SoL-Refiner (LTX-2.3 One-Step) — int8_convrot, image-server layout
Generic one-step refiner (SANA-Video, WAN, MiniMax-H3, ... drafts -> 1080p/2K): the draft is encoded, 2x
latent-upsampled with AdaIN, re-noised to sigma 0.725 and denoised in one transformer forward.
Repackaged from NVlabs' SoL-Refiner-LTX-2.3-One-Step
(project: SoL-Refiner) for image-server's LTX-2.5 ltx_core loader.
| File | Content |
|---|---|
diffusion_models/sol-refiner-ltx-2.3-one-step-transformer-comfy-int8-convrot.safetensors | DiT, 1344 block Linears int8 ConvRot (g256, per-row); text connectors, gates, adaLN, embedders bf16 |
text_encoders/gemma3-12b-with-proj-sol-refiner-ltx-2.3-one-step-comfy-int8-convrot.safetensors | Gemma 3 12B (all 336 LM Linears int8 ConvRot; embeddings + vision tower bf16) + caption projection (bf16) + embedded assets |
vae/ltx-2.3-22b_vae.safetensors | LTX-2.3 video VAE (identical to the refiner's) |
latent_upscale_models/ltx-2.3-spatial-upscaler-x2-1.1.safetensors | Lightricks LTX-2.3 x2 latent upsampler v1.1 (identical to the refiner's; not v1.0) |
Notes
- ⚠️ Always condition with
frame_rate=16, whatever the input fps (encode the output at the real fps). The model was trained on 16 fps drafts; passing 24 fps mis-scales the temporal RoPE and ghosts moving objects. - int8 sets follow Lightricks' LTX-2.x int8-convrot recipe (transformer) and Comfy's fp8 Gemma 3 set (text encoder),
produced with convert-to-quant 1.3.3; runs on
image_server_kernels.int8_linear. - Transformer math verified against the Diffusers reference within bf16 noise; int8 text encoder vs bf16: connector-output cosine >= 0.99994.
License: the source repository declares none. SoL-Refiner is a fine-tune of LTX-2.3, so the LTX-2.x Community License (Lightricks) applies; check it before commercial use.