ComfyUI-LTXV-TimeGated-LoRA v1.1

v1.1 envelope build: adds local / hold_strength envelope modes, q_curve ramps, a reusable data output and a CPU-rendered LTXV Envelope Curve Preview (CPU) node. transition_frames is only used by manual_frames and transition_frames legacy mode.

Node: LTXV Time-Gated LoRA (LTX 2.3)

Temporally apply visual LTX 2.3 LoRAs to selected regions of a single continuous video sampling run. Use it for effects such as a style appearing only in the middle third, a transformation beginning in the final quarter, or multiple independent scheduled LoRAs chained in one clip.

Scope

• Visual LTX 2.3 LoRA layers only; temporal audio gating is intentionally not supported in v1.0.
• effect_region mode for fast placement by halves, thirds or quarters.
• manual_frames mode for exact signed multi-segment schedules.
• Timing is derived from the required final video-only video_latent.
• Multiple instances can be chained additively.
• Low-rank temporal gating includes a safe mode and an experimental lower-VRAM in-place mode.

The validated diagnostic build is kept separately as v0.7.1; its diagnostic probe is intentionally not exposed in the production release candidate.

Placement

Place this node, or a chain of these nodes, last in the MODEL chain directly before the sampler or guider.

For a PromptRelay workflow:

... -> PromptRelayEncode -> LTX2SamplingPreviewOverride
    -> LTXV Time-Gated LoRA -> optional additional LTXV Time-Gated LoRA -> SamplerCustom

Do not put this runtime-gated node in the normal static LoRA group near the model loader. Static LoRAs and distilled LoRAs may remain in their original upstream locations.

video_latent timing and stacking

Connect the final video-only latent after image/FLF conditioning and before any audio/video concat node. The node now passes this latent through unchanged so stacked nodes can be wired cleanly.

Timing path:
LTXVImgToVideoInplaceKJ video_latent
  -> LTXV Time-Gated LoRA A video_latent
  -> LTXV Time-Gated LoRA B video_latent

MODEL path:
LTX2SamplingPreviewOverride MODEL
  -> LTXV Time-Gated LoRA A MODEL
  -> LTXV Time-Gated LoRA B MODEL
  -> SamplerCustom MODEL

Do not connect an audio+video combined latent to video_latent. The node derives the real LTX timeline from the video temporal dimension and reports the resolved timeline in schedule_info. LTX 2.3 temporal compression is fixed internally to 8.

Inputs and outputs

Inputs

• model: completed LTX 2.3 MODEL path to patch.
• video_latent: required final video-only timing reference.
• lora_name: compatible visual LTX 2.3 LoRA.
• strength_model: global LoRA multiplier. Effective strength is strength_model × schedule value.
• schedule_mode: effect_region or manual_frames.
• effect_region, strength_before, strength_during, strength_after: used in effect_region mode.
• segment_lengths, segment_strengths: used in manual_frames mode.
• transition_frames: crossfade width at effect boundaries.
• memory_mode: optimized_safe by default; optimized_low_vram_inplace is experimental.

The numeric strength widgets use practical 0.05 steps for mouse adjustment. segment_strengths remains a freely editable string for advanced schedules.

Outputs

• model: patched MODEL path.
• video_latent: unchanged timing latent passthrough for stacking.
• schedule_info: compact resolved-schedule and compatibility report.

effect_region mode

Example: apply a style only in the middle third.

schedule_mode:     effect_region
effect_region:     middle third
strength_model:    1.50
strength_before:   0
strength_during:   1
strength_after:    0
transition_frames: 16

Available regions include quarters, thirds, first half, center half and last half.

manual_frames mode

Use exact rendered-frame segment lengths and multipliers. Signed and decimal values are valid.

schedule_mode:      manual_frames
segment_lengths:    120,193,192
segment_strengths:  0,-1,1.5

The segment lengths must sum to the rendered frame length resolved from video_latent.

PromptRelay compatibility

For simple synchronization, leave PromptRelay segment_lengths empty. PromptRelay will evenly distribute its |-separated local prompts, which matches the region philosophy of this node when the prompt segmentation matches the selected granularity:

• halves: use 2 local prompt segments
• thirds: use 3 local prompt segments
• quarters: use 4 local prompt segments

Examples:

middle third   -> prompt A | prompt B | prompt C
second quarter -> prompt A | prompt B | prompt C | prompt D
last half      -> prompt A | prompt B

For stacked nodes using mixed region types, exact PromptRelay synchronization may require manual segment planning.

Trigger-dependent LoRAs

When a Style LoRA requires a trigger word or activation phrase, include it inside every active PromptRelay local segment in which that Time-Gated LoRA is enabled. Do not rely only on the global prompt for a scheduled style reveal. Some broad style LoRAs, such as Claymation-style LoRAs, may respond to a global style description, but may still require higher strength.

Best practices

Strength ranges

LTX 2.3 LoRA strengths are not inherently limited to 0–1. Depending on the LoRA, useful values may exceed 1.0. Slider or transformation LoRAs may use meaningful negative and positive values, and their two directions do not have to be equally strong.

If an effect is weak, do not assume the LoRA is broken at 1.0; test appropriate higher strengths while watching for artifacts.

