comfyui-video-resolution

Kirei ComfyUI Video Resolution

Choose a video canvas. Get dimensions aligned to your model.

One node for short-edge presets, megapixel budgets and custom dimensions—with profiles for MiniMax H3, Wan, LTX, Hunyuan Video, CogVideoX and Mochi. Outputs width, height and resolution_str. No models to download, no extra Python dependencies, and no GPU required for the selector.

Quick start · Presets · Model profiles · Inputs · FAQ

Installation

ComfyUI Manager: search for comfyui-video-resolution, install, then restart ComfyUI and refresh the browser.

Manual installation:

cd ComfyUI/custom_nodes
git clone https://github.com/hyukudan/comfyui-video-resolution.git

Requires Python 3.10+. Add Kirei ComfyUI Video Resolution from the node menu; it is in the video category. Search aliases include h3, minimax, video size and ltx2.

The Git repository is already listed in Manager. The first Registry release under ID video-resolution is being prepared; the features below describe version 1.2.1. Avoid installing a second copy alongside the existing one.

The display name includes Kirei; the repository name, Registry ID and internal node ID (VideoResolutionNode) remain unchanged for existing installations and workflows.

Quick start: MiniMax H3

1. Add Kirei ComfyUI Video Resolution and set model_profile to MiniMax H3.
2. Set resolution to H3 Native (768p), aspect_ratio to 16:9 (Widescreen), and keep scale = 1x and swap = false.
3. Connect width and height to the corresponding inputs of your H3 generation node. The result is 1344×768 with the default floor rounding.

flowchart LR
    R[Kirei ComfyUI Video Resolution] -->|width + height| H[H3 generation node]
    E[Optional H3 Prompt Enhancer] -->|enhanced_prompt| H

The selector calculates dimensions; it does not create latents, resize reference images, crop video, or configure generation. Your downstream node handles those jobs.

For a smaller test canvas, choose 0.3 MP and rounding_mode = nearest: the same 16:9 H3 setup produces 736×416. Lower pixel counts reduce the spatial workload, but are not a promise of a particular VRAM usage, speed or image quality.

Choose a sizing mode

| You want… | Choose | How it works | |---|---|---| | H3's native canvas | H3 Native (768p) + MiniMax H3 profile | 768 short edge, 1344×768 area budget, then native 32-pixel alignment | | Similar pixel area across aspect ratios | 0.3, 0.5, 0.7 or 1.0 MP | Derives width/height from area and aspect ratio | | Your own pixel budget | Custom MP | Set custom_megapixels between 0.01 and 16 MP; uses the same calculation as the fixed MP presets | | A familiar short-edge size | 480p, 540p, 720p, 768p, 1080p, 1440p, 2160p or 4320p | Sets the shorter dimension, then derives the longer one | | Your own width and height | Custom → manual | Uses both custom dimensions; ignores aspect_ratio | | A known width | Custom → from_width | Derives height from width and aspect ratio | | A known height | Custom → from_height | Derives width from height and aspect ratio |

All modes then apply scale, the selected model's alignment and swap. Custom dimensions are still scaled and aligned; “manual” does not bypass model rules.

For example, Custom MP = 0.4, MiniMax H3, 16:9 and nearest produces 832×480 at 1x. The MP field appears only in Custom MP mode. Its limits apply before scale/alignment; they are not limits on model memory or final pixel area.

Available aspect ratios: 1:1, 16:9, 19:9, 2.39:1, 9:16, 4:3, 5:4, 3:4, 4:5, 21:9, 9:21, 3:2 and 2:3. Portrait presets use the same short-edge convention as landscape presets.

