LTX-2 T2AV Benchmark

LTX-2 Text-to-Audio-Video Benchmark

Warm-cache generation times (second run, models already in VRAM). RTX PRO 6000 Blackwell 96GB.

Prompt

A close-up shot of a young woman with auburn hair sitting at a wooden desk in a warmly lit study. She looks directly at the camera and says in a clear, friendly voice: "Hello everyone, welcome to vMonad. Today I want to show you something really exciting about our new media generation platform." Soft window light from the left, shallow depth of field, cinematic 50mm lens, natural skin tones, warm color grading.

	Native Distilled	Kijai Simple (Distilled)	Kijai Upscale (Distilled)
T2AV	Best 14.8s 1280x704 / 802 KB / with upscaler	Ok 8.5s 640x352 / 781 KB / no upscaler	Good 29.5s 1280x704 / 2.6 MB / 2-stage upscaler
TA2AV	Best 28.7s 1280x704 / 1.1 MB / with upscaler Kino voice conditioned	Worst 9.1s 640x352 / 894 KB / no upscaler Kino voice conditioned	Ok 31.8s 1280x704 / 2.6 MB / 2-stage upscaler Kino voice conditioned
IT2AV	Good 23.9s 1280x704 / 1.7 MB / 9.0s / with upscaler Audio correct but not pleasant	Best 39.6s 1280x704 / 2.5 MB / 9.0s / no upscaler Best audio quality	Worst 29.1s 1280x704 / 3.4 MB / 9.0s / 2-stage upscaler Oversaturated, overexposed
ITA2AV	Best 24.1s 1280x704 / 1.8 MB / with upscaler Kino voice conditioned	Worst 9.2s 640x352 / 1.1 MB / no upscaler Kino voice conditioned	Worst 30.0s 1280x704 / 3.4 MB / 2-stage upscaler Better than simple but still poor
V2V Ext	RuneXX Workflow (Adopted) Flawless video extension using RuneXX workflow with LoRA at -0.3 + LTX2_NAG + LTX2AudioLatentNormalizingSampling. Chainable for infinite extension (tested 3 iterations: 5s -> 14.7s -> 24.4s -> 34s). Audio: LTX-2 native (gibberish) -- requires TTS mux for real speech. Saved to: comfynative/workflows/video_ltx2_v2v_extend_runexx.json
FFLF	Best (Candidate) ~42s 1280x768 / 7.6s / STGGuiderAdvanced + LTXVAddGuide. Lip sync works via cross-modal STG guidance (layer 29). Known issue: garbled text overlay artifacts (revisit prompt engineering). Saved to: comfynative/workflows/video_ltx2_fflf_stg_lipsync.json	Worst 123.4s 768x1280 (portrait!) / 8.0s / 2-stage upscale pipeline. Dimension swap bug: WIDTH/HEIGHT labels reversed in workflow nodes. Identity drift: upscale pass only conditions on first frame, loses last frame identity. Audio: garbage (CFGGuider + EmptyLatentAudio, no cross-modal STG coupling). Generates completely different person from input frames.
FMLF	Best 57.4s 1280x768 / 7.6s / Single-stage, 25 nodes. 3 chained LTXVAddGuide (first=1.0, mid=0.5, last=0.8) + CFGGuider cfg=1. Correct orientation, smooth transitions between keyframes. Saved to: comfynative/workflows/video_ltx2_fmlf_native.json	Worst 111.6s 768x1280 (portrait!) / 8.0s / 2-stage, 62 nodes. Same dimension swap bug as FFLF. Uses LTXVAddGuideMulti (kijai-specific).
V2A Foley	Best 24.7s 854x478 / 5.0s / Single-stage native pipeline. Better prompt adherence than Kijai. Prompt: "Starwars army of soldiers walking" Saved to: comfynative/workflows/video_ltx2_v2a_foley_native.json	Ok 60.9s 854x478 / 5.0s / 2-stage + LoopingSampler + AudioNorm. Weaker prompt adherence. Same RuneXX V2A Foley pipeline. Slower due to LTXVLoopingSampler refinement pass.
Talking Avatar	Best 54.1s 1280x768 / 7.6s / Single-stage, full resolution. LTXVImgToVideo + CFGGuider cfg=1 + Camera Control LoRA. Qwen3 TTS voice clone audio muxed post-generation. No transitions, strong identity lock, clean output. 23 nodes (vs 58 kijai). Only 1 third-party dep: TTS. Saved to: comfynative/workflows/video_ltx2_talking_avatar_native.json	Ok ~40s 640x384 / 6.0s / Stage 1 only (upscale bypassed). Lip sync works but low resolution, not sharp. Upscale pass causes dithering + destroys audio. Workflow had 2 broken connections in API JSON.	Worst 85.2s 1280x768 / 6.0s / 2-stage with upscale. Dithering artifacts from upscale pass. Model audio destroyed by stage 2 sampling.

