DEV Community

AI-KSK
AI-KSK

Posted on

LTX 2.3 Production Suite 2.0: Prompt Relay Plus Three-Tier Rendering, Five Graphs

#comfyui #ai #tutorial

The bottleneck in LTX 2.3 video was never the model — it was not having a repeatable production graph. Suite 2.0 is the answer: five graphs, Prompt Relay on all of them, three-tier rendering on all of them, and the final 390K VBVR reasoning model layered in. The author's own comparison shows the difference VBVR makes is compliance, not spectacle.

Evidence note: five 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

  • Replacing a pile of one-off LTX graphs with one consistent production set.
  • Multi-shot clips where each segment needs its own prompt and its own timing.
  • Digital-human talking footage where lip sync and identity both have to hold.
  • First/last-frame transitions, including a one-to-four variant for batch coverage.

How this differs from the previous build

This is the 1.0 to 2.0 jump, and the changes are structural.

  • Prompt Relay is on every graph. Total prompt, segment prompts and negative prompt are now standard inputs rather than an optional add-on.
  • Three-tier rendering is on every graph. Low-resolution pass, latent upscale, refine.
  • 390K VBVR is the new reasoning layer. The author tested the earlier 96K build; this one is the final release and the gain shows up as prompts being executed rather than ignored.
  • Every image-to-video graph converts to text-to-video with a bypass. Flip it to true and the image stops conditioning the run.
  • The negative prompt now routes through a NAG node so it has an effect at all.

Running it

  1. Pick the graph. Image-to-video and the audio digital-human graph are the two workhorses; the first/last-frame pair covers transitions.
  2. Set seed, image size (0.1-0.15 in the exposed control), frames and frame rate.
  3. Write three prompts: total, per-segment, negative. The author's own prompt app is a ComfyUI graph carrying a GPT-5 system prompt; feeding the prompt rules to any LLM works too.
  4. For the audio digital human, load the image and the audio, then set the trim length — which becomes the video length.
  5. For the four-image first/last-frame graph, set each image's weight and the frame it appears at, for example frame 30 or frame 210.
  6. Flip the bypass node to run the same graph as text-to-video.
  7. Run, then compare across seeds; the author's VBVR comparisons locked the seed to make the difference legible.

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-390K-R32.safetensors at 1.0 (96000 variant in the first/last-frame pair)
LoRA, crisp LTX2.3_Crisp_Enhance.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

The 96000 and 390K VBVR LoRAs are not interchangeable: the first/last-frame pair still ships the 96000 build, while the image-to-video and digital-human graphs use 390K.

Honest limits

  • The VBVR gain is compliance, not motion amplitude — the author is explicit that unseen prompts appear rather than dynamics improving.
  • With strong multi-image conditioning, VBVR's benefit shrinks: four inserted guides already constrain the result heavily.
  • Image-to-video with segment prompts can produce a talking-head result without an audio track, but that is not the same as an audio-driven digital human.
  • All comparisons are visual, with seed locked, and are not scored.
  • One graph is explicitly the no-audio variant; do not expect speech-driven output from it.

Episode resources

RunningHub workflows

Top comments (0)