Ideogram 4 Detailer Sampling Provider (Umbra)

Some model families sample against two models at once - a main model guided by a second one's predictions - and Ideogram 4 is one of them. Ideogram 4 Detailer Sampling Provider (Umbra) exists so the Umbra Image Detailer's crops use that dual-model guidance and Ideogram 4's own resolution-aware sigmas, instead of falling back to a generic img2img pass that ignores how the model actually wants to sample.

How it works

Wire its single sampling_provider output into the Image Detailer's sampling_provider input and the provider takes over per-crop sampling:

• DualModelGuider - sampling runs through ComfyUI's DualModelGuider, which needs two models: the main model and a model_negative. The second model provides the negative-direction guidance signal, a design that's specific to models trained with that guidance scheme. If this sounds like the "does this model need a negative model?" question - yes, and the node surfaces it as a required input precisely because forgetting it silently changes results.
• Ideogram 4 sigmas - each crop's sigmas come from Ideogram 4's own scheduler, parameterized by mu (default 0.5, range −10 to 10) and std (default 1.75, range 0.1–5). These two shape the noise schedule - mu biases where the schedule sits, std widens or narrows it. They're the model-family-specific knobs you'd otherwise have no way to reach from a stock detailer.
• Your sampler - the sampler_name dropdown (default euler) runs on top of those sigmas.

Inputs: model, model_negative (required - the guidance-negative model), sampler_name, mu, std. Output: sampling_provider → Image Detailer.

How to use it

1. Load both models - your main Ideogram 4 model and the model_negative one.
2. Add this node, wire both in. Start with mu 0.5 / std 1.75.
3. Add Image Detailer (Umbra UI), wire image/model/clip/vae/positive/negative/seed, and connect sampling_provider into the detailer's input.
4. Tune mu/std only if crops look off - they're the model-specific flavor dials. std up makes the schedule noisier/more exploratory; mu shifts the center.

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 listed in the Image Detailer article.

Gotchas

The required model_negative input is the thing most likely to trip you up - leave it unwired and the provider can't build its guider, so the stage fails rather than silently degrading. Don't confuse it with a negative prompt; it's a second model providing the negative-direction signal. And as with all the providers, this replaces only the sampling: detection, SAM masking, and paste-back still come from the Impact Pack machinery underneath, so the usual detector-model-file requirements still apply. Unplug the sampling_provider and the detailer silently reverts to classic sampling - easy to forget mid-experiment.