MOSS Storyboard Layout

A 4×2 storyboard grid of 16:9 panels looks like trivial arithmetic until you hand it to a diffusion model. The canvas must be a multiple of 16 or the VAE quietly truncates it. Every panel has to be identical in size to the pixel or your batch stacking throws. The leftover pixels from alignment need somewhere to go that isn't a visible asymmetry. And the crop downstream has to agree with all of it.

This node is that arithmetic, done once, in one place. It loads no model: instant, deterministic, runs on a machine with no GPU.

What it gives you

Tell it how many panels and roughly how big, and it computes the layout: canvas size, columns and rows, each cell's rectangle. Then it emits that geometry in four shapes, because four consumers want it differently.

The one you'll use most is spec: a short string like MOSS-SB1 panels=4 w=1024 h=576 cols=2 rows=2 pw=496 ph=272 gutter=8 margin=8 carrying the entire layout down one wire. Wire it into the Sampler, Crop or Paste and those nodes stop guessing. That's the plumbing-layer value-node pattern: one source fanned out, so five nodes can't disagree about the grid.

layout_json is the same format the host's own MOSS layout prompt emits - summary plus items with absolute-pixel bboxes from canvas centre. Straight into a CLIPTextEncode. elements is a different format for a different consumer: {type, desc, bbox} with 0–1000 normalised bboxes in [y0, x0, y1, x1] order, built for the host's Create Bounding Boxes node so you can see and edit cells on that canvas.

template is a preview image of the grid - grey cells at cell_gray (default 25) broken by your gutter. For your eyes; it does not enter sampling. report is one line: panel count, canvas, megapixels, how many prompt segments lined up, oversize warnings. The rest are plain numbers - width, height, columns, rows - for an EmptyFlux2LatentImage or a bounding-box node.

The inputs that matter

panels is what you came for: 3, 4, 6, 8 or 9. Then panel_aspect (ten choices, 16:9 through 2.39:1) and panel_edge, the panel's long edge. Default 512, and the tooltip is honest about why: eight panels at 512 lands the sheet near 1.3MP, inside MOSS's 1–2MP training range. Push it and you leave the range, small text first.

canvas_aspect sets orientation. auto follows panel orientation; set it explicitly and the same 6 panels become 3×2 or 2×3. To force an arrangement, fill columns (rows auto-completes) or both - and then they must multiply to panels or you get an error rather than a surprise.

gutter and margin are seam width and outer padding, both default 8. Their tooltips are worth reading: gutter = 0 and the model will happily draw across the boundary; margin = 0 puts content flush to the edge so your crop picks up a rim instead of picture.

prompt exists and you can leave it empty - this node only computes geometry.

The advanced pair, cell_gray and invert, are preview styling. label sets the bbox type: object for "this cell is a picture", character to nudge that the cell is a person.

Install

Manager, search MOSS Storyboard. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ZenHG/ComfyUI-MOSS-Storyboard

You want [MOSS SB] v0.2.2 in the startup log. The pack installs no pip dependencies - requirements.txt exists purely so Manager skips the pip step. But it needs host ComfyUI ≥ 0.42.0, and to sample anything downstream you need the MOSS trio: moss_image_sft_bf16.safetensors in models/unet/, qwen3.5_9b_int8_convrot.safetensors in models/clip/, flux2-vae.safetensors in models/vae/. This node runs fine without any of them.

Where people get burned

Wiring elements into Create Bounding Boxes and getting boxes everywhere. You need three wires: width → its width, height → its height, elements → its bboxes. Normalised 0–1000 coordinates need real pixel dimensions to resolve against, and that canvas node defaults to 1024×1024 while an 8-panel sheet is 2032×576. Skip the wires and you get silently wrong boxes - no error, just nonsense.

Confusing the two formats. elements (normalised canvas elements) and layout_json (absolute pixels, full layout prompt) are not interchangeable. Swapping them won't parse as you expect.

Extreme canvas ratios. The pack caps canvas aspect near 4:1, because past that the model paints one long strip and calls it a day. Three 16:9 panels in a row lands around 5.3:1 and is rejected - 3 panels default to a vertical stack for exactly this reason. Set columns = 3 for the row and expect to fight the composition.