MiniMax H3 Prompt Relay Resource Estimate / 资源预估 (Advanced)

Every extra reference image, reference video, and audio track you throw at H3 grows the packed sequence length, and sequence length is the number that decides whether a render fits your card. MiniMaxH3PromptRelayResourceEstimateT8Advanced is a model-free estimate of that number plus the relay's bounded explicit chunk-bias allocation - a "what am I about to ask the transformer to chew on" preview that runs without loading any weights.

The first sentence of its description is the disclaimer you should tattoo on the workflow: it estimates native H3 packed rows and the relay's explicit bias peak, and it does not estimate total VRAM and is never a 16GB safety certificate. The author's README goes further: it's a memory planning proxy that excludes model weights, full attention activations, VAE/CLIP, VBAR, and fragmentation. Use it to compare configurations against each other, not to sign off on "will this fit."

The inputs that matter

• prompt_relay_plan - the plan whose events and length drive the estimate.
• width / height - 736×416 default; the report gives you the 736×416, 1152×640, and 1920×1088 reference matrices.
• query_chunk_rows - 256 default, and the tooltip warns it must match the Prompt Relay Conditioning node or the estimate is meaningless. This one gets people.
• precision - bf16_fp16 default; only affects the explicit bias-size estimate.
• The reference inventory, all manual: keyframe_stills, reference_images_match, reference_video_count / reference_video_frames_each / reference_video_has_audio / reference_video_audio_seconds_each, standalone_reference_audio_count / standalone_reference_audio_seconds_each.
• additional_text_rows (256) and manual_extra_packed_rows (0) - the fudge factors for Qwen system/media tokens and max-size references that don't fit the categories above. The tooltips are honest about what they're for: "manual allowance."

Outputs: prompt_relay_plan (pass-through), estimated_seq_len, peak_explicit_bias_mib, summary_text, and report_json. It's an output node by design - drop it at the end of the planning side of your graph and read it.

Why you'd actually use it

Because the alternative is learning your sequence length the hard way - a CUDA OOM at step 37 after a 15-minute ramp. H3's packed rows grow with every reference, and relay adds an explicit bias on top; this node gives you a number to compare against your last successful render ("my previous clip was ~5000 rows at 1056×608, this plan wants 9000 - probably not on this card"). The author's own numbers: a real run recorded 5139 tokens on a 0.737MP canvas, and relay's bounded chunking means the bias peak stays predictable rather than a full S×S mask. It won't save you from fragmentation or a swap file, but it'll catch the config changes that quietly double your sequence length.

Install

Part of the MiniMax H3 Audio T8 pack:

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

or search "MiniMax H3 Audio T8" in ComfyUI Manager and restart. No pip deps, no weights - pure arithmetic. The pure timeline/resource preflight workflow is examples/workflows/14-prompt-relay/2026-08-20_H3_Prompt_Relay_Plan_Preview_Advanced_EXP.json; run it, read summary_text, and keep the 512MiB-margin habit from the pack's preflight culture.