Krea2 Full-Res Reference:
MickmumpitzPanoKrea2Reference helps keep a Krea-2 edit reference aligned with a large image canvas. The Krea-2 edit encoder described in the node source limits references to about one megapixel; above that size, the smaller reference can appear anchored in the upper-left of the generated frame. This node encodes the canvas again and replaces the existing reference with one at the canvas resolution, subject to a configurable memory-safety limit. It is particularly useful for high-resolution panorama or equirectangular outpainting where the reference needs to cover the whole target frame. The replacement can require more memory and processing as resolution increases.
Krea2 Full-Res Reference Input Parameters:
conditioning
Required CONDITIONING input. Connect the conditioning produced for your Krea-2 edit workflow. The node keeps that conditioning and replaces its reference_latents value with a newly encoded reference, so downstream generation can use a reference grid that matches the canvas rather than the smaller grid attached by the earlier encode. The source defines no numeric minimum, maximum, or default for this input; it must be connected.
vae
Required VAE input. Connect the VAE used to encode the canvas image into the reference representation stored in the conditioning. This encoding step is what creates the replacement reference. The source defines no numeric minimum, maximum, or default for this input; it must be connected.
image
Required IMAGE input. Provide the canvas you want to use as the reference. Its width and height determine the intended reference size and grid, unless the memory-safety limit reduces them. The node resizes dimensions to multiples of 16 when necessary and encodes the first three image channels, so an alpha channel is not included in the reference. For exact grid alignment, use an image with the intended target dimensions, preferably already divisible by 16. The source defines no numeric minimum, maximum, or default for this input; it must be connected.
max_megapixels
Optional FLOAT input with a default of 8.0, a minimum of 0.25, a maximum of 32.0, and a step of 0.25. This is a safety cap on the number of image pixels encoded for the replacement reference; the calculation uses 1,048,576 pixels per megapixel. If the image exceeds the cap, the node scales it down while preserving its proportions as closely as its 16-pixel dimension rounding allows. Raising the value can let a larger reference retain the target grid, but increases memory demands; lowering it can reduce memory use at the cost of exact grid alignment. The default permits a 2048 × 1024 canvas without cap-driven downscaling.
Krea2 Full-Res Reference Output Parameters:
conditioning
The single CONDITIONING output contains the supplied conditioning with reference_latents set to a list containing the newly VAE-encoded image. Connect it to the downstream part of the Krea-2 edit workflow that uses the conditioning. It is not a rendered image: it is the guidance data carrying the replacement reference. When the input image is at the target size and the safety cap does not reduce it, this reference is intended to cover the full target grid; if the image is capped or its dimensions are rounded, check the printed grid dimensions before assuming an exact match.
Krea2 Full-Res Reference Usage Tips:
- Feed the node the full canvas at the same dimensions as your target generation, and set ModelSamplingFlux to that resolution as noted in the node source. This matters especially for pixel-exact panorama outpainting.
- Use canvas dimensions divisible by 16 to avoid dimension rounding. The node prints the original and encoded sizes, grid dimensions, and token count so you can confirm what it actually used.
- Leave
max_megapixelsat8.0for a 2048 × 1024 canvas. If the log says the reference was capped, increase the limit only if your available memory permits and you need the larger grid. - If memory becomes a problem, lower the cap or work at a smaller canvas size. A reduced cap can keep the workflow running, but it may bring back the reference-to-target grid mismatch the node is meant to prevent.
- Place this node after the conditioning you want to modify. It replaces the conditioning’s existing reference latents rather than adding a second reference alongside them.
Krea2 Full-Res Reference Common Errors and Solutions:
[Mickmumpitz/Pano] Krea2FullResReference: reference … (capped to … MP!)
- Explanation: This is a status message, not an exception. It means the reference image exceeded
max_megapixelsand was reduced before encoding. Its grid may therefore be smaller than the target grid. - Solution: Compare the printed encoded dimensions with your target dimensions. Increase
max_megapixelswithin its allowed range if you need an exact match and have sufficient memory, or reduce the target canvas size.
Reference appears in the upper-left corner
- Explanation: The node source identifies this as the visible symptom of a reference grid smaller than the target grid. It can persist if the image supplied here is smaller than the target or the safety cap reduces it.
- Solution: Supply the full-size target canvas, check that the printed reference dimensions match it, and make sure the conditioning output from this node is the one used downstream.
Out-of-memory error during VAE encoding or generation
- Explanation: A larger encoded reference increases the amount of data the workflow must handle. The exact error wording depends on the VAE, model, device, and runtime; the node does not define a dedicated out-of-memory message.
- Solution: Lower
max_megapixelsor the canvas resolution, then rerun. Check the printed reference size to understand the resulting tradeoff in grid alignment.
Image dimensions change in the node log
- Explanation: The node rounds encoded width and height to multiples of 16, and may also downscale the image when it exceeds the safety cap. This is expected behavior rather than a dedicated error message.
- Solution: Prepare the canvas at target dimensions divisible by 16 and set the cap high enough for those dimensions if exact alignment is required.
Missing required input
- Explanation:
conditioning,vae, andimageare all required sockets. Any missing connection prevents the node from receiving the data it needs; the exact validation message is supplied by the host, not by this class. - Solution: Connect all three required inputs with their matching socket types, then run the workflow again.
