Preserve pinned official Ideogram 4 FP8 release, components, original terms and reference code
8ebf922 verified | # Safety | |
| Ideogram is committed to the **responsible development and deployment of | |
| open-source models**. Image generation models can be powerful general-purpose | |
| tools, and we believe the best way to release them openly is to pair the | |
| weights with concrete safeguards throughout the pipeline, from development | |
| through use. These safeguards include, but are not limited to, training data, | |
| model behavior, and runtime filtering. | |
| This document describes the mitigations applied to Ideogram 4 and the | |
| expectations we have for anyone redistributing or serving the model. | |
| ## Pre-training mitigations | |
| Before any training begins, we run our image dataset through filters that are | |
| designed to remove content across multiple NSFW categories. The | |
| filtered dataset is then used as the input distribution for pre-training. | |
| This measure is designed to prevent the base model from reproducing the | |
| filtered categories of content. | |
| ## Post-training mitigations | |
| After pre-training, we apply additional post-training procedures that are | |
| designed to further reduce the probability of the model generating NSFW | |
| content, including for prompts that explicitly request it. | |
| ## Inference-time filters | |
| The inference code in this repository ships with two runtime safety filters, | |
| both powered by [Hive](https://thehive.ai/): | |
| - **Prompt moderation** (`--hive-text-key` / | |
| `HIVE_TEXT_MODERATION_KEY`) — every prompt is screened by Hive's Text | |
| Moderation API before it reaches the model. Prompts flagged for unsafe | |
| categories are rejected and no image is generated. | |
| - **Output moderation** (`--hive-visual-key` / | |
| `HIVE_VISUAL_MODERATION_KEY`) — every generated image is screened by | |
| Hive's Visual Content Moderation API. Images flagged for unsafe categories | |
| are rejected before being returned to the caller. | |
| See `run_inference.py` for the exact integration points. | |
| > **We expect anyone serving or redistributing the model to use these | |
| > inference-time NSFW filters (or equivalent or stronger filters).** | |
| > Running the model without prompt and output moderation configured disables | |
| > runtime safety screening entirely and is not a supported deployment | |
| > configuration. If you redistribute or host this model, you are responsible | |
| > for keeping equivalent (or stronger) filters in place. | |
| ## Reporting violations and ongoing monitoring | |
| We may monitor Ideogram 4 public deployments and outputs from time to time, | |
| but we cannot catch everything on our own. If you observe violations of our | |
| usage policies — content the model should not have produced, deployments | |
| that have removed, overridden, or failed to deploy the required safety | |
| mitigations described above, or other misuse — please email us at | |
| **safety@ideogram.ai** with as much detail as you can share (prompts, image | |
| hashes, URLs, screenshots, deployment context). Reports help us improve | |
| both the model and the surrounding guardrails. | |