Super Input

Most node packs in 2026 are one clever node plus a README. This one is a small application. HenjiccSuperLoader - Super Input, Henjicc/Media - is the entry point, and if you only learn one node from the pack, learn this one, because the other nine all take its output.

The idea: instead of six LoadImage nodes and a text box, you get one panel. Drop, paste, upload or pick from a local library; reorder the strip; type @ in the prompt to insert a reference; and the node emits a single HENJICC_MEDIA bundle plus the compiled prompt, with the reference tags already numbered correctly.

Under the hood it's a JSON widget

The single input, media, is a multiline STRING whose default value is:

{"version":1,"profile":"general","items":[],"prompt":[]}

The panel is a frontend that writes into that string. Items carry a stable id plus a path; prompt references are stored as {ref: <id>, track: "main"|"audio"}, not as text - which is why dragging an item to a new position doesn't repoint your references. Arrows or Alt+arrow reorder the strip and the references follow the asset, not the slot.

Two hidden inputs (prompt and unique_id) let the node walk the real graph. On submit, the pack's on_prompt handler reads what the media output is actually connected to and resolves the profile: a Qwen text encoder downstream → Qwen labels, a Fibo 2.0 encoder → <image_1> style, a MiniMax adapter → <Picture 1> / <Video 1> / <Audio 1>. It also rewrites the compact UI output slots into stable indices after validating the whole mapping, so adding outputs to this node in a later version doesn't break workflows saved against the old one. That's a nicer bit of engineering than most packs bother with.

The same stamping trick makes caching honest: the fingerprint is a SHA-256 over the state and the bytes of every asset, so replacing a file invalidates the cache even though the JSON barely changed.

The five outputs

• media (HENJICC_MEDIA) - the bundle that feeds Reference Image Processing, MiniMax Media Conditioning, H3 Role Inputs, and both prompt assistants. It is not an IMAGE and cannot be plugged into one.
• prompt - your text with references compiled. String, so it feeds any text encode.
• has_images - boolean, true when at least one image is in the strip. Branch on it.
• first_last - boolean, true only in MiniMax first/last-frame mode. Handy for muting branches of a workflow.
• Preset size (HENJICC_PRESET) - a prompt preset's dimensions, wired into HenjiccResolutionBoard, where it quietly outranks the board's own settings.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/henjicc/ComfyUI-Henjicc.git ComfyUI-Henjicc
python -m pip install aiohttp av numpy Pillow

Or Manager → search comfyui-henjicc / Henjicc. Restart, then hard-refresh - the panel is a JS extension and a soft reload sometimes leaves you with a bare text box where the strip should be.

Where it bites

Mixed model families. Connect the same media output to a Qwen encoder and an H3 adapter and the profile walker sees two families and refuses: "Different model families are both connected; use separate Super Input nodes." Correct behaviour, and a surprise the first time.

Soundtracks. Video audio is off by default. Switch it on for a file with no audio stream and the node errors up front - deliberately, because a silent reference would otherwise be dropped without telling you.

Limits. 24 items, 512 MiB per file, 1 MiB of serialised state. Generous, not infinite.

Files versus references. Removing an item from the strip only unlinks it; deleting it in the library is permanent, and picking files from output/temp copies them into input/henjicc/imports. A dangling prompt reference blocks the run until you remove or replace it.

Nothing models this in the community yet. I went looking for Reddit signal on this pack and found none at all - the corpus has zero mentions of henjicc as of mid-2026. Read the source or don't trust the assumption; there are no tutorials to lean on, and also no cargo cult.