Smart Aspect Ratio

ComfyUI Smart Aspect Ratio

A ComfyUI custom node that converts a megapixel target and aspect ratio into pixel dimensions (width × height). The ratio can be fixed, randomized, or detected automatically from a connected prompt.

Type cinematic ultrawide city at night and the node hands your latent a 21:9 resolution. Write 16:9 anywhere in the prompt and it obeys exactly. Or skip detection entirely and use it as a clean megapixels-plus-ratio resolution picker and/or randomize the aspect ratio for each run.

---

Features

• 23 aspect ratios from 9:32 portrait to 32:9 landscape
• Prompt-based detection — connect your prompt text and the node picks a fitting ratio automatically
• Explicit override — writing 16:9 (or any accepted ratio) in the prompt forces it exactly
• Keyword matching — terms like cinematic, widescreen, portrait, or anamorphic map to weighted ratio pools
• Seed-controlled randomization — choose between a fixed or random ratio when nothing matches, reproducible via seed
• Editable keyword rules — click Edit Keywords on the node to change the mapping, no restart needed

Installation

ComfyUI Manager: search for Smart Aspect Ratio. <!-- keep this in sync with the registry DisplayName -->

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/dawncreatescode/comfyui-ResolutionAndAspectRatio

Restart ComfyUI. The node appears under the Resolution category.

Quick start

1. Add the node: Add Node → Resolution → Smart Aspect Ratio.
2. On your Empty Latent Image, connect width and height to this node.
3. (Optional, for prompt detection) Feed the same prompt string into both your CLIP Text Encode and this node's prompt input — e.g. from a String/Primitive node:

[String: prompt] ──┬──→ [CLIP Text Encode] ──→ ...
                   └──→ [Smart Aspect Ratio] ──width───→ [Empty Latent Image]
                                              ──height──→

4. Pick a megapixels target (1.0 ≈ SDXL native) and you're done. Without a prompt connection, the node is a simple ratio/resolution picker using the aspect_ratio dropdown.

How the ratio is chosen

Detection runs in strict priority order — first match wins:

1. Explicit ratio in the prompt. 16:9 photo of a harbor → 16:9 (from prompt). Always wins.
2. Keyword match. anamorphic close-up → 21:9 (anamorphic). When a keyword maps to several ratios (e.g. widescreen → 16:9 / 19:9 / 21:9), one is picked by weighted random using the seed.
3. Fallback. Nothing matched → the fixed aspect_ratio dropdown, or a random ratio if randomize is on.

The ratio output always tells you what fired — 21:9 (from prompt), 16:9 (cinematic), 4:5 (fallback). Connect it to a Show Text node while you're tuning.

Keyword modes — and why Limited is the default

| Mode | What triggers detection | |---|---| | Off | Nothing — fixed or random ratio only | | AR only | Only explicit ratios written in the prompt (16:9, 3:2, …) | | Limited keywords (default) | Explicit ratios + unambiguous format terms (widescreen, anamorphic, ultrawide, …) | | All keywords | Everything above + content words (cinematic, portrait, poster, phone, …) |

Content words are powerful but can be unintended: in portrait of a king, "portrait" describes the subject, not the orientation — yet All mode will dutifully make the image tall. Limited only reacts to words that unambiguously describe a format, which is why it's the default. Switch to All when you want maximum hands-off behavior and your prompts don't use those words for content.

Default keyword rules

| Words | Ratio pool | Active in | |---|---|---| | anamorphic | 21:9 | Limited + All | | imax | 4:3 | Limited + All | | ultrawide | 21:9 32:9 | Limited + All | | widescreen | 16:9 19:9 21:9 | Limited + All | | panorama, panoramic | 16:9 19:9 21:9 32:9 | Limited + All | | square, avatar, thumbnail | 1:1 | All only | | mobile, phone, smartphone | 9:16 9:19 9:21 | All only | | poster | 2:3 3:4 4:5 | All only | | cinematic | 16:9 21:9 32:9 | All only | | banner, wallpaper | 16:9 21:9 32:9 | All only | | landscape, horizontal, wide angle, wide shot | all landscape ratios | All only | | portrait, vertical, tall | all portrait ratios | All only |

Supported aspect ratios

Portrait: 2:3 3:4 3:5 4:5 5:7 5:8 7:9 9:16 9:19 9:21 9:32 · Square: 1:1 · Landscape: 3:2 4:3 5:3 5:4 7:5 8:5 9:7 16:9 19:9 21:9 32:9

Editing the rules

Click Edit Keywords on the node. Each rule has:

• Words — comma-separated triggers (case-insensitive, phrase-aware)
• Ratios — comma-separated pool, or $LANDSCAPE / $PORTRAIT for the full sets
• Weights — optional per-ratio probabilities (blank = equal chance)
• Category — limited (active in Limited + All) or all (All mode only)

Changes save to keywords.json in the node folder and apply immediately.

Node reference

Inputs

| Input | Type | Description | |---|---|---| | megapixels | dropdown | Pixel budget, 0.1–2.5 MP (1.0 ≈ SDXL native, ~0.25 ≈ SD 1.5 native) | | aspect_ratio | dropdown | Fallback ratio when detection is off or finds nothing | | divisible_by | dropdown | Round dimensions to a multiple of 8 / 16 / 32 / 64 — see below | | keyword_mode | dropdown | Which prompt keywords trigger detection (see modes) | | randomize | boolean | Pick a random ratio when nothing matches | | seed | INT | Fixed seed → reproducible picks; set control_after_generate to randomize for variety per run | | prompt | STRING (optional) | The text to scan — same string you feed your CLIP Text Encode |

Outputs

| Output | Description | |---|---| | width | Image width in pixels | | height | Image height in pixels | | ratio | Selected ratio + what triggered it, e.g. 16:9 (cinematic) |

Choosing divisible_by

Stable Diffusion latents require dimensions divisible by 8, so 8 is the hard minimum. Larger multiples (32, 64) are safer across models and avoid edge artifacts some checkpoints show at odd sizes — at the cost of landing slightly further from your exact ratio and megapixel target. Rule of thumb: leave it at 64; drop to 16 or 8 only if you need shapes as exact as possible and know your model handles them.

Troubleshooting

My keyword isn't detected. Check keyword_mode — content words like cinematic need All keywords. Also confirm the prompt input is actually connected; the node can't see text it isn't given.

Dimensions aren't exactly my ratio. That's divisible_by rounding. Lower it for a closer fit.

randomize gives the same ratio every run. The seed is fixed. Set the seed widget's control_after_generate to randomize.

Which ratio fired, and why? Connect the ratio output to a Show Text node — it names the trigger.

License

MIT