📂 H3 Library Load - LD

The other half of the "continue that take" workflow

H3 Library Load (📂 H3 Library Load - LD, under LD / PromptMaster / Long form, in PromptMasterLD) is the reload side of H3 Library Save. It loads a clip out of input/h3_extend/ - the library every H3 Master or H3 Library Save render writes to - and hands you back the three things a continuation needs: the frames, the audio, and the latent. Wire them into H3 Master's extend_video, extend_audio and extend_latent inputs and the next render starts from that clip's exact tail instead of from black.

The loading is just plumbing. The interesting design is the latent: that's the bit that makes the join seamless. A clip saved with its sampler latent gives the continuation the exact pixels and motion context the original sample produced; a clip saved without one still loads, you just lose that output and settle for a softer, re-encoded join.

The inputs

• clip - a dropdown of everything in the library, newest first. Every video H3 Master or H3 Library Save has written lands here.
• tail_seconds (default 0) - 0 means continue from the END of the clip, which is what you want most of the time. Set it to a positive value to continue from earlier instead: 2.0 drops the last 2 seconds and continues from there. The latent is cut to the same point automatically, so picture and motion context can't disagree. Useful when the tail of a take has a bad few frames you'd rather not inherit.
• path_override (optional) - H3 Studio's extend_path output. Wired, it overrides the dropdown and loads exactly the clip you attached in the Studio panel's EXT tab. This is the only way an uploaded file can reach the sampler, because the dropdown is built from the library at import time and an upload isn't in the library yet. Empty falls back to the dropdown.

Outputs

• frames (IMAGE) → H3 Master's extend_video.
• audio (AUDIO) → extend_audio. A clip saved without audio gives you nothing here - silent continuation.
• latent (LATENT) → extend_latent. Present only if the clip was saved with its sampler latent.
• report (STRING) - what it loaded and from where.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Lora-Daddy/PromptMasterLD
cd PromptMasterLD
pip install -r requirements.txt

Restart ComfyUI, or ComfyUI Manager → search "PromptMasterLD".

The gotchas

Three, all worth knowing before you queue. One: the dropdown is frozen at import time - if you save a clip after ComfyUI started, it won't be in the clip list until you reload the workflow or restart. That's exactly what path_override is for, so the Studio EXT tab path is the reliable route for brand-new uploads. Two: a clip whose audio was never saved comes back silent, and so does its continuation - check the library entry rather than assuming. Three: tail_seconds trims both the frames and the latent together, which is the correct behaviour, but it means you're not really "continuing from the end" anymore - you're continuing from a point inside the clip, so write the extension prompt for what's actually happening at that cut.