AIAny
AI Video2026
Icon for item

LTX-Video 2.3 22B — IC-LoRA: CrossView Prompt v0.9

Generates a new camera viewpoint from a reference video: an IC‑LoRA adapter for LTX‑Video 2.3 that re‑renders the same scene from a requested discrete camera angle while preserving subject and content. Trained on synthetic multi‑view data, proof‑of‑concept with limited viewpoint range and best for small, chained angle shifts.

Introduction

Controlling camera viewpoint after capture is a major practical limit in video generation workflows. This IC‑LoRA acts like a virtual second camera for LTX‑Video 2.3: feed a reference clip and a short, fixed camera‑angle prompt and it attempts to re‑render the scene from the new viewpoint while keeping the original subject and motion intact. The design trades free‑form prompts for a discrete, learned camera vocabulary to improve stability and repeatability in viewpoint changes.

Key Capabilities
  • Discrete camera‑delta conditioning: accepts prompts of the form crossview. new camera angle: {azimuth}, {elevation}, {distance} from a fixed 63‑phrase vocabulary, so outputs are predictable and reproducible for each camera shift.
  • Reference‑conditioned video‑to‑video re‑rendering: preserves scene content and subject identity while moving the apparent camera position, so you can generate nearby viewpoints without rebuilding the scene from scratch.
  • Proof‑of‑concept training and practical tips: trained on synthetic multi‑view pairs (294 pairs) and tested in ComfyUI; works most reliably for small, single‑step angle changes and can be chained for larger shifts.
  • Attention‑only LoRA applied to LTX‑Video 2.3 (22B): designed as an adapter rather than a full model replacement, so it requires the LTX‑Video base weights and may need increased LoRA strength or non‑distilled passes for best effect.
Who it's for and trade-offs

Great fit if you need predictable, repeatable small viewpoint edits on short clips (e.g., shots for VFX, product demos, or interactive camera correction) and you can run LTX‑Video 2.3 in a ComfyUI workflow. The adapter is lightweight to integrate but expects the base 22B model and specific prompt vocabulary. Look elsewhere if you need extreme viewpoint changes (views from behind), free‑form natural language camera descriptions, high frame‑rate production pipelines, or fully real‑world training coverage—the model was trained on synthetic hemispherical camera sampling (frontal sector, ~±60°) and is explicitly a v0.9 proof‑of‑concept with known generalization limits.

Information

More Items

Hugging Face
AI Model2026

A dynamically quantized GGUF build of Ornith-1.5-35B optimized for agentic code-fixing and multi-turn conversations: targets 4-bit/≈22GB deployments, includes a vision projector, a custom importance matrix and a concise chat template.

Hugging Face
AI Model2026

Open-weights LLM fine-tuned for phone-based voice agents that prioritizes low latency and reliable tool/function calling. Based on NVIDIA Nemotron 3 Nano (30B total, 3.5B active), supports very long contexts (262,144 tokens) and recommends temperature=0 with thinking disabled for deployment.

Hugging Face
AI Model2026

A large open-weights MoE language model for complex coding, long-horizon agentic workflows, and cyber/security evaluations; post-trained from the GLM-5 family with substantial gains over GLM-5.2. Provides FP8/BF16 checkpoints and native support for very long contexts (up to 1M tokens).