SoL-Refiner (LTX-2.5 for MiniMax-H3) — int8_convrot, image-server layout
One-step 1080p refiner for MiniMax-H3 videos: an H3 draft is encoded, 2x latent-upsampled, re-noised to
sigma 0.909375 and denoised in one transformer forward, then decoded by the refiner's diffusion decoder.
Repackaged from NVlabs' SoL-Refiner-LTX-2.5-for-MiniMax-H3
(project: SoL-Refiner) for image-server's LTX-2.5 ltx_core loader.
| File | Content |
|---|---|
diffusion_models/sol-refiner-ltx-2.5-h3-transformer-comfy-int8-convrot.safetensors | DiT, 1344 block Linears int8 ConvRot (g256, per-row); text connectors, gates, adaLN, embedders bf16 |
text_encoders/gemma4-12b-with-proj-sol-refiner-ltx-2.5-h3-comfy-int8-convrot.safetensors | the refiner's Gemma 4 12B (328 LM Linears int8 ConvRot) + caption projection (bf16) + embedded tokenizer/processor assets |
vae/sol-refiner-ltx-2.5-h3-video-vae-bf16.safetensors | LTX-2.5 video VAE encoder + the refiner's own diffusion decoder (NVlabs-modified, not Lightricks') |
latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors | Lightricks LTX-2.5 x2 latent upsampler (identical to the refiner's) |
Notes
- int8 layer set = Lightricks' official LTX-2.5 int8-convrot set (transformer) / int8 Gemma 4 set (text encoder),
produced with convert-to-quant 1.3.3 from the refiner's bf16 weights; runs on
image_server_kernels.int8_linear. - The text encoder is the refiner's shipped Gemma 4 — the pre-2026-08-17 Lightricks encoder, not the current one.
- Conditioning
frame_rate= the input fps (H3 drafts are 24 fps). One step, no CFG. Output video only (no audio). - Transformer math verified against the Diffusers reference within bf16 noise; int8 text encoder vs bf16: connector-output cosine >= 0.99995.
License: the source repository declares none. SoL-Refiner is a fine-tune of LTX-2.5, so the LTX-2.x Community License (Lightricks) applies; check it before commercial use.