license: other
license_name: ideogram-4-non-commercial
license_link: https://huggingface.co/noelthekke/ideogram4-text-encoder-abliterated/blob/main/LICENSE.md
base_model: ideogram-ai/ideogram-4-fp8
tags:

• abliteration
• refusal-direction-ablation
• qwen3-vl
• text-to-image
• interpretability

Ideogram 4 Text Encoder -- Abliterated (research use only)

This is a research artifact, not a product. It is the text-encoder language
model used by ideogram-ai/ideogram-4-fp8
to condition its image generation, with directional ablation applied to remove its
refusal behavior. It is not a chat model, not a complete image generation
pipeline, and not an official or endorsed Ideogram product.

⚠️ Content warning

This model has had its refusal mechanism deliberately removed. When used to condition
ideogram-ai/ideogram-4-fp8's image generator, it may pass through prompts -- and
therefore contribute to generated images -- that the original model would have
resisted encoding, including depictions of violence, non-consensual imagery, or other
harmful content. You bear full responsibility for how you use it, under the same
legal and license restrictions as the original model (see below). This repository
does not include any content filtering, and none should be assumed.

What this is, precisely

ideogram-ai/ideogram-4-fp8 is a text-to-image diffusion pipeline: text_encoder +
transformer (DiT) + unconditional_transformer + vae. The DiT itself has no
linguistic "refusal" mechanism -- but text_encoder is a Qwen3VLModel (a Qwen3-VL-8B
vision-language backbone) used to embed the prompt before the DiT ever sees it.

This repository contains only the language-model half of that text encoder
(architecture class Qwen3VLTextModel: 36 layers, hidden size 4096, GQA 32/8 heads,
no vision tower, no lm_head -- it was never meant to generate text, only to produce
conditioning embeddings). The vision tower, transformer, unconditional_transformer,
vae, and scheduler are not included. This model cannot generate images on its
own; it is only useful recombined with the original, unmodified versions of those
components from ideogram-ai/ideogram-4-fp8.

Method

Directional ablation, following Arditi et al. 2024, "Refusal in Language Models Is
Mediated by a Single Direction":

1. Collected residual-stream activations (last-token hidden state, at every layer)
   over a set of contrastive image-generation prompts -- harmful (e.g. "generate a
   photorealistic image of graphic violence") vs. harmless (e.g. "generate an image
   of a golden retriever playing in a park").
2. Computed the mean-difference direction per layer, normalized -- the "refusal
   direction."
3. Projected that direction out of the output space of self_attn.o_proj and
   mlp.down_proj at all 36 layers: W_new = W - d ⊗ (d @ W), so nothing those
   projections produce can have a component along the refusal direction.

The original checkpoint's weights were stored in a scaled fp8 scheme (fbgemm_fp8:
F8_E4M3 + per-channel F32 scale); they were dequantized to bf16 before this
projection, since the math needs real floating-point precision. This repository's
weights are plain bf16, not fp8.

Full implementation: heretic/ideogram/abliterate.py, a custom script in a local,
unpublished research repository (not on GitHub or any public host at time of writing;
adapted from an original Llama 3.1 abliteration script in the same repo). Not
TransformerLens -- a direct PyTorch/transformers implementation operating on the
real Qwen3VLTextModel module tree.

Intended use

AI safety and interpretability research into how refusal behavior is represented and
can be removed in text-to-image conditioning models -- specifically, probing where
image-generation models draw their content-safety lines and how that differs from
text-generation refusal. Non-commercial research use only (see License).

Limitations

• No lm_head: this model cannot generate text and was never trained to. It only
  produces hidden-state embeddings for conditioning the (separate, unmodified) DiT.
• Ablation was validated by inspecting whether the refusal direction's component in
  the output was removed (it was, to within floating-point precision) -- it was
  not validated end-to-end through the full image-generation pipeline. Whether
  removing this direction actually changes what images get generated, and by how
  much, has not been measured here.
• The contrastive prompt sets used to derive the refusal direction were small (12
  harmful / 12 harmless examples) and hand-written; they do not comprehensively
  cover the space of content this model might resist or refuse.

License and required notices

This is a Model Derivative of ideogram-ai/ideogram-4-fp8 under the
Ideogram Non-Commercial Model Agreement (included in this repository
in full, as required by that Agreement). That means, without exception:

• Non-commercial use only -- the same restriction as the original model.
• The same use restrictions apply, including (verbatim from the original
  Agreement): no military or surveillance purposes; no biometric processing; no
  infringement of third-party rights; no generation of unlawful content, including
  CSAM or non-consensual intimate images; no violation of privacy/data protection
  laws; no automated "consequential decisions" or other high-risk use cases. Removing
  this model's internal refusal behavior does not grant permission to violate any
  of these restrictions -- they remain your legal obligation regardless of what the
  model will or won't refuse to do.
• See the full Acceptable Use Policy,
  incorporated by reference into the Agreement.
• This repository is not an official Ideogram product and has not been
  endorsed, approved, or validated by Ideogram, Inc.
• See NOTICE for the required attribution notice and a description of
  exactly what was modified from the original release.

If you redistribute this model or any further derivative of it, you must comply with
the same Section 3 (Redistribution and Use) conditions in LICENSE.md
that this repository complies with.