JZL - 🌊 海螺H3视频参数Pro

"JZL - 🌊 海螺H3视频参数Pro" is the plain "海螺H3视频参数" node with exactly one thing added: a fifth output for the second-pass upscale. If you've ever run H3, sampled, then wanted a sharper pass, this is the node that hands you the multiplier.

The plain version computes Width, Height, Frames, and the reference-image scale factor. The Pro version adds upscale_scale (default 1.5, range 1.1–4.0) and outputs it as a fifth value - the number you feed into the MiniMax H3 Latent Upscaler (3D) scale input when you re-sample a clip at higher resolution. That's the whole difference. If you never do the two-pass latent-upscale dance, the non-Pro node is everything you need and Pro just adds a dangling port.

How it works

Everything underneath is identical to the base node: the official ResolutionSelector formula (megapixels × 1024², sqrt-scaled to the aspect ratio, rounded to multiple, default 32) for width and height, and the duration → frames conversion at H3's fixed 24fps snapped to the model's 17k+5 temporal grid (124 frames ≈ 5s). The Pro node simply passes your upscale_scale through as an output instead of discarding it.

Where it matters in the workflow: you do a first pass at, say, 1344×768, upscale the latent with a 3D latent upscaler, then sample a second pass at the larger resolution. The Pro node gives the upscaler its scale, and the pack's "海螺H3二采条件同步" (CondSync) node automatically reads the upscaled latent's actual dimensions so the first/last-frame keyframes line up - which is why the upscale_scale tooltip is honest that you don't need to route this number to CondSync yourself.

The inputs that matter

• aspect_ratio - ten presets, 1:1 to 21:9.
• megapixels - the area knob, 0.1–16, default 1.0.
• duration - 4–15 seconds; links with the Script Processor's per-segment duration.
• upscale_scale - the second-pass multiplier, default 1.5. Must match the align setting on your latent upscaler (both should round to 32).

Outputs: Width, Height, Frames, 参考图放大系数, 二采放大倍数 (the upscale scale).

How to install it

Same pack install as everything else - ComfyUI Manager, search "ComfyUI-JZL-MiniMax-H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/wjluoxiao/ComfyUI-JZL-MiniMax-H3

then restart. It's a pure math node, so it needs nothing beyond torch and a recent ComfyUI (the pack's V3 nodes need comfy_api.latest, the same build that carries the official H3 nodes). The node itself contains no model and no weights.

Common issues

The usual failure is a mismatch between upscale_scale and your latent upscaler's own scale - if the two-pass latent is upscaled 1.5× but you told the sampler 2×, the keyframes misalign and you get a smeared first frame. Keep Pro's multiplier and the upscaler's scale in lockstep. And remember the surrounding H3 context still applies: the weights are ~42.5GB and the MiniMax H3 Community License excludes the US, EU, UK, and South Korea from running them locally.