Video Enhance - Prepare Video and Anchors

Video enhancement in this pack isn't "upscale the pixels and hope." It's a pipeline: slice your source into guide anchors, enhance those anchors with Z-Image, let LTX use the enhanced anchors to rebuild the whole video, then restore the exact original resolution. Every one of those steps needs a clean, consistent starting state - and that's what VRGDG Video Enhance - Prepare Video and Anchors produces. It's the front door of the whole Video Enhance node set.

You feed it the raw frames from a VHS Load Video node, and it does the prep work in one pass: creates an independent full-frame Video Enhance job, builds a temporary LTX video at working resolution, and picks regularly spaced frames to become Z-Image anchors - while keeping the original frames untouched and stashed in the context so the final restore step can enforce the exact source resolution and frame count.

The inputs

• video_frames - the full IMAGE batch from VHS Load Video. The tooltip is emphatic: these original frames remain unchanged and are what guarantees the final output matches the source exactly.
• video_info (optional) - the VHS_VIDEOINFO output from the loader, so the temporary LTX video uses the exact source FPS instead of a guess. Connect it; the fallback_fps (default 24) only kicks in when you don't.

The rest are the settings that decide the shape of the job:

• anchor_interval (default 16 frames) - how often an anchor frame is pulled. First and final frames are always included. Smaller spacing = more LTX reference, more Z-Image runs. 16 is the recommended middle ground.
• anchor_width / anchor_height (defaults 768×432) - the resolution Z-Image enhances at. These don't change the final video; they're the guide-frame working size. Match the source aspect ratio.
• ltx_width / ltx_height (defaults 960×544) - the temporary LTX working resolution. Higher = more spatial detail but noticeably more VRAM and runtime. Final output is restored to the source regardless.
• dimension_multiple (default 32) - rounds all dimensions to a model-friendly multiple. LTX is strict: width and height must be divisible by 32, or it fails. The resolved values come out as outputs so loaders can match them exactly.
• fit_mode - Fit with letterbox (preserve all, black bars if ratios differ), Crop to fill (trim edges), or Stretch (usually not recommended - it changes aspect ratio).
• resize_method - Bicubic recommended; Area helps on big downscales.

The outputs

ltx_working_frames and anchor_images (the picked guides), plus anchor_count, anchor_indices, ltx_video_path, ltx_width, ltx_height, and fps - and the video_enhance_context that carries the untouched source frames through the rest of the pipeline. The outputs you'll actually wire downstream are anchor_count (into the VHS Meta Batch setup) and the context.

Where it fits and the gotchas

Order of operations: Prepare → Load Anchors (Meta Batch) → Z-Image enhance → Store Enhanced Anchors → Collect LTX Inputs → LTX → Restore Original Resolution.

The classic mistakes are all resolution math. Don't set ltx_width/ltx_height to your final deliverable - you're paying for it in VRAM and the output gets restored anyway. Keep the working LTX size modest. And don't fight dimension_multiple; off-multiple dimensions are exactly the kind of thing that makes LTX silently fail. Start with the defaults on a short clip before committing a long video - the author's own guidance, and it applies double here because this node sets up an expensive chain.

Install

Standard pack install (Manager → search vrgamedev, restart, hard-refresh). This stage assumes the Video Helper Suite (VHS) for loading frames and video info, and a Z-Image checkpoint for the anchor-enhancement stage that follows. The pack's Video Enhance workflows come bundled under Workflows/ and show the full wiring.