MiniMax H3 Load AV Latent

The save half of the equation is useless without a load half, and MiniMaxH3LoadLatent is it. It reads back a latent from disk - either one written by this pack's Save Latent, or a Contex Loop chain checkpoint - and hands it to the Partial Denoise Mask as the thing to pin and rewrite against. No decode, no VAE round trip, no loss: the sampler gets the exact latent the clip was generated from, which is the entire point of the exercise.

What it reads

The dropdown lists every .safetensors file it finds in ComfyUI's output directory under paths containing h3_latent or /checkpoints. That second pattern is deliberate: the Contex Loop chain already stores per-clip latents in its checkpoints, so if you generated through a chain you can open a finished clip straight from there without ever having added a Save Latent node.

Two formats, one detail that matters: chain checkpoints also carry the scene prompt in their metadata. Load Latent surfaces it on the second output, prompt (a STRING) - so reopening a chain clip doesn't mean hunting for the prompt that made it. Save Latent files from this pack won't have a prompt to give you (nothing records it at save time), so expect an empty string there.

Inputs and outputs

One input: latent, the file dropdown. If nothing has been saved yet, it reads (none saved yet) and you know where you stand. On the output side:

• samples - the reconstructed LATENT. If the file recorded a NestedTensor, it's rebuilt as one; the original dtype is restored from the metadata, so downstream nodes see the precision the clip was generated in.
• prompt - the scene prompt, for chain checkpoints.

One thing to appreciate about the implementation: the dropdown is a widget, but a graph POSTed to /prompt can carry any string in its place, and os.path.join happily discards a base directory when handed an absolute path. The loader guards against that with a containment check and raises outside the output directory rather than letting a workflow read arbitrary files. For a node that mostly sits quietly in the corner, that's thoughtful hardening - worth knowing if you exchange workflows.

The workflow shape

Pair it with Save Latent at the end of a run, then later:

MiniMaxH3LoadLatent ──► MiniMaxH3DenoiseMask ──► SamplerCustomAdvanced.latent_image

Same latent that made the clip, so a region rewrite lands on footage it already knows. This is the "reopen and edit yesterday's commercial" loop the whole pack is built around.

Install

ComfyUI Manager, search MiniMax H3 Inpaint Tools, or:

cd ComfyUI/custom_nodes
git clone https://github.com/panghea/ComfyUI-MiniMax-H3-Inpaint-Tools
cd ComfyUI-MiniMax-H3-Inpaint-Tools && pip install -r requirements.txt

Restart and it appears under MiniMax H3/latent. The output directory it scans is the standard ComfyUI one, so files saved by this pack land where the loader will find them without configuration.