A standalone ComfyUI node for preparing images for GPT Image 2.5 Sunburst and Flare. Calculates a valid output canvas, crops or pads without deliberate stretching, and returns the prepared image plus dimensions. No API calls or keys are required.

Automatic editing workflow

Use Gpt Image Auto Prepare before generation and Gpt Image Auto Restore afterward. These companion nodes have no sizing or positioning widgets. The original Gpt Image Size Calculator remains available for manual control and existing workflows.

Connect the original ComfyUI image to Auto Prepare's raw_image.
Connect its prepared image and width/height to the generation node. With NanoSeed, keep resolution at 4K.
Connect the generated image to Auto Restore's edited_image.
Connect the same original to Restore's raw_image, and Prepare's transform to Restore's transform.
Optionally connect an edit mask at the original image resolution to Restore's edit_mask.

Prepare starts with the verified NanoSeed 4K dimensions and repeats the size calculation until the canvas is stable when fed back through NanoSeed. This avoids a second rounding adjustment in the receiving node. It fits the whole original into that canvas with centered padding and records the exact integer content rectangle. Restore removes that padding and maps the result back to the original width and height using the recorded geometry. There is no automatic subject detection or manual crop selection.

For face replacement or another local edit, a raw-resolution mask makes preservation precise: pixels where the mask is exactly zero are copied unchanged from the original tensor. White selects the edit; gray blends it. The optional restored_edit output is the complete restored generated image before compositing. Without a mask, the whole generated result is restored, and unchanged background pixels are not guaranteed to match the original.

An optional mask connected to Prepare is resized/padded alongside its image. Its mask output follows ComfyUI's white-means-edit convention and is not necessarily a ready-to-upload API mask; check the generation node's mask conversion. Use the original-resolution mask for Restore, not Prepare's resized output. If no mask is supplied, Prepare outputs a mask selecting the image content and excluding padding.

Geometry restoration is not image registration. The output canvas matches the original, but model-induced movement or reshaping cannot be undone from dimensions alone. Resizing also cannot preserve all original pixel values inside the edited area. Use the exact same raw image and the matching transform for each edit. Restore rejects unexpected generated dimensions rather than guessing how the service transformed them. “Raw image” means the original ComfyUI IMAGE tensor, not a camera RAW file decoder.

One original can be broadcast across multiple generated variations. Otherwise batches must be compatible; masks may have one item or match the final batch. Channels must match between the original and edit. For three-channel API output, use a three-channel original.

Install

Clone this repository into your ComfyUI custom_nodes directory:

git clone https://github.com/comrender/ComfyUI-Gpt-Image-Size-Calculator.git

Restart ComfyUI and search for Gpt Image Size Calculator. Requires Python 3.10+ and uses PyTorch already supplied by ComfyUI; no additional runtime packages or API keys are needed. You can also download the repository ZIP and extract its folder into custom_nodes. This package is independent of Nano Banana and does not import or change its node.

Once published and indexed in ComfyUI Registry, its package ID is gpt-image-size-calculator, under publisher comrender.

Size rules

The OpenAI size documentation, checked September 9, 2026, specifies:

Width and height are multiples of 16.
Each edge is at most 3840 pixels.
Width/height is between 1/3 and 3, inclusive.
Area is between 655,360 and 8,294,400 pixels, inclusive.
Resolutions above 2560 × 1440 are experimental.

These are custom-size constraints, not a fixed list of aspect-ratio buckets. Ratios such as 1:1, 4:5, 3:4, 2:3, 9:16, 16:9, and 3:1 are possible when the actual dimensions meet every constraint. The size output describes the actual canvas ratio, not a rounded conventional label.

Verified NanoSeed 4K behavior

All 12 user-supplied Sunburst test outputs match the sizes calculated by the current local NanoSeed function, using 4K, GPT image size auto, width/height zero, and aspect ratio auto. Request sizes were reconstructed from that source code, not read from historical Fal request logs. The new default is NanoSeed 4K (verified). This verifies the tested workflow; it does not reveal the provider's hidden auto-sizing algorithm or establish Flare behavior empirically.

