Nodes/ComfyUI-Arisu-Nodes/MiniMax H3 Video Settings (Upscale)
ComfyUI Node

MiniMax H3 Video Settings (Upscale)

The second-pass numbers, derived for you

By swqa7697·Created 5 days ago·Updated about 12 hours ago· 4
MiniMax H3 Video Settings (Upscale)
    • width
    • height
    • length
    • upscale_factor
    • target_width
    • target_height
    aspect_ratio16:9 (Widescreen)
    megapixels1.0
    upscale_factor2.00
    duration5.0
    advertisefalse

    This is MiniMax H3 Video Settings for the two-sampler latent-upscale workflow, and it exists because that workflow has a bookkeeping trap in the middle of it. When you generate an H3 clip at one resolution and refine it with a second sampler at a higher one, every node in the chain needs to agree on two separate sizes: the generation canvas and the upscaled target. Mistype either and the second pass gets conditioning and a latent that don't match. This node derives both from the same inputs, so they can't disagree.

    Mechanically it's the plain settings node plus an upscale_factor: the canvas comes from aspect_ratio + megapixels on the 32-pixel grid, length from duration in seconds on the 17k+5 frame grid, and target_width/target_height are the canvas scaled by the factor and rounded back onto multiples of 32.

    The inputs and outputs that matter

    • aspect_ratio, megapixels, duration - exactly as on the plain settings node (default 16:9, 1.0 MP, 5.0 s).
    • upscale_factor - the multiplier the latent upscaler between your two samplers applies, default 2.0, range 1–8. This is the number you previously had to compute twice and keep in sync.
    • advertise - same behavior as the plain node, but it drives the Advanced hybrid node's target_width/target_height too.
    • Outputs: width, height, length (as before) plus upscale_factor (a FLOAT, for a latent upscaler's multiplier input) and target_width, target_height - the upscaled canvas, both multiples of 32.

    Wiring it

    Video Settings (Upscale) ─ width / height / length ─▶ Hybrid to Video (Advanced)
                             ├ target_width / target_height ─▶ Hybrid to Video (Advanced)
                             ├ upscale_factor ─▶ Latent Upscaler (multiplier)
                             └ target_width / target_height ─▶ Latent Upscaler (target dims)
    

    With a Minimax H3 Latent Upscaler (3D) in multiplier mode, its 32-pixel alignment matches this node's rounding; in target-dimensions mode, wire target_width/target_height straight into it so the upscaler and the Advanced node agree by construction.

    Installing it

    Same pack, same zero-dependency story: ComfyUI-Arisu-Nodes (swqa7697, GPL-3.0), needs ComfyUI ≥ 0.30.0 for H3 support. Manager → Install Custom Nodes → "ComfyUI-Arisu-Nodes", or clone into custom_nodes:

    cd ComfyUI/custom_nodes
    git clone https://github.com/swqa7697/ComfyUI-Arisu-Nodes.git
    

    Restart, then it's under Add Node → Arisu Nodes.

    Gotchas

    The one thing to keep an eye on is that the upscale factor only produces a sane target if the latent upscaler actually lands on it. A factor of 2.0 on 1344×768 gives 2688×1536 cleanly, but less convenient factors get snapped to the 32-grid on both sides - so if your second sampler throws a shape error, check that the upscaler's own rounding matches this node's before blaming the graph. Everything about advertising (root graph only, one advertiser per graph) carries over from the plain settings node unchanged.

    CategoryArisu Nodes/MiniMax H3

    Inputs (5)

    NameTypeDefaultDescription
    aspect_ratioCOMBO16:9 (Widescreen)Aspect ratio of the generation canvas.
    megapixelsFLOAT1.00.1–16Pixel budget of the canvas in megapixels (1 MP = 1024 x 1024), rounded to multiples of 32. The stock 1344 x 768 canvas is about 0.98 MP.
    upscale_factorFLOAT2.001–8Factor of the latent upscaler between the two samplers; the target size is rounded to multiples of 32.
    durationFLOAT5.00.2–150Clip length in seconds at 24 fps, snapped up to the model's 17k+5 frame grid (5.0 s = 124 frames).
    advertiseBOOLEANfalseDrive every MiniMax H3 Hybrid to Video node in this graph: their size and length widgets grey out at once, refuse links, and take these values. Off by default. Read by the frontend; the outputs stay available either way.

    Outputs (6)

    NameTypeDescription
    widthINTCanvas width in pixels, a multiple of 32.
    heightINTCanvas height in pixels, a multiple of 32.
    lengthINTFrame count on the 17k+5 grid, for the hybrid nodes' length input.
    upscale_factorFLOATThe factor, for a latent upscaler's multiplier input.
    target_widthINTUpscaled width in pixels, a multiple of 32.
    target_heightINTUpscaled height in pixels, a multiple of 32.