVR 240K and Three-Tier Rendering: Three Optimal Graphs

#comfyui #ai #tutorial

The complaint was that LTX 2.3 would not move — prompts described motion that never appeared. The 240K VBVR model is the fix, and this episode is the three graphs where the author considers the combination settled: image-to-video, digital human, and a four-image first/last-frame build with a crop group for grids.

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

  • Clips where the prompt asks for a specific action and the base model simply does not perform it.
  • Digital human where the identity has to survive clothing detail across a long take.
  • Four-panel first/last-frame transitions, including looping a transformation.
  • Poster-to-motion: making a still key art image animate in the way the copy implies.

How this differs from the previous build

Three upgrades stack here, and the author is careful about which does what.

  • Distilled 1.1 replaces 1.0. With 1.0, even VBVR-equipped runs still broke during motion; 1.1 follows prompts better, breaks less and produces more sensible audio.
  • VBVR 240K replaces the 96000 build. The gain is not stronger motion in the abstract — it is that prompt elements previously skipped now render.
  • A crop group was added for four-grid input. Generate a 2x2 grid in an image model, crop it into four, and feed the graph without an upscale step.
  • The image-to-video graph is deliberately minimal at 14 nodes: only the inputs you actually adjust are exposed.

Running it

  1. Start with the image-to-video graph. Set prompt, negative prompt, size constraint (0.1-0.15), frames and frame rate.
  2. Compare seeds with and without the VBVR LoRA to see the compliance difference on your own prompt.
  3. For digital human, load image and audio, trim the audio length, and set frame rate and prompt.
  4. For the four-grid graph, generate a 2x2 grid elsewhere, connect the crop group, then set each quadrant's weight and appearance frame.
  5. Add a loop when you need a repeating transformation.
  6. Keep three-tier rendering on — the author attributes a large part of the hit rate to it.

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-240K-R32.safetensors at 1.0
LoRA, accelerate ltx-2.3-22b-distilled-lora-384.safetensors at 0.5
LoRA, motion track ltx-2.3-22b-ic-lora-motion-track-control-ref0.5.safetensors at 1.0
LoRA, transition ltx2.3-transition.safetensors at 1.0 (four-grid graph)

The image-to-video graph loads no text encoder and no sampler of its own in the parsed data — it is a 14-node entry shell around the shared chain, which is why the exposed-inputs-only design matters.

Honest limits

  • VBVR makes hidden prompt elements appear; it does not make motion more dramatic than the prompt asks, and the author says so explicitly.
  • With heavy multi-image conditioning the VBVR gain narrows, because four guides already constrain the result.
  • The author tested the newer HDR model and found image-to-video loses similarity and video-to-video misbehaves — text-to-video only, for now.
  • The four-grid workflow needs an external image model to build the grid; the graph only crops it.
  • Comparisons are visual with seeds locked; there is no scored metric.

Episode resources

RunningHub workflows

Top comments (0)