MOSS Storyboard Sampler

Making a comic with a diffusion model was never really a model problem. It's a geometry problem wearing a creative hat. The genre's most-cited write-up - a 72-page graphic novel that took three months of semi-full-time work - puts prompting at about 35% of the effort, the rest going to layout and manual correction. This node attacks that other 65%: give it a model and a script, get back a clean grid and the panels, aligned to the pixel.

What it actually does

You write a storyboard script: one global paragraph for characters and style, then [PANEL 1]: through [PANEL n]: for the shots. The node splits that into panels, converts each into a MOSS-native layout box (the same JSON the host's layout prompt node emits), encodes, samples, and hands back both the assembled grid and the crops.

Two ways to generate, and it matters which:

整张 (whole canvas) - one pass over the entire grid. Fast, good for feeling out a composition or running 8–9 panels.

逐格 (per panel) - each panel is sampled independently at its own canvas size, then assembled and re-bordered. Costs ~N× the time, buys two things. First, cross-panel consistency: whole-canvas mode spreads the global description thin across N regions and re-imagines the character in each, so hats change colour and faces drift. Per-panel, every sample reads the whole global paragraph again - the author's measured variance on a repeated red hat across four panels fell from 0.0036 to 0.00004. Second, resolution: sample_edge (default 1024) sets the sampling canvas per panel, decoupled from the finished panel size. Measured: 496 soft, 992 optimal, 1984 worse because composition loosens.

One detail that explains odd behaviour: the prompt cleaner strips multi-panel phrasing - "exactly 4 panels", "thin white borders" - before per-panel encoding, because otherwise the model draws four panels inside your single-panel box - it did, in testing, with one bbox.

The inputs you'll actually touch

Wire model from a UNETLoader, clip from a CLIPLoader, vae from a VAELoader, and type your script into positive.

That CLIPLoader type must be moss_image. Get it wrong and nothing errors - you just get the wrong embeddings and a picture that means something else. Most common silent failure in the MOSS stack.

steps and cfg come in two tiers that don't mix: official weights want 30–50 steps at cfg 3–5; with a 4-step distillation LoRA, 4–8 steps at cfg 1.0. Running 35 steps on a distilled LoRA isn't an error, it just wastes a minute and can hand you mush. The inverse gives you the black frame everyone posts about.

denoise stays at 1.0. There's no image input; lowering it truncates the sigma schedule and effectively runs fewer steps, and report prints nominal vs actual.

Optional spec beats everything: wire it in from the Layout node and every geometry widget here (panels, panel_aspect, panel_edge, canvas_aspect, gutter, margin) goes dead. Number-one wrong-shape complaint.

Outputs: grid, panels (a batch, row-major), latent, spec (feed it to Crop and Paste), prompt_used (what actually got encoded - in layout mode, the generated JSON; useful when a panel ignores you), and report. From the source: latent is the real sampled latent in whole-canvas mode (wire your own VAEDecode); in per-panel mode it comes back empty.

Install

Manager, search MOSS Storyboard. Or:

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

Restart; look for [MOSS SB] v0.2.2 in the log. No line, no pack.

The requirements outweigh the install: ComfyUI ≥ 0.42.0 (older hosts have no moss_image clip type - nodes appear, encoding silently misbehaves), moss_image_sft_bf16.safetensors in models/unet/, qwen3.5_9b_int8_convrot.safetensors in models/clip/, flux2-vae.safetensors in models/vae/. The pack pulls zero pip packages; requirements.txt is deliberately empty.

Where people get burned

Black output. Almost always the steps/cfg tier mismatch - read report first.

OOM at the VAE decode, or faint seams across your panels. Same cause: the host fell back to tiled decoding and the boundaries cut through your grid. Search the log for retrying with tiled VAE decoding. Fixes in ascending cost: git pull the host (older builds estimated decode VRAM at 4× reality), halve panel_edge (quarter the decode memory), drop 8 panels to 4. Note --fp32-vae silently disables the memory-saving path.

Panels populated but wrong. MOSS layout following is officially experimental - boxes are a strong suggestion, not a contract. One shot per [PANEL n], and check the segment count in report.

Per-panel mode refusing to run. It needs a [PANEL n] segment for every panel and errors rather than padding with blanks.

Start from examples/moss-storyboard-4panels-ui.json - pre-wired, grouped, sane defaults.