HiDream O1 LoRA

Every LoRA loader in ComfyUI looks the same: pick a file, set a strength, wire it between the model loader and the sampler. This one is that shape too - with one difference that matters: it will happily load any safetensors you point it at, but a LoRA trained on SDXL or Flux will not do what you think it will on HiDream O1. The architecture is different enough that the weights are meaningless. If the LoRA isn't an O1 LoRA, you're just injecting noise.

What it does

HiDream O1 LoRA sits between the HiDream O1 Model Loader and the HiDream O1 Sampler (optionally before HiDream O1 Dev Smoothing). It takes the model handle, applies a LoRA state dict to it, and returns the patched handle.

• lora_name - dropdown that reads from ComfyUI/models/loras/, including files inside symlinked folders. Shows "None" when there are no LoRAs to pick.
• strength - the dial. Default 1.0, range −10.0 to 10.0, and 0 disables the LoRA entirely (it returns the model untouched). Negative strengths invert the direction of the learned change, which is occasionally useful for style. Most of the time you'll be between 0.4 and 1.2.

Mechanically it's a straightforward merge: the node loads the state dict once, caches it per file, and applies it to the transformer weights before sampling. Nothing exotic, which is the point - it's the same mental model as a Flux or SDXL LoRA loader, minus the CLIP-adjacent options nobody uses anymore.

The reality check

HiDream O1's LoRA ecosystem is essentially empty. The KB tracks this: no community fine-tunes, no LoRA scene, MIT license notwithstanding. So in practice the LoRAs you'll load through this node are the ones you trained with the pack's HiDream O1 LoRA Trainer, which saves into models/loras/<output_name>/. That's the loop: train with the trainer, pick up the result here, dial strength, sample.

Wiring it

HiDream O1 Model Loader → HiDream O1 LoRA → HiDream O1 Sampler

The output is still a HIDREAM_O1_MODEL handle, so you can stack the Dev Smoothing node after the LoRA and before the sampler:

HiDream O1 Model Loader → HiDream O1 LoRA → HiDream O1 Dev Smoothing → HiDream O1 Sampler

What to watch for

• Lock-in. A HiDream O1 LoRA won't run on Flux, Z-Image, or anything else, and vice versa. If your existing LoRA collection is full of Flux character LoRAs, leave this node with "None" selected.
• A LoRA that overbakes the image usually isn't a loader problem - retrain or drop strength to 0.5–0.7. The pack's training doc is blunt about this: sample several strengths before blaming the model.
• Strength 0 is a bypass, not a "no-op." It's the sanctioned way to keep the node in the graph without it doing anything, which beats deleting and rewiring when you're toggling.

If you're using this pack, you're probably here for O1's editing rather than its LoRA ecosystem. That's the right read - the node is ready for a LoRA scene that hasn't shown up yet, and it's the exact node you'll want the day your trainer finishes.