LTX-2.3 Image Upscaler - Apple Silicon / Mac (GGUF port)

A Mac / Apple Silicon port of JXL's excellent LTX-2.3 Image Upscaler (https://civitai.com/models/2741107). The original workflow ships as fp8 DiT + fp4 Gemma, which hit the MPS float8 wall and won't run on Apple Silicon. This version swaps just those three loaders for GGUF / split-file equivalents, so the full two-stage pipeline runs end-to-end on Metal: no NVIDIA GPU, nothing leaves your machine. It's the same upscaler under the hood. It renders a blur-to-sharp fade micro-video with the upscalify LoRA and keeps the sharpest frame, rebuilding real detail (pores, hair, texture) while holding identity. Base gen, latent upsample x2, refine, keep last frame, ~1920px. What changed vs the original (that's all): CheckpointLoaderSimple (dev-fp8) becomes UnetLoaderGGUF (LTX-2.3-dev-Q4_K_M.gguf); its VAE output becomes VAELoader (LTX23_video_vae_bf16); LTXAVTextEncoderLoader (gemma fp4) becomes DualCLIPLoaderGGUF (gemma Q4 + text_projection, type ltxv); LTXVAudioVAELoader (dev-fp8) stays the same node, repointed to the standalone LTX23_audio_vae_bf16. Everything else (the NAG, chunk-FF, AV, ManualSigmas and sizing nodes) is kept exactly as JXL built it, available on Mac via ComfyUI core + KJNodes + ComfyUI_essentials. Heads up, it's slow on Metal: about 25-35 min per image (two-stage 22B dev; the x2 refine is the cost). Great for local and private one-offs; use an NVIDIA box for turnaround. Tested on an M5 Pro (48 GB), ComfyUI 0.26, --lowvram (peak ~17 GB). Notably the LTX-2.3 two-stage ManualSigmas + SamplerCustomAdvanced pipeline has been reported to NaN at VAE decode on MPS. Using the dev GGUF, this port runs it clean. Full setup, the weight list, and a one-command runner ( run .py) are on GitHub: https://github.com/Bambushu/ltx-image-upscaler-mac All credit to JXL for the original workflow and the upscalify LoRA. This is only a Mac/GGUF port of that work. The weights and LoRA are not included here; grab them from the sources in the README and follow their licenses.