Ideogram 4 Debanner Patch

Ideogram 4 is the model that finally made text-in-image look boring-good: legible multi-line text, exact placement, the best character recall of any open checkpoint. It is also, every once in a while, the model that bolts an unwanted banner onto its own output. That artifact is what this node exists to remove, at generation time, without a second pass.

It's an experimental model patch, and it does exactly one thing. You feed it your Ideogram 4 model, set a strength, and get back a patched clone you sample with. There's no UI magic and no separate denoise - the correction happens inside the transformer as it runs.

How it works

Under the hood the patch loads a small correction file of learned "direction" vectors and rewires the last four transformer blocks (layers 25 through 28) so each one adds a scaled correction to its forward pass. A direction here is just that: a vector that nudges the model's activations away from the artifact, rather than a regenerated image. A wrapper also stamps the grid size and strength into the transformer options so the correction knows what geometry it's working on.

One detail the author is explicit about, and it's the part people get wrong: use this only for the first split-sigma segment, then continue with the original conditional model. The artifact forms early in sampling (high sigma), so the correction is most valuable there. Run the patch for the whole schedule and you risk over-correcting detail that was fine to begin with. The intended pattern is patch → sample the first segment of your schedule → switch back to the clean model for the rest.

The inputs that matter

• model - your Ideogram 4 MODEL. Nothing else will do: feed it a different architecture and the node raises ValueError because it needs a real Ideogram4Transformer2DModel.
• strength - how hard the correction pushes, 0 to 2, default 0.6. At 0 the node is a literal no-op and returns your model untouched, which is handy for A/B testing whether the patch is even helping.

Output is a single patched model.

Install

It ships inside ComfyUI-UtilsCollection, so install the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

then restart ComfyUI (or search "UtilsCollection" in ComfyUI Manager). You'll also need an Ideogram 4 checkpoint in models/checkpoints - the fp8 or nf4 release from HuggingFace, and yes, that means the non-commercial license applies to whatever you make with it.

Where people get burned

• The correction file isn't in the repo. The node reads models/ideogram4_correction_v1.safetensors from inside the pack folder, and as of the current main branch that file isn't actually tracked on GitHub. A fresh clone can hit a FileNotFoundError naming that exact path. This is genuinely experimental work-in-progress - if the bundle is missing for you, that's why, and it's not the model's fault.
• Wrong model connected → immediate ValueError, not a bad image.
• It's not a fix-everything switch. If you're not seeing a banner artifact, this patch costs you nothing but also fixes nothing. It's a scalpel for one specific failure.

Reach for it when Ideogram 4 does its thing and then tacks a banner on top. Otherwise, skip it - there are more fun patches to spend time on.