LTX-2.5 ships Native Resolution ComfyUI workflow for single-GPU 4K/8K video
linoy_tsaban · x · 2026-09-29
LTX officially released the Native Resolution two-stage video-to-video ComfyUI workflow for LTX-2.5, with Refine and Restore IC-LoRA weights live on Hugging Face.
- Stage one resizes the source video to an initial canvas to establish composition; stage two enlarges the latent, re-attaches the source as guide, and refines with Tiled Fusion.
- Tiled Fusion handles canvases larger than the IC-LoRA's trained window via overlapping spatial crops sharing one latent canvas and noise field throughout sampling — not independent tiles stitched together — reducing inter-region discontinuities.
- Designed for a single-GPU path with Full HD/4K/8K options, though VRAM and runtime depend on tile size, temporal window and decode settings; benchmark before committing a full clip.
- LTXVTiledSampler cannot substitute for LTXVTiledFusionSampler.
More drops are promised tomorrow.
More from Multimodal
- Seedance 2.5 'moving suitcase' video goes viral, prompt included — umesh_ai · 2026-09-30
- Kling 4 Flash vs Seedance 1.5: head-to-head video generation comparison — Smart-Proof-1861 · 2026-09-30
- Flux2Klein Enhancer ComfyUI Node Boosts Krea 2 Prompt Adherence for Free — No-Bad-4838 · 2026-09-30
- GEAR Uses Geometry as Address to Route Attention, Enabling Minute-Long Camera-Controlled Video Generation — Zesong Yang · 2026-09-30
- Caltech's DynaTokens teaches dynamics to camera-controlled video models at test time, NeurIPS accepted — georgiagkioxari · 2026-09-30
- AI-generated 'day in the life in Ancient Greece' video goes viral — venturetwins · 2026-09-30