DJZ Zero EDIT

The DJZ Zero EDIT node is a prompt autopilot for image-editing models. Give it two numbers and it hands you a complete, sentence-length edit instruction - "add a small knitted wizard hat on the subject's head, positioned naturally between the ears and angled slightly" - ready to feed straight into a FLUX.1 Kontext, Nano Banana, or Qwen-Image-Edit workflow.

Why this exists

The current crop of instruction-following editors are prompt-driven in a way SD 1.5 never was. You don't tag an edit, you describe it: "change only the blue sofa to a vintage brown leather chesterfield. Keep the pillows and lighting unchanged." The way prompt engineering has shifted, these models are LLM-encoded and want narrative instructions - a well-structured sentence beats comma soup.

That's great when you're deliberately editing one image. It's a pain when you're exploring. For a batch of variations - testing a look, generating assets, A/B-ing phrasings - hand-writing every prompt is the bottleneck, and the "random" ones you like? Good luck recreating them. That's the gap this node fills. It's deterministic procedural generation: the same seed plus the same prompt_index produces the same edit prompt, on any machine, forever. The pack calls this ZeroBytes "position-is-seed," a sibling of the older DJZ-ZeroPrompt idea, reworked for edit instructions.

How it works

It's almost embarrassingly simple under the hood, and that's a feature. The node hashes your coordinates with xxhash32 - edit_hash(seed, prompt_index, 0) picks a sentence template, then edit_hash(seed, prompt_index, 1..n) picks one phrase from each vocabulary pool. A modulo maps each hash into its pool, and Python's template.format() assembles the paragraph. No iteration, no sampling, O(1) every time.

The "4.4 billion combinations" number isn't marketing, it's just arithmetic: 7 templates × the pool sizes in the default profile. Little-endian packing plus xxhash's cross-platform determinism is what makes the same seed reproduce on a different machine - the whole point of the thing.

The inputs that matter

Three of them, plus two you'll rarely touch:

• profile - a dropdown of vocabulary packs from profiles/edit/. Seven ship with the pack, and they get weird: the default, plus ones themed after artists (Araki, Nihei) and a dragonsheaven profile.
• seed - the "world seed." Change it to hop to a different universe of edits.
• prompt_index - your position in that universe. Each index is a different edit; consecutive indexes give related but distinct instructions.

prefix and suffix just prepend and append raw text. "Raw" is the operative word - no spaces are inserted for you, so a carelessly written prefix can glue itself onto the first word of the prompt.

What comes out

One edit_prompt string, a single paragraph. Wire it into whatever prompt input your editor node takes - typically the text-encode node feeding Kontext, Nano Banana, or similar. Nothing else leaves this node.

Install

Two-minute job, and the rare ComfyUI install with no model downloads and no GPU requirement - this is pure string math.

cd ComfyUI/custom_nodes
git clone https://github.com/MushroomFleet/ComfyUI-ZeroEDIT-nodes
pip install xxhash

Then restart ComfyUI. Easier still: ComfyUI Manager → search "ZeroEDIT" → Install → Restart. xxhash is the entire dependency list, and Manager pulls it in for you.

Where people trip up

• The xxhash ImportError at startup is just the missing pip install. Fix it, restart.
• A broken profile doesn't error loudly. If the JSON has a typo, the node silently outputs [Error loading profile '...': ...] as your prompt, which then flows straight into your editor. Preview before you queue.
• Your custom profile isn't in the dropdown. The list is built when ComfyUI loads the node, scanned from profiles/edit/. Drop a new JSON file in and restart before expecting it to appear.
• "4.4 billion" is technically true and slightly silly. With only 7 templates and modest pools, big chunks of that space are variations on the same instruction. It's a deep well, not an infinite one of genuinely distinct ideas.

If you only ever hand-write your edits, you don't need this. If you're automating or batch-exploring, it's the rare prompt node that keeps every good idea reproducible. Full disclosure: it's a small, quietly distributed pack - you won't find much community chatter about it yet. Judge it on the workflow, not the following.