Ostris tests HiDream-O1 LoRA on AI Toolkit — calls it industry-changing
No VAE, no text encoder, train fast — Ostris says HiDream-O1 is a paradigm shift for diffusion.
Deep Dive
A user is running their first test of training a HiDream-O1 LoRA on AI Toolkit, describing it as the coolest model they've ever seen. It uses efficient pixel space, has no VAE and no text encoder, trains super fast, and is called
Key Points
- HiDream-O1 removes the VAE (variational autoencoder) entirely, simplifying the generation pipeline.
- No separate text encoder is needed, cutting compute and memory requirements.
- Training speeds are reported as 'super fast,' suggesting a major efficiency leap over current diffusion models.
Why It Matters
If validated, HiDream-O1 could make high-quality image generation LoRAs trainable on consumer GPUs, democratizing fine-tuning.