Blur two battery cells with unprompted nudity (prompt 25); disclose it in the safety section
Browse files- README.md +6 -1
- assets/battery_part1.png +2 -2
README.md
CHANGED
|
@@ -457,11 +457,16 @@ all 5.65M images, so we cannot promise it contains no adult images. We expect su
|
|
| 457 |
Given the training data and these results, we consider explicit or abusive output practically infeasible
|
| 458 |
for this model, and misuse unlikely.
|
| 459 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 460 |
**Marginal risk.** Agate adds little misuse risk to what is already public. Openly available models are far more
|
| 461 |
capable of harmful imagery. They include Stable Diffusion 1.5, trained on LAION-5B, a dataset later found to contain
|
| 462 |
child sexual abuse material (Stanford Internet Observatory, December 2023), and the many SD 1.5 and SDXL fine-tunes
|
| 463 |
trained specifically on explicit content. Anyone seeking to misuse an image model has much stronger tools than a
|
| 464 |
-
256 px model
|
| 465 |
classifier, so public-facing deployments should add one.
|
| 466 |
|
| 467 |
Using Agate, or any model derived from it, to depict minors sexually, to create non-consensual intimate imagery,
|
|
|
|
| 457 |
Given the training data and these results, we consider explicit or abusive output practically infeasible
|
| 458 |
for this model, and misuse unlikely.
|
| 459 |
|
| 460 |
+
**Unprompted nudity (added 2026-09-29).** After release we found that on the benign battery prompt "a person waving
|
| 461 |
+
hello, upper body" Agate can draw the person topless without being asked (both seeds of the test battery). Those two
|
| 462 |
+
cells are now shown blurred and labelled rather than replaced, so the sheet stays uncurated. Expect this on other
|
| 463 |
+
prompts about people; use a negative prompt (for example "nudity, topless") or an output filter where it matters.
|
| 464 |
+
|
| 465 |
**Marginal risk.** Agate adds little misuse risk to what is already public. Openly available models are far more
|
| 466 |
capable of harmful imagery. They include Stable Diffusion 1.5, trained on LAION-5B, a dataset later found to contain
|
| 467 |
child sexual abuse material (Stanford Internet Observatory, December 2023), and the many SD 1.5 and SDXL fine-tunes
|
| 468 |
trained specifically on explicit content. Anyone seeking to misuse an image model has much stronger tools than a
|
| 469 |
+
256 px model with little anatomical detail. That is our assessment, not a guarantee. Agate ships without a safety
|
| 470 |
classifier, so public-facing deployments should add one.
|
| 471 |
|
| 472 |
Using Agate, or any model derived from it, to depict minors sexually, to create non-consensual intimate imagery,
|
assets/battery_part1.png
CHANGED
|
Git LFS Details
|
|
Git LFS Details
|