Load Latent EXR

This is the read side of the pack's latent round-trip. SaveLatentEXR writes a latent tensor out as a 4-channel 32-bit float EXR; this node loads one back and hands you a LATENT that's ready to feed a sampler, an img2img pass, or a refine loop. The whole point is that nothing gets lost between save and load - no PNG quantization, no VAE encode/decode drift. If you saved it, you get it back exactly.

Reach for it when you want to interrupt a big batch, pick up a specific generation state days later, or move latents between projects without rerunning the sampler that made them. The VAE-decode round trip is lossy and costs you quality on every pass; skipping it entirely is the whole appeal.

How it works

EXR is the only mainstream format this works with, because latents are float values that go negative and carry way more precision than 8-bit pixels. The loader reads the file with OpenCV's unchanged float mode, reorders the RGBA channels back to the order the sampler expects (the exact inverse of what SaveLatentEXR does on the way out), and reshapes into the (batch, 4, H, W) latent layout ComfyUI wants. Then it wraps it in the standard {"samples": ...} dict and returns it as samples.

Like LoadEXR, it accepts a single file or a folder of them, with the same VHS-style batch controls: image_load_cap (0 = everything), skip_first_images, and select_every_nth. Outputs are samples (LATENT) and batch_size (INT).

Wire samples into the latent input of a KSampler and you're sampling from a restored state - which is a genuinely different workflow from loading a PNG and encoding it, because you're skipping a lossy encode too.

The 4-channel caveat, repeated

Everything about this node assumes the SD 1.5 / SDXL latent layout: four channels, (B, 4, H, W). That matches what SaveLatentEXR writes. It does not match Flux (16 channels) or the newer video models. If you save a Flux latent with this pack, the load will only bring back four channels and the rest is gone. For SD-family models it's exact; for anything else, check the channel count first.

Install

Part of spacepxl's ComfyUI-HQ-Image-Save: ComfyUI Manager → search ComfyUI-HQ-Image-Save → install → restart, or clone https://github.com/spacepxl/ComfyUI-HQ-Image-Save into custom_nodes/. Dependencies are imageio and opencv-python only, no models.

Pair it with the obvious

SaveLatentEXR on the way out, LoadLatentEXR on the way in, same filename conventions, same channel order. If you hit the "OpenEXR codec is disabled" error on load, that's your OpenCV build missing the EXR codec - the standard opencv-python wheel includes it, and the pack enables it at import.