Megapixels, scaling and rounding

• 1 MP = 1,000,000 pixels, matching the H3 Prompt Enhancer. Some core ComfyUI sizing nodes use 1024² pixels per MP instead.
• MP presets are approximate targets, not strict memory or pixel caps. floor rounds each dimension down; nearest picks the closest aligned dimension; ceil rounds up. Alignment can slightly change the aspect ratio.
• scale multiplies both dimensions: 0.5x gives roughly one quarter of the original pixel area, while 2x gives roughly four times the area.
• For H3 Native, native alignment happens first, then scale and final model alignment. Its area budget applies before rounding; the aligned canvas can slightly exceed it.

For compatibility, short-edge and Custom modes truncate fractional pixels after scaling and before alignment. MP modes keep those fractions until alignment. Consequently, legacy ceil can differ from rounding the untruncated scaled size.

Model profiles

Profiles set spatial alignment. They do not select a model, enforce its trained resolution range, or set duration, FPS or frame count.

| Profile | Width/height multiple | Intended use | |---|---:|---| | MiniMax H3 | 32 | H3's 16× visual VAE followed by 2×2 spatial patchification | | Wan 2.2 TI2V 5B | 32 | The 5B TI2V VAE family | | Wan 2.x | 16 | Existing Wan 2.1 / Wan 2.2 14B sizing | | Hunyuan Video | 16 | Original Hunyuan Video | | Hunyuan Video 1.5 | 16 | Native ComfyUI node's spatial step | | LTX Video / LTX2 | 64 | Conservative alignment for LTX workflows | | LTX 2.3 | 64 | Same alignment, explicit model label | | CogVideoX | 16 | Existing CogVideoX profile | | Mochi | 64 | Existing Mochi profile | | Custom | 8, 16, 32 or 64 | Manual divisible_by and optional legacy add_one |

Every named profile disables add_one. Temporal rules such as “4n+1 frames” must not be applied to width or height. Selecting a resolution preset does not automatically change the model profile—select both for H3.

H3 examples

All examples use model_profile = MiniMax H3, scale = 1x and swap = false.

| Resolution | Aspect ratio | Rounding | Output | |---|---|---|---| | H3 Native (768p) | 16:9 | floor | 1344×768 | | H3 Native (768p) | 9:16 | floor | 768×1344 | | H3 Native (768p) | 1:1 | floor | 768×768 | | H3 Native (768p) | 4:3 | floor | 1024×768 | | H3 Native (768p) | 21:9 | floor | 1536×672 | | 0.3 MP | 16:9 | nearest | 736×416 | | 0.5 MP | 16:9 | nearest | 928×544 | | 0.7 MP | 16:9 | nearest | 1120×640 | | 1.0 MP | 16:9 | nearest | 1344×736 | | 768p | 21:9 | floor | 1792×768 |

H3 Native and plain 768p are different: the native mode also applies an area budget, so an ultrawide canvas becomes shorter than 768 pixels. It follows ComfyUI's adapt_canvas. Higher resolutions remain selectable, but are not H3's separate hosted 2K regeneration pipeline.

Live size preview

The panel below the controls shows the final width × height, actual MP and effective alignment. For example: 736 × 416 and 0.306 MP · multiple of 32. It uses the same Python calculation as execution through your ComfyUI server; no separate browser rounding rules or external services are involved.

Changes are debounced briefly. If any input is connected from another node, the panel shows Connected inputs instead of presenting local widget values as the result. Disconnect the inputs to resume the live preview. Values from connections are still handled normally during graph execution.

The panel is informational: it adds no output sockets or saved widget values. For the legacy Custom-profile offset it explicitly shows grid 32 + 1 (legacy) rather than claiming the final dimensions are divisible by 32.

With MiniMax H3 Prompt Enhancer

The Prompt Enhancer is optional. Send its enhanced_prompt to H3 and use one source of dimensions: either the enhancer's width/height outputs or this selector's outputs.

For matching dimensions, use MiniMax H3 + 1x + swap=false and the same aspect ratio in both nodes. Their shared ratios are 16:9, 9:16, 1:1, 4:3, 3:4 and 21:9; the enhancer's auto falls back to 16:9, not the aspect of an input image.

