LTX 2.3 Finetuned I2V

LTX 2.3 Finetuned Advanced I2V with identity & voice conditioning. Upload an image, write a prompt, generate.

All generated content must follow Huggingface policy.

reference image

prompt

preset (sets all lora, targeting, and sigma defaults)

enable prompt relay (timeline-based prompts)

negative prompt

duration (seconds, up to ~1000 frames)

1 41

last frame (optional)

keyframe strength (applies to last frame + all middle frames)

0 1

add intermediate keyframes

enable loras

i2v checkpoint influence

0 1

target megapixels

snap to multiple

32 64 (recommended)

custom resolution (overrides megapixels)

face mode

anchor only auto face manual bbox

manual bbox

likeness guide

0 1

likeness anchor

0 1

latent anchor

0 0.5

first frame strength

0 1

enable K/V identity conditioning (experimental)

K/V strength (0 = off, 1 = funpack default, >1 = stronger identity)

0 2

enable scene chaining (bypasses pass 2 for v1)

scene chain prompt

max scene chunks (free-tier test: keep at 2)

2 4

scene overlap frames (8 = safer first test)

0 24

carry previous-scene midpoint as guide

mid-scene guide strength

0.25 0.5

audio reference (voice ID transfer)

identity guidance scale (lower if audio problems)

0 10

isolate voice from background (stem separation, slower)

normalize reference audio (caps to 10s, boosts quiet clips)

audio reference (~4s clip recommended)

similarity threshold (lower = corrects drift earlier, catches face changes on angles; too low can distort anatomy)

0 1

anchor cache step (0 = auto-align to frame count; controls when identity locks)

0 12

cache warmup (affects sustained identity over duration; 50/400/1000 behave differently)

10 2000

energy threshold (latent anchor sensitivity)

0 1

refine sigmas

single pass (skip refiner)

hide sensitive content

enhance prompt budget (seconds)

20 540

generation budget (seconds, 0 = automatic)

0 540

seed

randomize seed

export all current settings to a json file, or import a previously saved profile. images, prompts, and audio files are not included.

profile name (optional)

import profile (.json)

output

status

used seed

Built with Gradio logo