What this does

A 2-stage ComfyUI pipeline that pairs MiniMax H3's ref2va generation (strong subject/identity fidelity from reference images, native synced audio) with LTX-2.5's fast distilled sampler as a refine/upscale pass. Built using quantized (int8) models with dynamic VRAM loading — runs comfortably on lower-VRAM cards (3060 / 4070 / 5060 class), not just high-end GPUs.

Stage 1 — MiniMax H3 ref2va: generates your base clip from reference images (face/identity, outfit, environment) with native audio, exactly like MiniMax's own reference workflow.

Stage 2 — LTX-2.5 refine: MiniMax's decoded frames are re-encoded into LTX-2.5's own latent space (the two models don't share a latent space, so this bridges through pixel space) and run through LTX's fast distilled sampler for detail/sharpness — either:

- Native 2x upscale using LTX's own latent upscaler model, or

- Custom target resolution (1920x1080, 1280x720, 1080x1920, or any arbitrary size) via a pixel-space resize before re-encoding

MiniMax's original native audio is passed straight through to the final output — LTX's audio path isn't used at all.

What's tunable

- Sigma schedule + video_cfg on the refine pass — trade off sharpness vs. staying faithful to MiniMax's exact output

- Resolution mode toggle — native 2x model upscale vs. exact custom resolution

- Full prompt/negative-prompt control on the refine pass separately from your MiniMax generation prompt

Not included in this version (coming later)

- RIFE motion pre-interpolation — reduces motion artifacts (ghosting/warping) on fast-action shots by halving the per-frame motion delta the video VAE sees.

- ReActor face-lock — locks final output identity to a reference photo, for cases where the base generation is too low-res for the refine pass to reliably hold facial identity on its own.

Want early access to the upgraded version with RIFE + ReActor before public release? I'm posting it early for supporters here: https://www.patreon.com/c/iiTzMYUNG

Requirements

- MiniMax H3 ref2va models (unet, clip, dual VAE)

- LTX-2.5 models (distilled transformer, video VAE, gemma clip, x2 latent upscale model)

- Standard ComfyUI, no custom node packs required for this version

- Runs on 8-12GB VRAM cards (3060/4070/5060 class) thanks to int8 quantized models + dynamic VRAM loading — no high-end GPU required

Notes

Built and iterated through real trial-and-error on sharpness/identity/resolution tradeoffs — the note nodes inside the workflow explain the reasoning behind each setting if you want to retune it for your own use case.