Ostris/AI-Toolkit now supports HiDream O1 training with token-only input
No text embeddings needed — tokens go straight into the model via Ostris toolkit.
Deep Dive
Ostris's GitHub repo for HiDream-O1-Image includes a note: disable caching text embeddings because "There are not text embeddings. Tokens go directly in." ComfyUI checkpoint versions are available, and a test ComfyUI workflow exists, but as of the post, there is no official workflow template.
Key Points
- Ostris AI-Toolkit now supports HiDream O1 training with direct token input, skipping text embeddings
- Disabling cached text embeddings reduces memory overhead, tokens feed straight into the model
- ComfyUI checkpoints and a test workflow are available, but no official template yet
Why It Matters
Eliminating text embeddings simplifies fine-tuning pipelines and could lower hardware requirements for custom image model training.