The calculation preserves orientation, clamps the long/short ratio to 3, scales toward an 8,294,400-pixel budget with a 3840-pixel maximum edge, and tries the floor/ceiling multiples of 16 on each edge. It selects the largest valid area among those four candidates, using ratio error as a tie-breaker. This is a local rounding rule, not a search over every possible valid canvas. NanoSeed submits an explicit size even when its GPT image size control says auto.

The repeated 3678 × 4598 input produced 2576 × 3216 both times. The earlier reported 2592 × 3200 result was not reproduced for that input. The initial area-first heuristic matched only 4 of the 12 tests and is retained only as an explicitly experimental alternative. Exact dimensions can still reproduce the earlier target.

Crop mode fits the selected canvas by trimming edges before resizing; padding retains the full frame. Integer crop boundaries introduce less than a pixel of rounding error relative to the ideal crop. Dimensions alone do not prove whether generated content was cropped, stretched, or recomposed by the model.

See the complete test analysis.

Controls

| Control | Behavior | | --- | --- | | NanoSeed 4K (verified) | Default. Matches all 12 supplied test sizes using NanoSeed's four-candidate calculation. Fixed 8,294,400-pixel budget and 3840-pixel edge; pixel_budget and max_crop_percent are ignored. | | Max area (experimental estimate) | Selects the largest valid canvas within the allowed aspect crop loss; default tolerance is 2%. Reproduces the observed example but does not predict provider auto sizing. | | Closest aspect | Favors the closest ratio among canvases using at least 95% of the ideal feasible area. Allows a small area sacrifice to preserve composition. | | Exact dimensions | Uses target width/height exactly after validating all size rules. Pixel budget and crop tolerance do not apply. | | pixel_budget | Maximum area for Max area and Closest aspect modes; defaults to 8,294,400. It is a node policy, not a quality setting. Smaller images may be upscaled. | | max_crop_percent | Aspect tolerance for the area-first calculation, not a hard limit on final crop. Inputs outside 1:3–3:1 must first fit the supported range. If no candidate meets tolerance, the closest available ratio is selected. | | Crop | Removes edges to fit the target, then resizes. | | Pad | Resizes the full frame to fit, then adds black padding. | | crop_x / crop_y | Position from 0 (left/top) through 0.5 (center) to 1 (right/bottom); also positions content in Pad mode. |

Quality is deliberately separate: high, xhigh, and max are generation settings, not resolution presets.

Connect to generation

Connect the prepared IMAGE to the editing node and use the width and height outputs as the explicit generation dimensions. For direct OpenAI requests, size is the returned WIDTHxHEIGHT string. For Fal, pass image_size: {"width": width, "height": height}. Preprocessing alone cannot force a downstream node or service using auto to keep this canvas.

For a NanoSeed workflow, select NanoSeed 4K (verified), connect the prepared image and both dimension outputs, and keep NanoSeed at 4K. Its size calculation preserves the selected sizes in the tested cases. To connect width/height to widget-based inputs, use ComfyUI's convert-widget-to-input option on the receiving node.

Other modes may select targets NanoSeed recalculates; exact general passthrough requires support in the calling node. NanoSeed was not modified by this package.

Outputs: prepared image, width, height, size string, reduced aspect-ratio string, and an information string describing crop/padding and sizing limitations. Batches share the same crop and canvas. No subject detection is performed.

Verification

Run python -m unittest discover -s tests -v with a Python environment containing PyTorch. Tests cover supplied output sizes, constraints, crop/pad geometry, and image tensors. The 12 live results were generated by the user with NanoSeed; the new calculator's full ComfyUI UI integration has not been exercised here.

Publishing releases

pyproject.toml contains the Comfy Registry metadata. Registry publication additionally requires a publisher API key saved in GitHub Actions as REGISTRY_ACCESS_TOKEN. Creating the file or repository alone does not publish a registry listing.

Update the semantic version and changelog, push the changes, then create a GitHub release for that version. The publishing workflow runs tests on Python 3.10 and 3.12 before uploading to Comfy Registry. It can also be started manually under Actions. Each version must be unique; do not republish an existing version. Registry indexing may take additional time after a successful upload.

License

MIT. Maintained by comrender. Not affiliated with OpenAI or Fal.