HiDream O1 Dev Smoothing

Grid artifacts. If you've run HiDream O1 Dev and squinted at a faint tiled pattern across the image, you've met the thing this node exists to fix. The model works on a patch grid, and distilled Dev checkpoints have a habit of showing the seams between patches - the release-thread reports on Dev FP8 mention "bad grid noise" as a known sore spot. HiDream O1 Dev Smoothing is the author's answer: it runs a few extra shifted predictions in the last denoise steps and blends them back in so the seams stop being visible.

What it does

It slots in between the model loader (or LoRA) and the sampler, and returns the same HIDREAM_O1_MODEL handle - just with a smoothing recipe attached. The node gates itself to Dev and Dev-2604 model folders; wire it after the Full model and it raises a clear error. That's a feature: Full is the undistilled model and doesn't need it.

How it works

HiDream O1's diffusion runs over a fixed patch grid. Each patch is predicted independently-ish, and at the seams the predictions don't line up, which reads as faint grid lines or repeated texture. During the last steps denoise steps, this node takes the latent, shifts the patch grid, runs the model again at that offset, shifts back, and blends the two predictions. Misaligned predictions average out the seam.

The controls:

• steps (default 4) - how many final denoise steps get the extra forwards. 0 disables smoothing entirely. Keep it small; each smoothed step costs an extra model forward.
• strength (default 0.5) - how hard the shifted prediction gets blended in.
• schedule - constant, linear, cosine, front_loaded, late. How the blend strength ramps across those final steps. constant is the sane default.
• shift_mode - rotate cycles through several patch offsets across steps; static uses one; all uses all of them every step (slowest, most aggressive).
• adaptive_threshold (default 0) - above 0, the node estimates seam intensity each step and skips smoothing when it's below the value. Nice safety valve; 0 just always smooths.
• multiscale - adds a smaller patch-grid offset on top. Only tick it if rotate alone leaves seams.
• cfg_aware - also smooths the unconditional CFG branch when CFG is active. Costs extra forwards; on Dev, CFG is ignored anyway, so leave it off.

Wiring it

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

When to actually reach for it

• You're on Dev or Dev-2604 and see the grid. Turn steps on (default 4 already is) and see if the seams go.
• Seams persist → raise strength toward 0.7–0.8, try all shift mode, or tick multiscale.
• Seams are mild or only in some regions → set adaptive_threshold around 0.05–0.1 and let the node skip where it doesn't matter. Saves the forwards.

If you're on the Full model, don't add this node at all - it'll error, and even if it didn't, you'd be paying extra forwards for nothing. The README's own guidance is that Dev is the variant that benefits; Full's artifacts (when it has them) are a different animal and usually show up with reference-image editing, not as patch seams.