Descriptions:
Nerdy Rodent walks through an updated ComfyUI workflow for LTX Video 2.3 that integrates several newly supported capabilities: SAM3 (Segment Anything Model 3) for per-frame video segmentation, an inpainting LoRA using the IC LoRA guide node, a body positivity LoRA from Berg Stall with an Apache 2.0 license, frame interpolation models, and the APG (Adaptive Projected Guidance) node for steering generations away from undesired cartoon-style outputs.
The tutorial is built around the presenter’s “rodent method” — organizing workflows into color-coded modular groups to minimize spaghetti node layouts and make components easy to swap. Key configuration decisions are shown in detail: SAM3 segmentation prompts are capped at 32 tokens with numeric element counts to control mask generation; IC LoRA guide nodes replace standard add-guide nodes for inpainting LoRAs; denoise levels (demonstrated at 0.42 for moderate change) and scheduler choice (linear quadratic shown) control how far stage-two processing diverges from the source video. Switching between encoded video and empty latent as the stage-two input is demonstrated as a way to allow aspect ratio changes and more dramatic prompt-driven transformation.
Direct before/after comparisons show APG’s effect on cinematic quality versus its absence, and the inpainting workflow’s ability to replace specific masked regions (face, window) while leaving the rest of the scene intact. Completed workflows are available to channel supporters via Patreon.
📺 Source: Nerdy Rodent · Published May 02, 2026
🏷️ Format: Hands On Build
