Most I2V merges simply average weights or load large Lora adapters; LTX2.3 tries a different trade-off: layer-scaled merges across multiple checkpoints to preserve conditional responsiveness while keeping a compact checkpoint suitable for BF16/FP8 workflows. The practical result is a model that behaves more predictably under stronger, structured prompts — but it requires explicit direction for frame-to-frame motion and audio cues.
Key Capabilities
- Layer-scaled merged weights: merges are performed with per-layer scaling rather than a straight weight average, which helps retain conditional behavior from source checkpoints and reduces the “washed-out” effect common in naive merges. So what: you get better adherence to prompt instructions for motion and scene evolution.
- Multi-format checkpoint support (BF16, FP8): provides BF16 checkpoint and FP8 variants (including an fp8_mixed_learned option) so you can run compact, faster inference or fine-tune smaller-footprint versions. So what: enables deployment on setups that favour lower-precision runtimes while keeping acceptable fidelity.
- Prompt-first behaviour: the model has limited self-directed reasoning — it expects precise, prescriptive prompts for the initial frame, subsequent motions, and audio. So what: with careful prompt engineering you can control fine-grained temporal changes; without that the model will produce generic or stagnant outputs.
- Compatibility notes: relies on Sulphur-2-base and other merges; Kijai-split files exist for a 10Eros FP8 transformer structure and some assets must be placed under diffusion_models. So what: integration requires knowing which file-structure variant you need for your chosen runtime.
Who it's for and trade-offs
Great fit if you are experimenting with multimodal I2V pipelines and want a merged checkpoint that keeps prompt control at the center of generation. It’s useful for researchers and creators who can invest in prompt design and want BF16/FP8 options for constrained inference environments. Look elsewhere if you need plug-and-play, highly autonomous storyboarding — LTX2.3 deliberately shifts responsibility for motion and audio direction to the prompt. Also avoid applying large distilled Loras on top of this model: larger distilled Loras can degrade the model’s fine-tuned conditional behaviour; prefer the smaller cond_safe distilled Loras if augmentation is needed.
Where it fits
Positioned as a middle-ground merged I2V checkpoint: more controllable than a straight weight merge and lighter than heavyweight multi-stage stacks. Best used as the core generator in research prototypes or creative pipelines where prompt engineering and checkpoint precision choices (BF16/FP8) are part of the workflow.
