H3 Super: Draft to LTX Refiner / H3草稿转LTX二阶段 (Advanced/T8)

NVIDIA's H3 Super Acceleration is a two-stage pipeline with an unusual shape, and you need to understand it before you can use any of the Sol Engine nodes. Stage 1 is your normal H3 run but cut short - a 4-step draft that's fast and rough. Stage 2 is where the magic happens: the draft's decoded frames get handed to LTX-2.5, a completely different model, which refines them with a 3-step distilled pass at higher resolution. This node is the boundary between the two stages. It takes your decoded H3 draft frames and prepares them for the LTX refiner.

Three things happen in prepare_h3_draft_for_ltx_refiner:

1. Trim to the LTX 8n+1 frame grid. LTX video operates on frame counts that are multiples of 8 plus 1 (17, 25, 33...). Your H3 draft is probably 124 or 243 frames - not LTX-shaped. The default trim_to_8n_plus_1 policy trims the tail to fit without adding frames. There's a preserve_all_exp option, and it's explicitly for runtimes that accept a non-8n+1 length; the official route trims.
2. Aspect-preserving center-crop to half the requested output size. This is the part that feels backwards until you see the next stage: you target 1920×1088, so the frames get cropped to 960×544. Why? Because the official pipeline runs the LTX-2.5 x2 latent upsampler after encoding, so the refiner works on a latent that's already been grown. Crop half now, upscale 2x in latent, refine at full size.
3. Output frames for ComfyUI's full LTX-2.5 video VAE encode. The crucial detail: this node prepares frames for the full LTX VAE encoder, not the fast TAEHV preview codec. The refiner was trained on the full VAE's latent distribution, so feeding it TAEHV-encoded latents would be like giving an SD model latents from a different VAE - structurally wrong. The description hammers this: "TAEHV is only used for the final fast decode."

And the audio? H3 audio must bypass Stage 2 on a separate wire. The frames input's tooltip is unambiguous: "AUDIO is intentionally not accepted." Your H3 audio is trimmed to match the final duration and remuxed at the end - it never passes through LTX.

The inputs that matter

• frames - decoded H3 draft frames (IMAGE only).
• target_width / target_height - the final output size you want (defaults 1920×1088). The node crops to half of this for the encoder.
• frame_policy - trim_to_8n_plus_1 (default) or preserve_all_exp.
• fps - the source H3 frame rate, used only to compute the exact audio-bypass trim duration so the remuxed audio matches the trimmed video.

Outputs: ltx_encoder_frames (feed to the full LTX VAE encode), encoder_width / encoder_height (the crop size), kept_frames, dropped_tail_frames, output_duration_seconds, and report_json. The output durations matter because you'll use them to trim the H3 audio to match.

Where it sits

This is one node in the 22-sol-engine-h3-super workflow, which strings together: H3 draft → decode → this node → full LTX-2.5 VAE encode → x2 latent upsampler → the LTX-2.5 3-step refiner setup node → TAEHV fast decode → remux with the bypassed H3 audio. Use the pack's example workflow; this is not something to free-form. The whole Stage-2 model bundle is ~45GB (LTX-2.5 transformer, Gemma text encoder, distilled LoRA, video VAE, latent upscaler, TAEHV) from the t8star HuggingFace org.

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 models go in their standard ComfyUI folders (diffusion_models, vae, text_encoders, latent_upscale_models, loras, taehv), preserving the t8star repo's directory structure.