I2V image anchor can suppress transformations

In I2V workflows, a strong image anchor can substantially suppress visible LoRA transformations. With nodes such as LTXVImgToVideoInplaceKJ, a high value such as strength_1 = 1.0 may pull the appearance back toward the initial image late in denoising, even when early sampler previews already show the LoRA effect.

For strong style or transformation effects, try a lower anchor first, for example 0.2–0.5, before pushing LoRA strength further.

Transition frames

• 8–16 frames: good starting point for ordinary style/effect reveals.
• 0 frames: intentionally hard switches or diagnosis.
• 16–32 frames: stronger transformations that need a gentler blend.

At 24 fps, 8–16 frames correspond to approximately 0.33–0.67 seconds.

Frame rate and valid LTX length

For convenient whole-second durations, prefer 24 fps: whole seconds map cleanly to the LTX 8n+1 frame convention when the workflow adds the final +1 frame.

Memory and compatibility

• Start with memory_mode = optimized_safe.
• Try optimized_low_vram_inplace only after an OOM in safe mode, especially for long clips or stacked nodes, and compare output stability.
• Each additional scheduled LoRA adds compute and VRAM pressure.
• SageAttention is usable, but keep allow_compile = false when using this node.
• torch.compile / TorchCompileModelAdvanced is not currently supported.
• Temporal audio LoRA gating is not supported in v1.0.
• A real LoRA may influence later video semantics after its active region even when the gate itself is temporally correct.

Validated demo scenarios

Age Transformation Reveal

A single continuous I2V run showing an adult appearance, a younger reveal and an elderly reveal using a time-gated age/slider LoRA with three PromptRelay segments. A reduced I2V image anchor was critical for visible transformation.

Claymation Style Reveal

A realistic scene reveals a Claymation style only in the middle third, then returns to realism. The sign's blank reverse side acts as a natural occluder/reveal transition. This workflow is the recommended release-candidate smoke test.

Envelope diagnostics / curve preview

This development build contains three nodes:

• LTXV Time-Gated LoRA (LTX 2.3) — productive model patching node.
• LTXV Temporal Envelope Inspector (v1.1rc4-dev) — non-patching audit node. It resolves the frame profile, latent profile, regions, samples and warnings.
• LTXV Envelope Curve Preview (CPU) — renders a small timeline/strength analyzer image from the Inspector's combined data output.

Recommended diagnostic wiring:

video_latent
  → LTXV Temporal Envelope Inspector
      data → LTXV Envelope Curve Preview (CPU) → Preview Image
      report_md → Show Text / PreviewAny

The data output is a single custom payload (LTXV_ENVELOPE_DATA) so the preview node does not need separate cables for timeline, profile, regions and warnings. Rendering is CPU-side via PIL/numpy and returns a normal ComfyUI IMAGE.

Installation

Extract the folder into ComfyUI custom_nodes, restart ComfyUI, and add a fresh instance of LTXV Time-Gated LoRA (LTX 2.3) to the workflow. Because v1.1rc4-dev adds envelope diagnostics and preview outputs, replace older node instances instead of relying on stored widget/output indexes.

v1.1 visual envelope preview

v1.1 adds a CPU-rendered curve preview that shows the actual local LoRA envelope over the full video timeline.

The preview is useful for checking selected regions, comparing stacked Time-Gated LoRA nodes, debugging flat versus q_curve, and verifying data / data_2 overlays before sampling.

Example: late q-curve gate

A delayed q-curve where the LoRA stays inactive for most of the clip and ramps strongly into the last third.

Example: three-step flat envelope

A flat local envelope using one node as a 0 → 0.5 → 1 step controller.

Example: stacked Time-Gated LoRA nodes

Two Time-Gated LoRA nodes can be stacked in the MODEL path. Their data outputs can be connected to the Curve Preview / Analyze node as data and data_2 to compare both envelopes in one plot.

Comparison videos

The repository also includes two three-up comparison videos:

• Cucumber comparison — No LoRA / Standard LoRA / Time-gated LoRA
• Rain comparison — No LoRA / Standard LoRA / Time-gated LoRA

The cucumber comparison is a deliberately playful A/B showcase. The rain comparison is useful as a production/stress-test example: it demonstrates a rain/particle LoRA and also shows why stochastic, high-frequency detail LoRAs can be more sensitive to temporal gating than semantic or stylistic transformations.

Production workflow example

The repository includes a production/testing workflow:

workflows/production/PromptRelay_LTXV_TimeGated-LoRA_Rain_StepGate_v1.1.json

This is not a minimal install test. Bypassed nodes are intentionally kept for A/B testing, alternate gate setups, preview inspection, last-frame export, upscaling, and production debugging. Input images and third-party LoRAs are not included.

Note on stochastic / particle LoRAs

Time-gated curves are especially useful for semantic or stylistic transformations such as realism ↔ claymation, age sliders, mood/look changes, or character/style intensity. Fine stochastic detail LoRAs such as rain, snow, dust, film grain, water ripples, or particle effects may flicker under gradual curves because the model keeps renegotiating high-frequency details over time. For these LoRAs, test lower lora_strength, flat instead of a long q_curve, a non-zero strength_before, or constant full-clip application.