MiniMax H3 Image • Resolution Preset

H3ImageResolutionPreset - "MiniMax H3 Image • Resolution Preset" - is the version of the resolution calculator that respects your time. Where its Advanced Resolution sibling makes you think about megapixel targets and grid multiples, this one hands you named profiles tuned to H3 and says "pick one." It's the node the bundled workflows use, and it's the right first stop for anyone who doesn't care about canvas math.

The profiles

Pick an aspect_ratio (or source image to copy a connected image's ratio), pick a resolution_profile, and it returns a width and height already rounded to H3's 32-pixel grid. The profiles are worth knowing by name:

• fast preview | 0.40 MP - quick drafts, low VRAM.
• balanced | 0.70 MP
• high | 0.90 MP
• native detail | 0.98 MP - the default, and the one to reach for. This is H3's native canvas (about 1344×768).
• high-res | 2.00 MP and up - ultra 4.00, ultra+ 8.00 - where the README's warning kicks in: higher profiles raise VRAM/RAM and decode cost, and "H3 learned detail does not necessarily scale proportionally." A 2 MP canvas can help small or distant details in some images; it is not a general quality upgrade.

Inputs and the oversize warning

Inputs: aspect_ratio, resolution_profile, plus three optionals - source_image (only for the source image ratio), custom_megapixels (only when the profile is custom megapixels), and limit_to_native_area. That last one is where this node differs from the advanced one: here it defaults to OFF, so if you choose high-res or ultra, you'll get the oversize canvas. The resolution_info string will tell you exactly how far past native you are ("UNLOCKED oversize (~2.0× native pixel area; VRAM and attention cost rise sharply") - that's the pack warning you, not flavor text. Flip limit_to_native_area on if you want it conservatively capped near the 768×1344 area.

Outputs are width, height, and resolution_info, all meant to feed the prepare nodes (Text to Image, Image to Image, Reference Edit, or Advanced Combined). The prepare nodes accept these INTs directly into their width/height inputs - that's the cleanest way to guarantee a legal canvas.

Quality tiers aren't what they look like

A small opinion: don't treat the resolution presets as quality tiers. They're cost and canvas-size tiers. The model's real estate is around the megapixel mark, so native detail isn't the "good" setting, it's the native setting, and everything above it is you paying more for the chance that distant details sharpen up. For most prompts, the money is better spent on the base 20-step sampling profile than on a 4 MP canvas.

Install

Install is the pack only: ComfyUI Manager (search MiniMax H3 Image Studio) or git clone https://github.com/astropuzzo/ComfyUI-MiniMax-H3-Image-Studio.git into custom_nodes, then restart. Requires ComfyUI 0.30.0+; the H3 models are only needed once the node is wired into a full workflow.