MiniMax H3 Still Preflight (EXP/T8)

The whole point of this node is to fail before the expensive part runs. MiniMaxH3StillPreflightT8 checks your experimental Ref2VA still-image setup - canvas, target frame count, reference strength, the model and VAE you've loaded - and tells you whether the sample is going to work, instead of letting you find out after a multi-minute H3 generate spits out a 32-pixel-wide smudge or a hard error.

H3 stills are a workaround, not a native mode. The model is trained on video grids, so the "still" route drives a one-frame (or micro-video) target through the Ref2VA path, and the contract is finicky: wrong canvas multiple, a target the model can't represent, reference images that don't match the edit image, and it all falls apart quietly. This preflight exists precisely because that contract is easy to get subtly wrong.

The settings that decide everything

• canvas_mode - from_edit_image inherits the canvas from your edit image and is the safe default; custom lets you type width/height yourself (must be multiples of 32, and the node enforces a ~1920×1088 pixel cap).
• target_mode - how many frames the "still" actually generates: direct_1_frame, micro_video_5_frames, short_video_22_frames, or trained_124_frames. This is the frame count you're committing to, so pick based on how much generation time you want to spend.
• reference_strength - how hard the model holds onto your reference image (0.999 default, same knob the Visual Reference Strength node exposes).
• audio_target - defaults to generate_and_discard, which is the honest way to treat audio when you only want a picture.

The optional model, video_vae, edit_image and ref_images inputs let it actually check that your loaded weights and references are compatible rather than just doing arithmetic. If you're going to leave any inputs connected, edit_image is the one that matters - it's the anchor everything else is validated against.

Outputs:

a ready boolean (the thing you gate a switch on), warning_count, and report_json with the details. Wire ready into a switch or a bypass and you've got a workflow that only runs the sampler when the contract is sound.

Installing:

this is a node in the MiniMax H3 Audio T8 pack. ComfyUI Manager → search "MiniMax H3 Audio T8", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart ComfyUI after. No extra pip packages, no auto weight download - models are on you.

Common trap:

treating the preflight as a rubber stamp. It checks the contract, not your taste - ready=true means the sample should run and decode, not that the composition is good. And if you see "prompt media tag validation failed," it's not this node's fault: check that your reference tags (<Picture 1>, etc.) and inputs actually line up before you blame the check. Also remember H3's local weights are geofenced out of the US, EU, UK and Korea by the Community License - preflight won't run without weights, so sort that out first.