AD_MinMax_FL2_mul

AD_MinMax_FL2_mul doesn't generate anything, and that's the whole point. Its source docstring says it plainly: "Queue-stage FL2VA conditioning for an external sampler and video output." Where AD_MinMax_FL2_generate bundles sampling and merging into one node, this one is the conditioning half - it turns your prompt, reference images, and stage plan into a RUN_CONTEXT that you hand to a sampler (or this pack's controller stack) yourself. You reach for it when you want the stage-based FL2 prompt logic without letting the pack own your sampler choice.

How it works

Like its generate sibling, it lives in the flow-stage pipeline. You feed it the same prompt, width, height, length, and a JSON stage_prompts list; it runs once per stage and updates the context for that segment. The interesting output is trim_frames (default 22): the number of motion-context frames reused from the previous continuation clip and removed from the new one. That's the seam-healing trick - H3 carries forward the tail of the last clip as context so the next segment starts from real motion instead of a blank slate. 22 is the sweet spot for most 5s H3 clips; raise it if you see flicker at cuts, lower it if your segments feel stuck in the previous pose.

Note what's not here: no seed input. It's deliberately popped out of this node - the seed belongs on your sampler, which is part of the design, not a bug.

The inputs that bite

• stage_prompts - JSON array, one prompt per stage. The only thing you must get right.
• default_trim_frames - 1 to 56, default 22. Sets trim_frames for every stage that doesn't override it.
• media / media_1…64 - and here's the trap: FL2 conditioning only accepts image, latent, or text. The source raises "AD_MinMax_FL2_mul Media only accepts image, latent or text" if you wire in video or audio, and it's strict about counts too: at most two ordered images (first-frame + last-frame references) and one context latent. Anything else is an error, not a warning.

Optional context and model pass through to the sampler, and stage_info/stage_data hook it into the flow system.

Outputs

• context (RUN_CONTEXT) - wire this into your KSampler-style node or the pack's sampler/controller.
• trim_frames (INT) - the context-frame count for continuation.
• text (STRING) - the resolved FL2 prompt this stage encoded, handy for debugging exactly what the model saw.

Install and issues

Same pack install as everything else in ComfyUI-Apt_Preset: Manager → search the pack title, or clone into custom_nodes and run install.bat. Needs MiniMax H3 weights and a current ComfyUI. If your context errors with "missing: clip, vae, audio_vae," the RUN_CONTEXT you're feeding it wasn't built by the pack's loader/controller - it needs all three, so route it through the pack's context chain. And the usual H3 license reminder applies: the open weights are territory-locked out of the US, EU, UK and Korea.