H3 Native matches the enhancer's Auto sizing in 0.15.0; MP presets with nearest match its custom MP sizing. 0.15.0 is being prepared and is not yet published; Registry version 0.14.1 uses older sizing rules. The packages install separately and do not import each other.

Inputs and outputs

| Input | Purpose | |---|---| | resolution | Native, short-edge, MP or Custom sizing | | aspect_ratio | Frame shape; ignored in Custom/manual | | scale | 0.25x, 0.5x, 0.75x, 1x, 1.25x, 1.5x or 2x | | model_profile | Chooses the spatial alignment rules | | custom_mode | manual, from_width or from_height; only for Custom resolution | | custom_width | Width reference; unused in from_height | | custom_height | Height reference; unused in from_width | | divisible_by | Manual alignment; only for the Custom model profile | | rounding_mode | floor, nearest or ceil; floor is the existing default | | add_one | Legacy +1 after alignment; only for the Custom model profile | | swap | Swaps the final dimensions | | custom_megapixels | Optional, last input; 0.01–16 MP, default 0.5; used only in Custom MP |

| Output | Type | Example | |---|---|---| | width | INT | 1344 | | height | INT | 768 | | resolution_str | STRING | 1344x768 |

The node returns the final dimensions when the graph executes. resolution_str can also feed a compatible text display; the live panel is separate from these outputs.

FAQ

Why does 720p give me 704 pixels? 720 is not divisible by 64. The existing default—LTX profile plus floor rounding—therefore returns 1280×704 for 16:9. Use the profile for your actual model; use nearest if you prefer the closest aligned size rather than always rounding down.

Where did the Custom controls go? The node shows only controls used by the selected mode and adjusts its height automatically:

• Presets: aspect ratio, scale, model profile, rounding and swap.
• Custom/manual: width and height replace the aspect ratio.
• Custom/from_width or from_height: aspect ratio plus the reference dimension.
• Custom MP: aspect ratio plus the megapixel budget.
• Custom model profile: also shows divisible_by and the legacy add_one.

Hidden values remain saved and return when you switch back. Connected inputs remain accessible; if a mode selector is connected, its potentially relevant controls stay visible because the mode will be resolved during execution. This compact layout uses ComfyUI's current widget visibility support; refresh the browser after updating.

Why does the preview ask me to restart? The preview endpoint is registered when ComfyUI loads the node. After updating, restart ComfyUI and refresh the browser. If the server cannot be reached, the panel says Preview unavailable rather than leaving an old size visible.

Why did Custom add one pixel? The legacy Custom-profile default is add_one = true. Set it to false unless your downstream workflow explicitly requires that behavior. Named profiles always keep it off.

Can I use 4K or 8K with any model? These are size presets, not model capability claims. The selector neither checks available VRAM nor guarantees that the generator will accept or perform well at that resolution. Start with the model's documented canvas.

Compatibility and development

Version 1.2.0 keeps VideoResolutionNode, all previous combo values, the existing input order/defaults and all three output sockets. custom_megapixels is optional and appended after the old inputs; it defaults to 0.5 when omitted. Existing selector workflows do not need migration. The calculation uses only the Python standard library; the preview route uses ComfyUI's existing web server dependencies.

Run the tests from the repository root:

python -m unittest discover -s tests -v
node --test tests/test_ui.mjs

Sizing references: H3 architecture, ComfyUI H3, Wan, Hunyuan Video.

<details> <summary>Maintainers: publishing to Comfy Registry</summary>

The default branch is master, Registry ID is video-resolution, and publisher is hyukudan. Configure the repository secret REGISTRY_ACCESS_TOKEN for that publisher. The publish workflow runs tests before upload when pyproject.toml changes on master, or when manually dispatched. Each release needs a new semantic version. Verify that it becomes active in Registry after the workflow completes.

Manager already lists the Git repository; do not submit a duplicate registration. See the official publishing instructions.

</details>

License

MIT.