H3 Studio • Single Image Output

Here's the thing nobody warns you about with MiniMax H3 for images: you don't get one image out of a sampler. You get a batch of frames - 5, 9, 13 or 20 of them, depending on your frame profile - because H3 denoises a temporal packet and each frame is a candidate still. H3StudioFrameSelector is the node that turns that pile of frames into the one image you actually wanted.

It takes the decoded frames batch (from H3 Studio · Native H3 VAE Decode) and selects a single still. The default strategy is decode_recommended, which is the lazy-correct choice: it uses the recommended_index output from the decode node. Connect the decode node's recommended_index into this node's recommended_index input and it just takes the decoder's recommendation - the mode-aware pick that usually corresponds to the most settled frame.

When you want more control

The other nine strategies exist for when the default isn't doing it for you:

• first / middle / last / manual_index - fixed positions in the batch. manual_index is zero-based and only used with the manual_index strategy.
• best_quality / stable_quality / sharpest - metric-based scoring across the batch.
• most_similar_to_source - needs a source_image connected; picks the frame closest to your source, useful for image-to-image where you want the result to stay anchored.
• balanced_edit - a blend of source similarity and quality; similarity_weight (default 0.6) sets the balance, and it only matters when source_image is connected.
• decode_recommended - the default, described above.

The scoring strategies respect a range you can narrow: candidate_start and candidate_end are fractional bounds (0.0 → 1.0) over the decoded batch, and skip_first_frames excludes that many initial frames from scoring. If you know the first few frames of a profile are always unstable, skip them rather than fighting the scores. top_k (default 4) controls how many top frames the debug output carries.

What comes out

The main output is selected_image - your still. The rest are debugging aids:

• selected_index and selected_score - which frame won and why.
• score_report - a string breakdown of the scoring.
• candidate_batch_debug - here's the gotcha: this is empty by default unless you enable emit_candidate_batch. Flip it on when you're tuning a strategy and want to see every candidate frame; leave it off for real runs to avoid carrying a full image batch through the graph for nothing.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/thaakeno/ComfyUI-MiniMax-H3-Studio.git
cd ComfyUI-MiniMax-H3-Studio
python -m pip install -r requirements.txt

Restart and hard-refresh the frontend. No extra dependencies beyond the pack. My honest advice: run it on decode_recommended first, and only reach for the metric strategies when the recommended frame is consistently the wrong one - the scores are fast, but the default is usually the one the H3 decoder itself thinks is settled.