Super Director Tools · Asset Prompt

"Super Director Tools · Asset Prompt" takes a description you typed and turns it into a much longer description a diffusion model will actually obey, for three specific jobs: a character reference sheet, a seamless 2:1 panorama, or a 3×3 asset sheet. It calls no model, no LLM, no translation API. It's a local string builder with a dictionary of templates. That's the whole trick, and it's a good one.

Why this exists

If you've chased character consistency, you've met the turnaround sheet - front/side/back in one image, used afterwards as an IP-Adapter reference, a ControlNet input, or a LoRA training set. It's the workflow the community kept landing on: CharTurnerV2 built a whole model around it, Multi-View Character Creator pushed it to +775, and the Kontext-era sheet LoRAs made it cheap again.

The catch is that hand-writing sheet prompts is tedious and you always forget a constraint. You ask for three views and get two. You get three views with different belt buckles. You get a nice sheet with "FRONT" written across it in Comic Sans. The prompt you need is long, boring and highly specific, and you end up pasting the same 150 words every time.

This node is that 150 words, parameterised. Pick a mode, fill in your description, and it staples on the constraints that make sheets work: same character at the same scale, front to back left to right, head to feet fully visible, no text or logos, identity and costume locked across views. Your own text is preserved verbatim - it doesn't paraphrase you, and it doesn't translate you.

What it does mechanically

The pack ships template dictionaries: per-option lines for style, background, expression, environment, time_of_day and lighting in English and Chinese, a character_layout block, a character-consistency block, and a long panorama block spelling out 360°×180° equirectangular, 2:1, seamless left/right wraparound, correct pole stretching, and - pointedly - "do not substitute a perspective view, fisheye view, cropped wide image or little-planet projection."

Only the fields relevant to the current asset_type get appended. Switch from Character to Panorama and the irrelevant widgets hide, but their values are remembered, so switching back doesn't lose your work. None adds literally nothing, which is the correct default: a vague extra sentence is worse than no sentence.

The fields you actually set

custom_prompt is the one that matters - your character, prop or scene, in whatever language you want. asset_type picks the job: Character, Panorama, or Nine-grid asset · 九宮格資產. Then character_layout (Three views, Three views + face close-up, Single full body) if you're doing a character, or grid_asset_type (Character / Scene / Prop) if you're doing the nine-grid - that one decides what the nine panels are, from nine character views down to nine prop angles.

language only controls the appended preset text, not your text. Leave it on English unless you're deliberately prompting in Chinese. aspect_ratio is the other one worth noticing: Panorama is always 2:1 and hides it; for a nine-grid the ratio applies to the whole sheet, so 16:9 means each cell is 16:9 too.

The single output is prompt (STRING), and it goes wherever a text prompt goes - CLIP Text Encode's text, usually. If the field renders as a text box with no dot to drag from, right-click and convert it to an input first. That's a ComfyUI thing, not this pack's fault.

The size trap

Big one: the output is a string. It cannot change your image dimensions. Ask for a 2:1 panorama all you like - if your latent is 1024×1024 you'll get a square image with a stretched horizon. Set 2048 × 1024 yourself. For a nine-grid matching a 1:1 sheet use 1536 × 1536; for 16:9 use 2304 × 1296.

And a prompt can't guarantee a model obeys. On older SDXL checkpoints, 3×3 grids come out as nine vague smudges or one hero shot with a vague border. Edit models and modern instruction models are much better at this. Hand-write nothing, but do look at the output.

Install

ComfyUI Manager → search Super Director Tools → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/mch000534/ComfyUI-DirectorTools

Restart ComfyUI and refresh the page. There is nothing else to do - the pack's requirements.txt literally reads "No additional dependencies," it borrows ComfyUI's own torch/numpy/Pillow, and there are no model downloads. Just make sure you don't end up with ComfyUI-DirectorTools/ComfyUI-DirectorTools/ from a bad unzip.

Where people get stuck

Nothing runs: prompt isn't connected downstream, or the workflow has no output node. Prompt in Chinese but preset text in English: that's by design. And if a nine-grid generation comes out as nine separate images - it didn't, the node never generates anything. It emits text; the sampler does the rest.