MiniMax H3 Image • Single Image Output

H3ImageFrameSelector - "MiniMax H3 Image • Single Image Output" - is the last node in the pipeline, and the one that answers the question everyone asks five seconds in: which of these 5 frames do I actually keep? H3 generates a whole temporal packet for a still, so somewhere a decision has to be made. This node makes it, either from a recommendation or by scoring the frames itself.

How selection works

The default strategy is decode_recommended, which trusts the recommendation from the Exact Frame Decode node. That's the sane starting point: wire recommended_index from H3ImageDecode into the optional recommended_index input here and let the packet's metadata decide. If nothing's connected, it falls back to frame 0.

When you want control, the strategy dropdown is where the depth is:

• Fixed picks: first, middle, last, and manual_index (which activates the manual_index integer).
• Scoring picks: sharpest (edge detail only), best_quality (sharpness + contrast + exposure), stable_quality (favors sharp, clean, temporally stable frames), balanced_edit and most_similar_to_source (both need source_image connected; similarity_weight, default 0.6, balances source similarity against raw quality in balanced_edit).

There are also knobs for the scoring range - candidate_start/candidate_end as fractions of the batch, skip_first_frames to ignore an unstable opening frame, and top_k for how many candidates candidate_batch_debug returns.

Outputs, in the order you'll use them: selected_image (the single kept still by default), candidate_batch_debug (the ranked top-k as an image batch), selected_index, selected_score, and score_report (a human-readable summary of what it picked and why).

Scoring is a diagnostic, not a fix

Here's the honest take: the scoring modes are diagnostics, not magic. If your prompt was weak or your edit instruction was vague, no frame is going to be good, and a scoring strategy can't invent one. The README says it plainly - these modes "cannot correct weak conditioning or an unclear edit prompt." Don't reach for best_quality hoping it fixes a bad generation; re-roll the seed or fix the prompt. The one genuinely useful non-default: emit_candidate_batch. Flip it ON and selected_image returns the entire decoded batch instead of one still, so a connected Preview Image or Save Image node shows or saves every frame. That's great for eyeballing candidates once, then you turn it back off - it deliberately retains more RAM/VRAM.

Install

Install is the pack: ComfyUI Manager (search MiniMax H3 Image Studio) or clone into custom_nodes, then restart. No extra dependencies beyond ComfyUI 0.30.0+, the H3 models, and the decode node upstream. If decode_recommended isn't in the strategy list, your node definitions are stale - update and do a full ComfyUI restart, not just a refresh, or the old selector will keep rejecting the recommendation input.