AD_MinMax_FL2_generate

If you've ever tried to make a longer MiniMax H3 video by chaining clips, you know the pain: each segment drifts, the seams show, and by segment four the thing has forgotten what it was doing. AD_MinMax_FL2_generate is this pack's answer - a H3 (Hailuo) sampler designed around the pack's "flow stage" system, where a long scene is split into N stages, each stage gets its own prompt, and the node generates every segment and merges them into one continuous video. The FL2 part is the conditioning trick: "first/last" reference images, so Picture 1 anchors the 0:00 mark and Picture 2 pins the final frame. That's how you keep a character and a scene stable across segments instead of letting each one reinvent itself.

How it works

It's a "queue-stage FL2VA sampler." It reads the stage_info socket that comes from flow_stage_begin, figures out which segment of the run it's on, samples that segment, and only when the last stage fires does it produce the merged output. Your per-stage prompts go in as a JSON list in stage_prompts (["shot one...", "shot two..."]). Frame count matters: length gets snapped up to the H3 grid of 17k+5, so 124 frames ≈ 5 seconds at the default 24 fps. Don't fight the snap - just treat 124 as "about 5s" and adjust up.

The second pass is where this node earns its keep. second_pass_mode gives you three mutually exclusive options:

• None - single pass, fastest, usually enough for drafts.
• refine - a low-denoise second sampling pass (refine_denoise 0.3, refine_steps 8 by default) to clean up the image. The tooltip on refine_model is the important bit: leave it on None, because "None or a 1x model" is what preserves latent continuity across stages.
• latent_scale - upscale between two sigma ranges. split_step is the SplitSigmas point: high-sigma steps run before the latent scaling, low-sigma steps after, and latent_scale (default 1.3) is the multiplier. This needs the MiniMax H3 3D upscaler, selected in latent_model.

Inputs and outputs that matter

You'll actually set prompt, width/height (512×768 default), length, seed, and stage_prompts. Everything else has sane defaults. Optional inputs worth knowing: context (the pack's RUN_CONTEXT bus - if you use the loader/controller stack, wire it), model to override the sampler's model, and the media/media_1…64 inputs. FL2 mode is strict about media: at most two ordered images (first frame + last frame) and one context latent - video or audio references will error.

Outputs: denoise_latent1 (this stage's LATENT - decode it to see where you are), segment_video (the current stage's clip), merged_video (the full stitched run, populated on the last stage), and text.

Install and the gotchas

Install the pack (ComfyUI Manager, search "ComfyUI-Apt_Preset", or git clone https://github.com/cardenluo/ComfyUI-Apt_Preset into custom_nodes), then run install.bat and restart ComfyUI. You also need MiniMax H3's weights on the machine plus a ComfyUI build that includes H3 support - if you hit the error "This ComfyUI build does not provide MiniMax H3 support," that's your ComfyUI being too old, not the node.

The big one people forget: the MiniMax H3 Community License geofences the weights out of the US, EU, UK and South Korea - running local H3 in those regions isn't licensed, full stop. And the latent_scale path silently needs the minimax_h3_latent_upscaler_3d models in ComfyUI/models/latent_upscale_models; the dropdown placeholder literally tells you where to put them.