HiDream O1 Detailer Sampling Provider (Umbra)

HiDream's models are famously heavy, and the O1 generation (the pixel-space family HiDream.ai opened under MIT in 2026, per the KB's panel) samples in a way that needs two things stock detailers won't give you: a noise-scale patched model and an LCM-style sampler with its own noise schedule. HiDream O1 Detailer Sampling Provider (Umbra) packages both into a single sampling_provider for the Image Detailer, so each detail crop gets the sampling HiDream O1 actually wants instead of a generic pass.

How it works

When you wire this into the Image Detailer's sampling_provider input, the provider takes over per-crop sampling:

• ModelNoiseScale patch - the noise_scale input (default 7.5, range 0–64) patches the model's noise scale before sampling. This is the HiDream-specific adjustment that aligns the model's internal noise handling with your sampling setup; if crops come back undercooked or overcooked, this is the first dial.
• CFG guider - unlike Flux-style no-CFG sampling, HiDream O1 keeps a classic CFGGuider with positive, negative, and CFG (the CFG value comes from the detail stage's crop context, so set it on the stage).
• LCM sampler with noise clipping - sampling runs through ComfyUI's LCM sampler, and the three noise inputs (s_noise, s_noise_end, both default 1.0, range 0–64) plus noise_clip_std (default 2.5) control the noise schedule and its standard-deviation clamp. These are the knobs that tame the LCM-style noise behavior on a model that's sensitive to it.
• Scheduler - the scheduler dropdown (9 options, default normal) builds the sigmas for each crop at the stage's step count and denoise.

Inputs: model, noise_scale, scheduler, s_noise, s_noise_end, noise_clip_std. Output: sampling_provider → Image Detailer.

How to use it

1. Load your HiDream O1 model (and CLIP/VAE for the detailer).
2. Add this node, wire the model in. Start with the defaults - 7.5 noise scale, s_noise 1.0/1.0, clip 2.5 - and only adjust once you see results.
3. Add Image Detailer (Umbra UI), wire everything including this node's sampling_provider into its sampling_provider input.
4. Run and inspect. Noise-parameter tuning is iterative: if crops look too smooth/latent-y, raise noise_scale or s_noise; if they look noisy or scrambled, lower them or tighten noise_clip_std.

Installing it

Part of Umbra-Nodes, the ComfyUI companion pack for Umbra Studio (NocturneLabs' open-source local AI suite). ComfyUI Manager → search "Umbra Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nocturne-Ai-Labs/Umbra-Nodes

Restart ComfyUI. The detailer still requires Impact Pack + Impact Subpack and the detector/SAM model files from the Image Detailer article.

Gotchas

These noise parameters are real and they matter, but they're also the fiddliest inputs in the whole pack - expect to tune them per model version. Start at defaults, change one variable at a time, and write down what worked. And remember HiDream O1 is a pixel-space model, which is exactly why it needs this non-standard path; if you force it through a stock detailer pass you'll get the "works but clearly not right" look that makes people blame the model. The provider also assumes the stage-level CFG is set sanely - it feeds the guider directly.