H3 Super: LTX-2.5 Low-Sigma Identity Refiner / LTX低Sigma保脸细化 (Advanced/T8)

The official H3 Super Acceleration refiner works - most of the time. But it fully denoises the LTX latent, and "fully denoise" can mean the refiner rebuilds a face into something that's clearly no longer the person from your H3 draft. That's the exact complaint the pack's identity-preservation experiment is built around. This node is the low-sigma alternative: same three Euler updates, same complete terminal denoising, but the schedule starts much lower, so the refiner has far less room to rebuild identity.

The default schedule is [0.5, 0.412, 0.350, 0] - versus the official route's [0.909375, 0.725, 0.421875, 0]. Starting at 0.5 instead of 0.91 means the refiner inherits more of the original latent and mostly polishes rather than regenerates. It's the difference between "add detail on top of what's there" and "denoise from scratch." For a face that keeps drifting into someone else, low-sigma is the lever.

Before you use it, the honest framing

The README is careful here, and you should be too. This is not the official NVIDIA sigma-aligned route, and the pack does not claim it's better for all content. It exists so you can A/B: when the official fully-denoised version changes a face too much, run this on the same seed and compare. The identity_preserve_0p5 schedule mode is a preset; manual_exp lets you type your own sigmas (comma-separated, finite, strictly descending, ending at zero) if you want to probe the space. The tooltip's warning is exact: the preset always uses 0.5, 0.412, 0.350, 0 - no surprise schedule drift.

Same inputs as the official refiner node otherwise: your LTX-2.5 dev model with the distilled refiner LoRA applied at 0.8, enabled, attention_backend, min_tokens, kernel_precision, verbose. Outputs are the patched model, sampler (Euler), sigmas, refiner_lora_strength, and report_json.

The attention question, stated plainly

The default attention_backend here is dense_reference - dense ComfyUI attention - and the README explains why: it isolates the sigma change as the only variable in your A/B. The auto_sol_attn_conservative_exp option routes through Sol-Attn but with a deliberately conservative constant tau 1.0 policy instead of remapping your custom knots onto the official per-step tau schedule. That's an experiment inside an experiment. Start dense, compare, and only add Sol-Attn if you want a second axis of fiddling.

Install

Same pack, same drill:

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

or ComfyUI Manager → "MiniMax H3 Audio T8" → restart. Update ComfyUI itself first. Same Stage-2 model bundle as the official route, from the t8star HuggingFace org, in the standard ComfyUI folders. The H3 audio still bypasses on its own wire - this refiner never touches it.