H3 Super: LTX-2.5 3-Step Refiner Setup (Advanced/T8)

This is the heart of the H3 Super Acceleration Stage 2. After your H3 4-step draft has been trimmed, cropped, encoded with the full LTX-2.5 video VAE and grown 2x in latent space, this node configures the LTX-2.5 refiner that actually adds the detail back. It's the "3-step" in the pipeline's name: three Euler updates at the official sigma schedule [0.909375, 0.725, 0.421875, 0], taking the latent from heavily-noised to fully-denoised.

Before you do anything: apply the official LTX-2.5 distilled refiner LoRA to your LTX model at strength 0.8. The node's model input tooltip says it plainly - it expects "LTX-2.5 dev transformer with the distilled refiner LoRA applied at 0.8" - and it even hands you the strength back as an output (refiner_lora_strength, always 0.8) so you can't lose track of it. This is distilled-LoRA-on-dev-model, the same pattern the LTX ecosystem standardized on: a fast LoRA stacked on the base checkpoint rather than a separate distilled checkpoint.

What it configures

• model - your LoRA'd LTX-2.5 dev transformer. Output is the patched model.
• sampler - always Euler (hardcoded to the official contract).
• sigmas - the official four-sigma schedule above.
• attention_backend - the interesting knob. Default auto_sol_attn tries to route self-attention layers 1-47 through a loaded ComfyUI-SolAttn_triton backend with tau 1.0/1.25/1.5, while keeping layer 0 dense. If Sol-Attn isn't installed or loaded, it falls back to dense without blocking - no error, no crash, just a note in the report. dense_reference skips Sol-Attn entirely for clean A/B comparisons.
• min_tokens (4096), kernel_precision (bf16_official default), and verbose are advanced; leave them alone unless you know.

Outputs: patched model, sampler, sigmas, the refiner_lora_strength (0.8), and report_json telling you which attention path actually ran.

The honest picture

Sol-Attn (from the ComfyUI-SolAttn_triton project) is an optional dependency, and the pack treats it exactly that way - nice when present, invisible when not. The README is clear that the validated-first default for identity work is dense attention, and the only A/B guarantee is the sigma schedule and the LoRA. If you want to know whether Sol-Attn is actually helping you, run dense_reference on the same seed and compare. Also note: the LTX-2.5 Stage-2 refiner works on the full LTX VAE's latent distribution, which is why the upstream Draft-to-LTX node feeds it full-VAE latents and why the TAEHV encode node is explicitly banned from this path.

Install

Same pack:

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. The Stage-2 model bundle (~45GB total) comes from the t8star HuggingFace org - LTX-2.5 dev transformer, Gemma text encoder, the distilled LoRA, LTX-2.5 video VAE, the x2 latent upscaler, and TAEHV - placed in the standard ComfyUI model folders. Sol-Attn is a separate install and genuinely optional.