Native Distilled

Kijai Simple (Distilled)

Kijai Upscale (Distilled)

T2AV

Best

14.8s

1280x704 / 802 KB / with upscaler

8.5s

640x352 / 781 KB / no upscaler

Good

29.5s

1280x704 / 2.6 MB / 2-stage upscaler

TA2AV

Best

28.7s

1280x704 / 1.1 MB / with upscaler
Kino voice conditioned

Worst

9.1s

640x352 / 894 KB / no upscaler
Kino voice conditioned

31.8s

1280x704 / 2.6 MB / 2-stage upscaler
Kino voice conditioned

IT2AV

Good

23.9s

1280x704 / 1.7 MB / 9.0s / with upscaler
Audio correct but not pleasant

Best

39.6s

1280x704 / 2.5 MB / 9.0s / no upscaler
Best audio quality

Worst

29.1s

1280x704 / 3.4 MB / 9.0s / 2-stage upscaler
Oversaturated, overexposed

ITA2AV

Best

24.1s

1280x704 / 1.8 MB / with upscaler
Kino voice conditioned

Worst

9.2s

640x352 / 1.1 MB / no upscaler
Kino voice conditioned

Worst

30.0s

1280x704 / 3.4 MB / 2-stage upscaler
Better than simple but still poor

V2V Ext

RuneXX Workflow (Adopted)

Flawless video extension using RuneXX workflow with LoRA at -0.3 + LTX2_NAG + LTX2AudioLatentNormalizingSampling.
Chainable for infinite extension (tested 3 iterations: 5s -> 14.7s -> 24.4s -> 34s).
Audio: LTX-2 native (gibberish) -- requires TTS mux for real speech.
Saved to: comfynative/workflows/video_ltx2_v2v_extend_runexx.json

FFLF

Best (Candidate)

~42s

1280x768 / 7.6s / STGGuiderAdvanced + LTXVAddGuide.
Lip sync works via cross-modal STG guidance (layer 29).
Known issue: garbled text overlay artifacts (revisit prompt engineering).
Saved to: comfynative/workflows/video_ltx2_fflf_stg_lipsync.json

Worst

123.4s

768x1280 (portrait!) / 8.0s / 2-stage upscale pipeline.
Dimension swap bug: WIDTH/HEIGHT labels reversed in workflow nodes.
Identity drift: upscale pass only conditions on first frame, loses last frame identity.
Audio: garbage (CFGGuider + EmptyLatentAudio, no cross-modal STG coupling).
Generates completely different person from input frames.

FMLF

Best

57.4s

1280x768 / 7.6s / Single-stage, 25 nodes.
3 chained LTXVAddGuide (first=1.0, mid=0.5, last=0.8) + CFGGuider cfg=1.
Correct orientation, smooth transitions between keyframes.
Saved to: comfynative/workflows/video_ltx2_fmlf_native.json

Worst

111.6s

768x1280 (portrait!) / 8.0s / 2-stage, 62 nodes.
Same dimension swap bug as FFLF.
Uses LTXVAddGuideMulti (kijai-specific).

V2A Foley

Best

24.7s

854x478 / 5.0s / Single-stage native pipeline.
Better prompt adherence than Kijai.
Prompt: "Starwars army of soldiers walking"
Saved to: comfynative/workflows/video_ltx2_v2a_foley_native.json

60.9s

854x478 / 5.0s / 2-stage + LoopingSampler + AudioNorm.
Weaker prompt adherence. Same RuneXX V2A Foley pipeline.
Slower due to LTXVLoopingSampler refinement pass.

Talking Avatar

Best

54.1s

1280x768 / 7.6s / Single-stage, full resolution.
LTXVImgToVideo + CFGGuider cfg=1 + Camera Control LoRA.
Qwen3 TTS voice clone audio muxed post-generation.
No transitions, strong identity lock, clean output.
23 nodes (vs 58 kijai). Only 1 third-party dep: TTS.
Saved to: comfynative/workflows/video_ltx2_talking_avatar_native.json

~40s

640x384 / 6.0s / Stage 1 only (upscale bypassed).
Lip sync works but low resolution, not sharp.
Upscale pass causes dithering + destroys audio.
Workflow had 2 broken connections in API JSON.

Worst

85.2s

1280x768 / 6.0s / 2-stage with upscale.
Dithering artifacts from upscale pass.
Model audio destroyed by stage 2 sampling.

Source Image (I2AV row)

RuneXX Community Workflows

View RuneXX LTX-2 Workflows Benchmark (16/22 succeeded)

All 22 workflows from RuneXX/LTX-2-Workflows on HuggingFace, auto-converted and submitted with same test inputs.

All workflows use LTX-2 19B distilled FP8 model at 25fps. Audio generated natively by LTX-2 from prompt text (not TTS). Timing is wall-clock for the second run with models warm in VRAM. First run includes model loading. T2AV = text only. IT2AV = image + text input. ITA2AV = image + text + audio input (Kino voice conditioned).
ITA2AV uses MelBandRoFormer vocal extraction + SetLatentNoiseMask to preserve voice through sampling.