- Reject invalid configs in validate_inputs instead of mid-execution:
Strict-mode multiple_of above max_res, out-of-range min_res/max_res/
multiple_of, and unknown constraint modes
- Share the resolution and multiple_of bounds between define_schema() and
validate_inputs() so the widget and the queue-time check cannot drift
- Cap the min-res pixel budget at an absolute 64 MP, floored at max_res
squared; the multiplier alone permitted ~12 GB outputs at max_res=8192
- Warn on large upscales in Strict mode as well as min-res mode
- Pin GitHub Actions by commit SHA on node24-native majors; drop issues:write
from the publish workflow and set contents:read on both
- Pin Comfy-Org/publish-node-action to the commit @v1 resolves to; the 1.0.1
tag calls `comfy tracking disable`, removed from comfy-cli
- Add requirements-ci.txt with pip caching; add ruff config and a CI lint step
- Add schema-contract, tooltip, budget, crop-guard and validation tests
- Replace typing.Tuple with tuple
- Bump version to 2.4.1
- Refuse extreme-aspect-ratio upscales that would exhaust memory (1x690 at
default settings asked for 704x485760, ~4 GB per batch item)
- Raise instead of returning a zero dimension when multiple_of exceeds max_res
in Strict mode, and when a thin strip truncates to zero at multiple_of=1
- Drop **kwargs from validate_inputs, which disabled ComfyUI's own range and
combo-option checks for every input on the node
- Resize once on the crop path instead of computing and discarding a resize
- Compare exact aspect ratios rather than 4-decimal rounded ones when deciding
whether to crop
- Reject unknown constraint modes rather than applying neither limit
- Log the reason when the pack fails to import
- Bump version to 2.4.0
- Add .github/workflows/test.yml: runs the pytest suite on push/PR
across Python 3.10/3.11/3.12. comfy_api is mocked in conftest.py,
so no ComfyUI install is required.
- execute(): pass a zero-dimension (malformed/empty) input tensor
through unchanged instead of crashing the resizer.
- execute(): warn when "Prioritize Min Resolution" forces a large
upscale (>4x) on an extreme aspect ratio, pointing at the strict
mode as the size-capping alternative.
- Add regression tests for the zero-dim passthrough and the
extreme-upscale warning.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SqaVwWKNn9bsmmhUkJXhZ2
Nearest-multiple rounding could drop a dimension below min_res in
min-res mode (e.g. min_res=1100, multiple_of=256 -> 1024); bump such
dimensions to the next multiple instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Add resize_method input (lanczos default, bicubic, bilinear,
nearest-exact, area) backed by comfy.utils.common_upscale, with
F.interpolate fallback outside ComfyUI; clamp bicubic/lanczos
overshoot to [0, 1]
- Fix rare off-by-one where the crop-path intermediate resize could
truncate below the target dimension, yielding output 1px smaller
than the reported width/height
- Replace StrEnum with (str, Enum) so the pack imports on Python 3.10
- Remove unused numpy dependency from requirements.txt and pyproject
- Switch print() to logging (crop notices at debug level)
- Bump version to 2.3.0, sync README (inputs, technical details,
version history), add requires-python
- Stop tracking .claude/settings.local.json; expand tests to 36
covering execute() output shapes, all resize methods, crop positions
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Critical fixes:
- Fixed inverted cropping logic that caused odd dimensions and math errors
* When original is wider: now correctly resizes by height and crops width
* When original is taller: now correctly resizes by width and crops height
* This fixes the 1792x1024 → 1280x731 bug (now correctly becomes 1280x736)
Configuration updates:
- Changed default multiple_of from 8 to 2 (more appropriate for modern models)
- Updated all SDXL-specific references to be model-agnostic
- Added mentions of modern models (Flux, Stable Diffusion, etc.)
Documentation improvements:
- Updated README to remove dated SDXL references
- Updated tooltips and descriptions to be model-agnostic
- Changed example workflows to use generic "diffusion model" terminology
- Changed io.Input to io.Combo.Input for constraint_mode and crop_position
- Added options= parameter for combo box inputs
- Changed all Output definitions to use display_name= parameter instead of positional argument
This resolves the error: "Input.__init__() got an unexpected keyword argument 'default'"
Changes:
- Set crop_as_required default to True (was False)
Rationale: Most users need exact dimensions for image-to-video and
other strict requirements. Cropping is now the default behavior for
immediate compatibility with downstream processes.
- Increase max resolution limit from 16384 to 65536 (64k)
Rationale: No technical reason to limit at 16k. Users working with
high-resolution content need flexibility up to 64k range.
- Update node description to reflect crop-enabled-by-default behavior
- Comprehensive README overhaul:
* Added "Why Use This Node?" section explaining use cases
* Detailed documentation of all inputs with ranges and recommendations
* Comprehensive outputs section with descriptions
* Added multiple workflow examples (I2V, SDXL, batch, portraits)
* Advanced tips section for edge cases
* Technical details on algorithms and validation
* Installation instructions for ComfyUI Manager and manual
* ComfyUI v3 compatibility notes
* Version history and support information
No version bump - still v2.1.0 as functionality unchanged
Features:
- Actual image resizing using torch bilinear interpolation
- Optional cropping to preserve aspect ratio when rounding causes deviations
- Configurable crop position (center, top, bottom, left, right)
- Resized image now output first, followed by original passthrough
- Input validation with detailed error messages
- Comprehensive tooltips on all inputs and outputs
- Reorganized UI with logical grouping of parameters
- Enhanced description with usage tips
- Max value constraints added to prevent unreasonable inputs
Technical improvements:
- Use torch.nn.functional.interpolate for high-quality resizing
- Smart crop algorithm that resizes first then crops to preserve quality
- Proper tensor shape handling [batch, height, width, channels]
- Aspect ratio deviation warnings when crop is disabled
- min/max bounds on all numeric inputs
Breaking changes:
- Output order changed: resized_image is now first output
- Output names simplified (width/height instead of constrained_width/constrained_height)
- execute() signature expanded with crop parameters
Bump version to 2.1.0
- Convert node to use comfy_api.latest with io.ComfyNode base class
- Replace INPUT_TYPES() with define_schema() returning io.Schema
- Convert execution method from process() to execute() classmethod
- Implement ConstraintModeEnum as io.ComboInput for dropdown options
- Update all inputs to use v3 type system (io.Image, io.Int, etc.)
- Update all outputs to use io.NodeOutput wrapper
- Replace NODE_CLASS_MAPPINGS with comfy_entrypoint() pattern
- Add ConstrainResolutionExtension class for node registration
- Update __init__.py to export comfy_entrypoint
- Add comfy_api dependency to pyproject.toml and requirements.txt
- Bump version to 2.0.0 to reflect major API change
EnragedAntelope