Three months of Claude-guided DiT fine-tuning fails, non-technical dev begs Reddit for help
Old_Manufacturer7905 · reddit · 2026-10-15
A non-technical Reddit user spent three months on L40S GPUs trying to fine-tune a real-time video-editing DiT (RV2V) model with Claude handling all the technical work, testing every available model tier. With a small 250-clip dataset, every LoRA run performed worse than the base model. His takeaway: "Claude can't help me train a model" — a textbook case of the limits of fully LLM-delegated deep learning.
More from Multimodal
- New paper: Image Tokenizers Should Be Trained Between Images cuts artifacts, speeds training — ostrisai · 2026-10-15
- Harmonic representations in symbolic music transformers are localized and controllable — kastnerkyle · 2026-10-15
- Open-source ff-tracking hits 407 stars in 6 days as community ports and remixes it — PurzBeats · 2026-10-15
- Purported Gemini 4 builds a shakable New York snow globe entirely from code, zero assets — prasenx · 2026-10-15
- This week in AI film history: Firefly Video launch, Veo 3.1 and more milestones — DavidmComfort · 2026-10-15
- Viral Argon gummy bear video was fake: real Gemini 4 Argon results much better — socialwithaayan · 2026-10-15