AD_MinMax_Ref2_generate

The FL2 generate node is great at one specific trick: anchor the start and end frame. AD_MinMax_Ref2_generate is the one you grab when "start and end" isn't enough - it's the Ref2VA sampler, meaning the whole video can lean on reference media, and the node gives you up to 64 reference slots to prove it. It's the same staged, self-merging pipeline as FL2, but the conditioning model is different: instead of a first/last image pair, you feed it images, videos, even audio, and H3 treats them as the thing the target clip should conform to. This is your character-consistency and video-to-video workhorse.

How it works

Same stage mechanics: stage_prompts is a JSON list, one prompt per stage; length snaps to the 17k+5 grid (124 ≈ 5s at 24 fps, per the author's own tooltip); and when the last stage fires you get merged_video out. The references come in through media plus media_1…64, each with a matching media_type_N string that tells the node what it's looking at. Unlike FL2, this node's media ports accept IMAGE, VIDEO, AUDIO, and LATENT - audio references are genuinely useful here because H3 generates native synchronized audio.

Two inputs are worth calling out before anything else:

• ref_image_size - match scales references to your canvas area (fast, good enough for most shots); max aligns to a 2048 short edge for better identity but slower. When the whole point of your shot is a recognizable face, max is usually the difference between "close enough" and "that's her."
• single_long_video_split / single_long_audio_split - for the long-reference workflow. Flip one on, feed a single long video (or audio) in, and the node splits it by stage; its embedded audio drives H3 during generation, while the final merge reuses the original continuous soundtrack. That's how you do a whole-scene remake with one source clip instead of juggling dozens of slices.

Then the standard second-pass trio (second_pass_mode → None/refine/latent_scale, with refine_model/refine_denoise/refine_steps or latent_model/latent_scale/split_step), which works exactly like the FL2 node.

Outputs and wiring

denoise_latent1 (current stage latent), segment_video (this stage's clip), merged_video (all stages, stitched on the last run), text. Decode merged_video at the end and you're done - that's the whole pipeline.

Install and the practical gotchas

Install via ComfyUI Manager (search "ComfyUI-Apt_Preset") or clone + install.bat. You need the MiniMax H3 weights and a ComfyUI build with H3 support; without it you'll get a "does not provide MiniMax H3 support" runtime error, which is a ComfyUI-version problem, not yours. The latent_scale path additionally wants the H3 3D upscaler in ComfyUI/models/latent_upscale_models - the dropdown placeholder shows you the path. And the standing H3 caveat: the open weights' community license excludes the US, EU, UK and South Korea, so check your region before building a workflow around this node.