its impossible with my slow internet connection. The upload would take about 4 hours.. way beyond my time limit...

I might have an idea :-) I am just gonna publish the workflow... So you can built it on your own :-)

Lupus_dH https://civitai.com/articles/17666

The instructions are inside. Sorry for this.. but I just want to play safe here..

whozzat I have to think about some solutions.. but one thing is clear before I am not 100 percent sure what I should do.. I am gonna play safe.. If there is a lawyer: Is sharing a workflow to built the checkpoint ok? Its no checkpoint on its own.. :-) What about LORAS?

https://www.taylorwessing.com/en/insights-and-events/insights/2025/07/update-ai-act The problem is the documentation. I am not creating checkpoints to earn money from it or fame. I am creating my checkpoints for fun. Tell me is it fun to document every step? To make 1000 percent sure that all training images are comply to any copyright rules? (Everyone doing photography knows that this is a complex thing.. even with your own images...) So basically you have to tell the source from every image.. easy? No.. Because you have to write it in a list.. Training a full checkpoint requires a lot and I mean A LOT of images...

This is and was the biggest issue with Colossus Project Flux.. Documenting everything afterwards will be a colossal work (if any EU guy reads this: I make some efforts to get everything complient until 2027.. for sure just give me that time :-)

There are more questions:

What if you want to merge your checkpoint with another one.?. Like I did with the de-distilled model.. have you to ducument internal merges as well? What about LORAS? And don't get me started with the guidelines... which aren't fully clear yet..

Everything what the EU does ends up in such a mess.. And no the law doesn't to exemptions for non profit hobbyiest like me... Actually it would be much better to go fully open source and commercial.. So I only have to share everything.. including all technical data.. (So that everyone can just copy it.

And well the fines can be heftly.. we talking about millions here..

dispose608 The only practical way is the following: First wait what the EU is doing. The law is that bad that it will hinder further developments in the industry. This will lead sooner or later to changes in the law. Second and this is what I also looking into it is too look for stuff they may have forget about. eg. legal loopholes.

For example: Are comfy workflows even in this law? (If there is a lawyer here.. tell me if I am right)

Let's say I am gonna write an tutorial where I can show you some stuff like block merging Flux 1dev into Flux Krea. So this will get educational. even scientific. You may be able to craft your own checkpoint afterwards out of example Colossus V12 (which was released before 2nd August and Krea... ;-). Btw private checkpoints are still allowed... So you can create them on your own computer.

And outside of the EU you can still share it. But don't let me know or mention me if my tutorial inspired you to share your own checkpoint ;-)

So any lawyer here could say if I am on the right track..

The is what Grok4 say about it:

Yes, based on the current provisions of the EU AI Act (which entered into force in August 2024 and is being phased in through 2026), sharing ComfyUI workflows and tutorials about creating AI checkpoints appears to be allowed, with some important caveats around how they're used or distributed. I'll break this down step by step, focusing on why this is the case and any potential limitations.

1. Understanding Key Terms in the EU AI Act

AI Systems: The Act defines an AI system as software that can generate outputs (e.g., images, text, predictions) influencing the environment. ComfyUI itself is an open-source graphical user interface (GUI) for running AI models like Stable Diffusion, but workflows are essentially configuration files (JSON-based node graphs) that define processes for using existing models—they don't qualify as standalone AI systems or models.

General-Purpose AI (GPAI) Models and Generative AI: These include foundation models like those used in Stable Diffusion. Generative AI (e.g., for image creation) isn't classified as "high-risk" by default but must comply with transparency rules, such as labeling outputs as AI-generated and disclosing if copyrighted data was used in training.d4c984 Providers of generative AI tools must ensure compliance with EU copyright laws, but this applies more to the models themselves than to workflows or tutorials.

Checkpoints: These are fine-tuned model weights (e.g., from training on specific data). Creating and sharing checkpoints could involve GPAI models, but if done under open-source licenses and not for prohibited or high-risk uses, they're generally exempt from heavy regulation.

2. Are ComfyUI Workflows Restricted?

Workflows in ComfyUI are not AI models or checkpoints—they're reusable setups for processing inputs through existing models (e.g., to generate images). They're more akin to code scripts or recipes than regulated AI components.

The Act doesn't explicitly restrict sharing such workflows. Since they're open-source and don't inherently fall under prohibited AI practices (e.g., they don't deploy subliminal techniques or enable social scoring), sharing them is permissible.8ff524f5ae7d Even if a workflow could be used to create a checkpoint, the workflow itself isn't the checkpoint.

Potential Caveats:

If a workflow is designed specifically for high-risk applications (e.g., biometric identification in public spaces or emotion recognition at work), it could trigger obligations if deployed commercially.

For generative uses, any outputs created via the workflow should be marked as AI-generated to comply with transparency rules.fb024f

If the workflow involves training on copyrighted data without permission, that could violate EU copyright laws separately from the AI Act.43071c

3. Are Tutorials on Creating Checkpoints Allowed?

Tutorials are educational content, not AI systems, so they're not directly regulated by the Act. Sharing knowledge about creating checkpoints (e.g., fine-tuning models) is generally fine, especially in research, open-source, or non-commercial contexts, which are exempt.393dee

This aligns with the Act's support for innovation and open-source development. Communities like those around Stable Diffusion continue to share tutorials freely, and there's no indication of restrictions on this.b1035f7f4ec7

Potential Caveats:

If the tutorial promotes prohibited AI practices (e.g., creating deepfakes for harmful manipulation), it could indirectly fall under scrutiny, though the Act targets the AI systems themselves, not the teaching materials.

4. Overall Assessment and Recommendations

You're correct: Workflows aren't checkpoints, so sharing them (and tutorials) doesn't trigger the same regulations as distributing full AI models. The open-source nature of ComfyUI provides strong protections under the Act's exemptions

Thanks for sharing this. Yes I am working on Krea. But more or less privately since yesterday. I might share the results here and maybe the workflow to scientifically educate you maybe with a workflow that contains some examples how to block merge a Flux 1 checkpoint like Colossus with Flux Krea. You also then could hit run to get built your own checkpoint locally on your computer. But again only for educational reasons ;-) . God beware someone would share the "inspired" results here.. The EU would not be happy 😁. I mean if you living outside the EU it should be fine..