Ideogram 4 censorship bypass: Two sigma-hacks using LCM sampler

A Reddit user has published two working methods to disable the content censorship in Ideogram 4, the AI image generation model. The first method involves shifting only the first sigma step by +0.005 (or +0.01 in some cases), which slightly moves the starting trajectory away from what the model expects while keeping all other sigma steps unchanged. The second, preferred method increases the initial noise by 2x (or 3x when needed), producing a similar effect. Both methods require an LCM sampler to correct the trajectory after the initial deviation—other samplers will break the image. The user also outlines the necessary ComfyUI nodes: Noise Math (More Math addon), SamplerLCMCustom (Extra Samplers), Custom Sigmas (KJ Nodes), and Sigmas2 Mult (RES4LYF).

It's important to note that these hacks only remove the gray censorship square, not all behavior issues. Ideogram 4 still struggles with very short prompts; one-word prompts may no longer trigger censorship but the model may fail to follow them accurately. Prompts composed of several sentences work much better even without JSON formatting. The user recommends a workflow of two sequential samplers: the first runs the high-sigma stage with the modified noise/sigma setup using LCM, then passes output to a second sampler for remaining steps with a normal sigma range. The split point in SplitSigmas can be adjusted between 1 and 3 for best results. The user encourages continued exploration for more robust solutions.