What makes a model “uncensored”
Limits in image AI come from three places: the training data, a safety fine-tune that teaches the model to avoid some subjects, and a filter around the model that refuses or rewrites prompts. Closed models such as GPT Image, Nano Banana and Seedream have all three, run by their providers.
An uncensored model is an open model released without the last two. It is still shaped by what it was trained on — which is why some open models are great at anime and weak at photos, and the other way round.
Chroma: the general-purpose uncensored model
Chroma is an 8.9-billion-parameter open model based on FLUX.1 [schnell], released under Apache 2.0. Its authors describe it as fully uncensored and trained on about 5 million curated images covering photos, anime and art [1]. On Kenerate it renders up to 1536 px.
Write for it in plain sentences: subject, clothing, setting, light, camera. It follows literal descriptions closely, so say what you want rather than what you don't.
Pony XL: anime and stylised characters
Pony Diffusion V6 XL is an SDXL model trained on about 2.6 million images, split roughly evenly between safe, questionable and explicit ratings, and it reads quality and rating tags in the prompt [2]. Kenerate runs Prefect Pony XL, a model in the Pony family. A tag prompt is built in this order:
- Quality: score_9, score_8_up, score_7_up
- Rating: rating_safe, rating_questionable or rating_explicit
- Source: source_anime or source_cartoon
- Then tags for character, outfit, pose, place and light
Neta Lumina: clean anime illustration
Neta Lumina is an anime model from Neta.art, fine-tuned from Lumina-Image-2.0 on over 13 million anime-style images and released under Apache 2.0 [3]. Its Gemma text encoder understands English, Chinese and Japanese prompts.
Draft uncensored, finish polished
The comparison on this page runs the same prompts through Seedream 5.0 Pro, Chroma, Pony XL and Neta Lumina. A common workflow is to explore ideas on an open model, then re-render SFW versions on Seedream or GPT Image when you need the extra polish.
