ComfyUI Workflow: HiDream O1 Detailer (Two-Pass)
HiDream O1 is not a base text-to-image model. It is a two-pass refinement workflow: feed it an already-rendered image (from any model, FLUX, SDXL, or a previous HiDream generation), and it runs a second HiDream-E1 pass focused on detail recovery, face sharpness, and small-text legibility.
The first pass establishes the base composition. The second pass, gated by the O1 detailer node, only resamples the regions where the detailer detects low confidence: faces, hands, small text, sharp edges. This is faster than a full second generation and avoids changing parts of the image that were already good.
Key benefits
- Two-pass detail recovery on top of any base model, without re-rendering the full image.
- Selective resampling on faces, hands, and small text, where most generation artifacts hide.
- HiDream-E1 was trained on these specific failure modes.
- Compatible with SDXL, FLUX, and HiDream base generations.
- Single workflow JSON, no custom nodes required beyond the HiDream loader.
How it works
- Connect the base generation: pipe the output IMAGE from your existing generation (or load an image from disk) into the first stage of the workflow.
- Set the detailer region: the O1 detailer node auto-detects low-confidence regions. You can constrain it to faces only or run it on the full frame.
- Configure the second pass: denoise 0.2-0.4 keeps the composition intact. Higher denoise lets the detailer change more.
- Queue the prompt. The second pass typically takes 30-50% of the original generation time.
- Compare side-by-side. The detailer pass is most visible on faces and small text.
The detailer is for the case where the base model got 95% of the image right but the eyes are slightly off, the logo text is garbled, or a hand has a missing finger. HiDream O1 fixes those without forcing a full reroll.