SoL-Refiner LTX-2.3 MLX BF16
AI2Apps video upscaling checkpoint, derived from NVIDIA SoL-Refiner / Lightricks LTX-2.3. Original: Efficient-Large-Model/SoL-Refiner-LTX-2.3-One-Step at c69c2a543997fe12c1ae24c776df6188e5d2248a.
Modified: unused audio and vision tensors removed; remaining tensor payloads preserve original BF16 bytes. No quantization. Safetensors shards are repacked and are intended for the AI2Apps MLX implementation, not generic Diffusers loading.
Standard upscaling needs transformer/, vae/, latent_upsampler/, scheduler/, model_index.json, default-prompt-context.safetensors, context-metadata.json and META/. It uses the fixed prompt recorded in context-metadata.json (about 28.71 GB). Custom prompts additionally need text_encoder/, connectors/ and tokenizer/ (about 28.33 GB extra). All files together are about 57.04 GB. Existing video files can be reused by checksum.
9/33-frame 128-to-256 tests on one portrait video, with empty and default prompts, produced exactly equal decoded tensors to the original checkpoint. Broader quality and installation validation are separate.
See META/CHECKPOINT-TERMS.txt and META/NOTICE.md for original license terms and required attribution. Users must review applicable terms before downloading through AI2Apps.
- Downloads last month
- 33