MiniMax H3 RefDelta Comparison Replay

What this is

This is the pack's diagnostic workhorse: a disk-backed sampler that replays exact, previously captured sampler states through whatever MODEL you have loaded, then hands back the original captured fused result. The name is doing real work - it's a comparison replay, and it's how you answer a genuinely hard question about the RefDelta fused checkpoint.

The question: the fused checkpoint claims to be FL2VA plus a rank-1024 approximation of the Ref2VA delta. But how good is that approximation, actually? This node measures it, on the same latent, same seed, same sigma, same state - a controlled experiment instead of eyeballing two videos.

The three-pass workflow

A complete comparison needs three runs of the same workflow, each loading one full H3 model. There's no shortcut; the memory hog is the point. Loading one model at a time is exactly what makes this the preferred diagnostic path over the old simultaneous dual-model guider.

Pass 1 - fused capture. Use the [Advanced/Diagnostic] MiniMax H3 RefDelta Sampler with the fused model loaded, calibration_capture = true, and a calibration_id you'll reuse. For a clean vector-field comparison, keep the sampler quiet:

calibration_capture = true
calibration_id = int8-convrot-r1024-test-01
adaptive_order = false
stochastic_adaptation_strength = 0
trajectory_correction = false

This writes exact packed sampler states, fused x0 values, sigmas, actual/forecast classification and telemetry to ComfyUI/output/refdelta_calibration/<calibration_id>/.

Pass 2 - FL2VA. Switch the MODEL loader to FL2VA, replace the sampler with this node, keep the same calibration_id, and set comparison_label = fl2va. Only captured actual steps get evaluated, at the exact captured state/sigma. This is the base-model same-state pass.

Pass 3 - Ref2VA. Load genuine Ref2VA and set comparison_label = ref2va. This requires the completed FL2VA pass. Now the node can decompose the true Ref2VA − FL2VA field against the fused fused − FL2VA field - cosine, magnitude, projection, residual and orthogonal components, for both video and audio streams.

The inputs

Only three, which is refreshing:

• calibration_id - must match the ID used during Pass 1 capture. The files live on disk, so a typo here means "nothing found."
• comparison_label - fl2va or ref2va. This labels which model's pass you're recording.
• telemetry_prefix - filename prefix for the comparison's own telemetry.

Output is a single SAMPLER. Importantly, it always returns the captured fused final state, so the original Continuum trajectory survives across all three diagnostic passes - you're not destroying your reference run by replaying against it.

What it's for (and not for)

Read this carefully: the FL2VA/Ref2VA metrics are explanatory. They tell you how much the rank-1024 approximation retains (the pack's own profile metadata notes the compressed matrices keep roughly half the original delta energy, with the exact pieces dominating overall). They are not a quality oracle, and they never feed production stochastic control or scheduler density automatically. If a tool ever makes you choose a scheduler based on Ref2VA distance, that's the trap this pack explicitly warns against.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-MiniMax-H3-RefDelta-Solver.git

Restart ComfyUI. No pip dependencies; you do need the fused checkpoint plus stock FL2VA and genuine Ref2VA H3 models on disk, one loaded per pass.

Common issues

• Missing capture. The replay needs the fused capture from Pass 1 to exist under ComfyUI/output/refdelta_calibration/<calibration_id>/. No capture, no replay - the node fails rather than guessing.
• Ref2VA before FL2VA. The Ref2VA pass requires the completed FL2VA pass. Do them in order.
• Spectrum is rejected on comparison passes. The replay evaluates actual model calls only; forecasting would break the same-state guarantee.
• It's not a scheduler tuner. The numbers are explanatory diagnostics, not a reason to re-densify your sigma schedule. Production scheduling stays on BasicScheduler with beta.