Hugging Face releases an end-to-end CogVideoX-5b 3D-style fine-tuning guide with LoRA weights
RisingSayak · x · 2026-11-17
An end-to-end fine-tuning guide for 3D-styled video (3DGS dissolve) is now on Hugging Face, built on CogVideoX-5b.
- Experimental checkpoint finetrainers/3dgs-v0 fine-tuned from THUDM/CogVideoX-5b on the finetrainers/3dgs-dissolve dataset, with a LoRA variant provided
- Trigger word 3Ddissolve generates videos of 3D characters/objects dissolving into swirling red sparks
- Training code is open source (a-r-r-o-w/finetrainers); inference uses diffusers in bfloat16
- Authors note it's experimental with known poor generalization
HF team members also referenced earlier similar fine-tuning guides for 3D models.
More from Multimodal
- AI-Generated Spanish Sci-Fi Trailer 'Enki Chronicles' Gets Director's Cut — Excellent-Emphasis25 · 2026-11-17
- Krea2TextEncoder hands-on: reference-based gens with zero sampling overhead and minimal prompting — ThetaCursed · 2026-11-17
- Koyal pivots from NeurIPS audio-to-video paper to full-blown AI film company — DavidmComfort · 2026-11-17
- PAN2 lands native ComfyUI support, positioned as a world model for VLA experiments — Altruistic_Heat_9531 · 2026-11-17
- Dreamina Marketing Studio turns a product photo into an ad video in a few clicks — AIwithGhotai · 2026-11-17
- Midjourney v8.2 tip: low --stylize plus --raw makes prompts actually stick — michaelrabone · 2026-11-17