SoL-Refiner on MiniMax-H3 — LTX-2.3 one-step vs LTX-2.5 H3 refiner (blind, 1080p)

NVlabs SoL-Refiner turns a low-res draft into 1080p with one denoising step. Two checkpoints on the same 10 MiniMax-H3 clips (5 T2V from minimax-h3-prompt-format, 5 I2V from minimax-h3-i2v-prompt-format; first 5 s = 121 frames): SoL-Refiner-LTX-2.3-One-Step (paper model, generic; σ 0.725 + AdaIN) and SoL-Refiner-LTX-2.5-for-MiniMax-H3 (H3 extension; σ 0.909, NATTEN diffusion decoder). A/B order varies per clip. Pick a pair per card — the slider keeps the timestamp. 1:1 = native pixels. Judge on shippability and faithfulness to the input, not on resemblance. Votes stay in this browser — use download votes.

RTX 5090, 32 GB · DiT int8 ConvRot (ISK) · TE/decoders bf16 refiner prompt = the clip's RAW prompt ⚠️ LTX-2.3 refiner run at frame_rate 16: at the input's 24 it ghosts moving objects ⚠️ neither HF repo declares a license

refiner	refine / clip (median)	DiT 1 step	peak VRAM
LTX-2.3 one-step (frame_rate 16)	27.5 s	7.3 s	28.0 GiB
LTX-2.5 H3 refiner	49.3 s	7.4 s	28.6 GiB