LTX-2.3 Surface Realism LoRA — video-branch, rank 32
Kills the plastic. LTX-2.3 (and the image models feeding it first frames) defaults to clean, smooth, showroom-new surfaces: pristine asphalt, unmarked walls, grounds that read as rendered geometry. This LoRA is an always-on materials prior: cracked wet asphalt stays cracked and wet, concrete gets pores and rust streaks, floors scuff, weeds mat, wood grains. It pushes the render toward surfaces that have existed in weather.
No trigger word — by design
The training captions name only material, condition, framing, and light — they deliberately contain zero realism vocabulary ("photoreal", "detailed", "8k" never appear). The realism therefore trains as unconditional LoRA behavior instead of a phrase you have to remember to include. Load it at 1.0 and write your prompts normally; concrete material nouns in your prompt ("scuffed linoleum", "rust-streaked concrete") still help composition, as they always did.
Voice/audio safe by construction
1,152 LoRA tensors, all in the video branches — attn1 (480), attn2 (480),
ff (192). Zero audio-touching tensors: dialogue, voices, and accents
render byte-identical with or without this LoRA. Verified alongside an
audio-branch accent LoRA in the same stack with no interference.
Usage
- Strength 1.0 in any LTX-2.3 LoRA loader.
- Compatible with the LTX-2.3 22B family: stock distilled 1.1 and the JoyAI-Echo surgical merges (bf16 / fp8 / INT8 / GGUF via fuse-at-load).
- If you render dialogue: keep your workflow at 24 fps — not for this LoRA's sake, but because off-24 fps drifts every LTX voice toward British/Australian regardless of prompt (verified 2026-07-30).
Training
- ai-toolkit on an RTX 3090, rank 32 / alpha 32, 3,000 steps, lr 1e-4,
qfloat8. Base: stock
ltx-2.3-22b-distilled-1.1. Loss 1.01 -> 0.52. - Data: 1,076 hand-curated surface plates (from a 1,900+ candidate pool) generated with Z-Image Turbo specifically for this training — grounds, walls, floors, and weathered materials under varied light. Every image is self-generated; no third-party photography.
- Module scope excludes every audio branch (
ignore_if_contains: ["audio"]), which is what guarantees the audio lane stays untouched.
License
LoRA weights follow the LTX-2 Community License of the base model. All training imagery self-generated.