A manga software maker withdrew its image generator in days

The process wall

On 29 November 2022, Celsys announced that Clip Studio Paint's version 1.13.0, due 6 December, would ship an experimental 'Image Generator palette' built on the Stable Diffusion model. Three days later, on 2 December, a second notice reversed it: 'After the initial announcement, we received a lot of feedback from the community and will no longer implement the image generator palette.' For a tool used daily by manga and illustration artists, the process wall was not the feature itself but its removal before a single release shipped it, a rare case of a generative feature cancelled between announcement and release date.

What the documents show

The November announcement is specific about method and scope: the palette would use Stable Diffusion 'as-is,' Celsys stating it 'is not providing any user data to train the model' and 'has not collected, and will not collect or use any form of data without clear consent from the user.' It also commits to a sunset, the feature 'will be discontinued in 1.0 after the release of Clip Studio Paint Ver. 2,' alongside plans for an anonymous survey, a public archive of prompts and outputs, and research into 'authentication features for artists to prove ownership of their artwork.' The December withdrawal notice lists the objections Celsys says it 'took to heart' verbatim, including that 'current image generation AI exploits other artists' intellectual property,' that 'using artist's work that is not opted-in to a data set is ethically unacceptable,' and that unresolved legal or moral rights risk was 'unacceptable' in an app used daily. Celsys states plainly: 'we had neglected to take the necessary considerations.'

Authorship and limitation

The mechanism named in the announcement is a stock, unmodified base model rather than a Clip Studio-trained system: Celsys was reusing Stable Diffusion 'as-is,' not fine-tuning it on Clip Studio's own users' work, which is why it could truthfully say no user art trained the feature while objections still focused on the model's own training data, gathered elsewhere and outside any Clip Studio user's consent. The withdrawal notice does not resolve that upstream question; it removes Celsys's feature rather than adjudicating what Stable Diffusion was trained on. The company frames this as a listening failure on its part, not a finding about infringement by anyone.

What survives

Editorially, what survives for an illustrator is the list of stated concerns itself, a rare instance of a vendor publishing the objections that changed its plans rather than a generic statement of regret. Worth noting too is what the company did not reverse: the planned authentication and ownership-proof research is not described as cancelled, only the generator.

• Compare the seven listed objections against a current tool's own generative feature before adopting it.
• Check whether a still-planned 'prove ownership' authentication feature from Celsys was ever released.
• Read the November announcement's technical details against the December apology to see what was walked back and what was left unaddressed.

Read together, the two notices are an unusually transparent record of a specific product decision being made, contested, and reversed inside a single week, with the company's own reasoning published rather than paraphrased.