topabaem/LTX-2.5-Text-Encoder-4bit-8GB-Demo

LTX-2.5 with a 4-bit text encoder

Text-to-video, image-to-video and video-to-video driven by `topabaem/LTX-2.5-Text-Encoder-4bit-8GB` — the Gemma4-12B text encoder compressed from 26.264 GB to 8.46 GB with no compute-capability floor, paired with the distilled 22B DiT as Q3_K_M GGUF.

Video and audio are generated together; the clips carry their own soundtrack.

What is actually being demonstrated

This Space exercises the packed encoder in a complete LTX-2.5 pipeline. The serving stack also includes pinned ComfyUI and LTX VAE revisions, a third-party Q3_K_M DiT, custom audio-video guide handling, adaptive scheduling and an optional refinement pass. A generated clip is therefore an end-to-end smoke test, not an isolated proof of encoder quality.

Live prompts use the gemma4-raw-intermediate-slots-v1 conditioning contract: an explicit BOS token, a 1024-token left-padded Gemma forward, extraction of the valid-token suffix, and the raw intermediate hidden-state slots returned by the model (whose final slot is already final-normalized). The static sample pairs were generated before this correction. They remain historical comparisons of two checkpoints under the same older preprocessing, but they do not validate the corrected live path; a corrected BF16 oracle comparison is not yet published.

Image- and video-to-video need one thing ComfyUI does not ship: LTXVAddGuide cannot take a combined audio-video latent, because it calls torch.cat on a NestedTensor. The model, sampler and mask plumbing all support it already, so ltx_av_guide.py unwraps the pair, runs the stock node on the video half and re-wraps. The guide maths stays the vendor's.