VALYDA AI PROTOCOL · EU AI Act · Broadcast (ARD, ZDF, RTL, BBC and others)

If you deliver to a broadcaster, the game is different from posting on social. Delivery contracts - ARD, ZDF, RTL, BBC, Netflix - don't just want a label; they want a full production record, right down to the prompt. In plain text. Per clip. And since Article 50 of the EU AI Act applied on 2 August 2026, that transparency duty covers every broadcaster, public, private, or international. Chasing all that metadata by hand, per shot, is how delivery weeks die. The Broadcast node from the Valyda-Protocol pack does it automatically, every render.

It's the sibling of the Creator node, but it asks for more and gives you more. Four required fields - Production, Scene, Reason_for_AI_Use, Disclosure_Type - plus timecodes, source footage, credits, and a broadcaster dropdown with 84 entries covering the whole ARD network (BR, NDR, WDR…), ZDF, ORF, SRF, Deutsche Welle, and an "andere / other" option with a name field for anyone else.

The trick that makes it special

Pick ARD Degeto or ARD Das Erste and the node doesn't just write the protocol - it also generates the Erklärungsbogen KI, the AI declaration sheet straight out of the Degeto delivery contract, already filled in. That's normally a form you'd be filling by hand at 2am. Other broadcasters have their own paperwork, so for them the main protocol is the basis; the author notes that explicitly rather than pretending one form fits everyone.

Same honest mechanics as the rest of the pack: it reads the executed workflow, not the canvas JSON, then hashes every model file it finds (SHA-256, cached), pulls seed, sampler, steps, resolution, frame rate, and logs the prompt. Whatever you type is marked as your input; whatever's missing says so. Nothing invented.

Inputs worth your attention

• Store_Prompt - defaults to show prompt here, the opposite of the Creator node. That's deliberate: German public broadcasters require the prompt in plain text, and it cannot be recovered afterwards. Hiding it would be the wrong default for this node.
• Prompt_Assignment - some workflows don't reveal which text block is the prompt and which is the negative. automatic leaves both in the document unlabelled; "main text is the prompt" records your statement that the long text is the prompt. It's marked as your call, not the tool's claim.
• Timecode_In / Timecode_Out - 00:21:48:11 style; dots or commas get turned into colons. Leave In empty until the edit is locked - it also sets the ordering in the project record.
• Broadcaster_Name - for "Redaktion Fiktion" style precision; the document shows it as one line, ARD Degeto · Redaktion Fiktion.
• Output_File - the saved file's full path, so the record's checksum binds to that exact deliverable.

Connect video, images, or audio (one is enough). Outputs are two strings - pdf and akte - the file paths, written to ComfyUI/output/valyda/<production>/ alongside the JSON record.

Install and gotchas

ComfyUI Manager: search "Valyda-Protocol". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/RebellionPictures/Valyda-Protocol
pip install -r Valyda-Protocol/requirements.txt

Dependencies are just reportlab and Pillow - no models, no GPU, no network, no API key. The same install covers all three nodes in the pack, so if you have Creator installed, this one's already there.

The usual traps: a muted node writes no record and can't say so; Production must be spelled identically on every clip or records scatter into separate folders; if your ComfyUI build errors on the in-node preview, flip Preview_In_Node to off (the record is unaffected). And remember the sheet is a self-declaration, not a proof of authenticity - that's stated on every page, and it's what keeps it defensible.