MOSS-Image Official Sampler:
MOSSOfficialSampler generates image latents using the pinned official MOSS image pipeline. Its description specifies an Euler sampling schedule aligned to the nearest training grid, with noise, classifier-free guidance (CFG), and sampling updates handled in FP32 precision. In practical terms, it performs the image-generation sampling stage while giving you control over the generation settings. Its result is a latent representation, which is not yet a viewable image.
MOSS-Image Official Sampler Input Parameters:
dit
Connect a MOSS_IMAGE_DIT model here. This is the model the sampler uses during generation, so you need a compatible upstream model connection for the node to run. There is no numeric minimum, maximum, or default; this required input must receive the specified connection type.
conditioning
Connect MOSS_IMAGE_CONDITIONING here. This supplies the conditioning information used during sampling, allowing the generation to be guided by the preparation performed upstream. The source does not define the contents of that conditioning, so its exact effect depends on what you connect. There is no numeric minimum, maximum, or default; this required input must receive the specified connection type.
seed
Sets the random starting point for sampling. Reusing a seed is useful when you want to compare changes to other settings without also deliberately changing the starting noise, although identical results may depend on factors outside this node. The default is 42; the minimum is 0 and the maximum is 18446744073709551615. The interface also enables control_after_generate for this parameter, so check its generation control if you intend to keep the seed unchanged between runs.
steps
Sets the number of sampling steps. This controls how many updates the sampler requests; increasing it generally means more work, but the source does not establish a particular quality improvement or ideal value. The default is 50; the minimum is 1 and the maximum is 1000.
cfg
Sets the classifier-free guidance strength used in sampling. It determines how strongly guidance influences the generation; changing it can alter how closely the result follows the supplied conditioning. The default is 3.0; the minimum is 0.0 and the maximum is 30.0, with an interface step of 0.1. The source does not specify a universally best value.
shift
Controls a sampling-schedule setting. The node description identifies a nearest-training-grid Euler schedule, but the provided source does not show how shift changes that schedule internally; treat it as an advanced setting and compare results before adopting a new value. The default is 5.0; the minimum is 0.01 and the maximum is 100.0, with an interface step of 0.1.
width
Sets the requested output width. A larger width requests a wider generation and, when combined with height and batch size, may increase the resources needed to run it. The default is 1024; the minimum is 16 and the maximum is 4096, with an interface step of 16. The provided class does not specify how the runtime handles dimensions that a connected model cannot support.
height
Sets the requested output height. A larger height requests a taller generation and, when combined with width and batch size, may increase the resources needed to run it. The default is 1024; the minimum is 16 and the maximum is 4096, with an interface step of 16. The provided class does not specify how the runtime handles dimensions that a connected model cannot support.
batch_size
Sets how many latent results to request in one sampling run. Use 1 when testing settings; a larger batch requests more results at once and may require more memory. The default and minimum are 1; the maximum is 16.
uncond_skip_layers
Provides a string specifying layers to skip for the unconditional part of sampling. The default string is 22. The node description states that an empty layer list disables skipping, so you can leave this field empty when you do not want that behavior. No minimum, maximum, or list syntax is declared in the provided interface, and the runtime parser is not shown; use only a format you have confirmed your installation accepts.
slg_layers
Provides a string specifying layers for the SLG setting. Its default is the empty string ""; the node description states that empty layer lists disable skipping. The provided class does not explain SLG’s internal behavior or declare a list syntax, minimum, or maximum. Keep it empty unless you have a validated layer configuration for your setup.
slg_scale
Sets the scale associated with the SLG settings. The default and minimum are 0.0; the maximum is 30.0. The source does not show how this scale is applied internally, so change it only when you are intentionally testing an SLG configuration, and compare against the default.
slg_start
Sets a starting value for the SLG settings within the interface’s 0.0 to 1.0 range. The default and minimum are 0.0; the maximum is 1.0. The provided source does not define the precise runtime interpretation of this value, so keep the default unless you have documentation for the connected runtime’s SLG behavior.
slg_stop
Sets a stopping value for the SLG settings within the interface’s 0.0 to 1.0 range. The default and maximum are 1.0; the minimum is 0.0. The provided source does not define the precise runtime interpretation of this value, so keep the default unless you have documentation for the connected runtime’s SLG behavior.
MOSS-Image Official Sampler Output Parameters:
latent
The single MOSS_IMAGE_LATENT output contains the result of the sampling run in latent form. It represents generated image information for use by a compatible downstream node, but it is not itself a displayed or exported image. Connect it to a downstream component that accepts MOSS_IMAGE_LATENT and performs the next step in your workflow.
MOSS-Image Official Sampler Usage Tips:
- Start with the supplied defaults, including
50steps,3.0CFG,5.0shift, and a batch size of1. Change one setting at a time so you can identify what affected the result. - For quick comparisons, keep the seed fixed while adjusting CFG or steps. Check the seed’s after-generation control before each comparison, because the interface enables that control.
- When a run needs too much memory, try reducing
batch_size,width, orheight. Those settings directly control how many results and what dimensions you request. - Leave
slg_layersempty when you are not using a confirmed SLG configuration. To disable the layer skipping described by the node, use an empty layer list rather than the default22inuncond_skip_layers. - Connect both required upstream types before running, and plan for a compatible downstream latent-processing step. This node outputs a latent, not a finished image file.
MOSS-Image Official Sampler Common Errors and Solutions:
Missing required input: dit
- Explanation: The sampler cannot run without its required
MOSS_IMAGE_DITmodel connection. This is a likely interface validation issue; the provided source does not specify the exact error wording. - Solution: Connect a compatible
MOSS_IMAGE_DIToutput toditand run the workflow again.
Missing required input: conditioning
- Explanation: The required
MOSS_IMAGE_CONDITIONINGconnection is absent. This is a likely interface validation issue; the provided source does not specify the exact error wording. - Solution: Connect a compatible
MOSS_IMAGE_CONDITIONINGoutput toconditioning.
Type mismatch on dit or conditioning
- Explanation: An upstream socket with a different type may not connect to, or be accepted by, one of the sampler’s required inputs. The exact message depends on the host.
- Solution: Check the socket types:
ditrequiresMOSS_IMAGE_DIT, andconditioningrequiresMOSS_IMAGE_CONDITIONING.
Value outside the allowed range
- Explanation: The interface declares bounds for its numeric settings. For example,
stepsmust be from1to1000,batch_sizefrom1to16, and bothslg_startandslg_stopfrom0.0to1.0. The host’s exact validation message is not shown in the source. - Solution: Restore the affected setting to its documented range. If you are unsure which value caused the issue, reset recently changed numeric settings to their defaults and try again.
Invalid layer-list value
- Explanation: A nonempty
uncond_skip_layersorslg_layersstring may fail if the runtime does not recognize its format. The provided source does not include the parser or an exact error message, so no particular separator or syntax can be guaranteed. - Solution: Clear the relevant field to disable skipping, or use a layer-list format verified for your installed runtime. For
uncond_skip_layers, the supplied default is22.
Out of memory during sampling
- Explanation: A request with larger dimensions or batch size may exceed the resources available to your setup. This is a possible runtime failure, not an error string specified by the node class.
- Solution: Reduce
batch_sizefirst, then try smallerwidthandheightvalues and rerun.
No viewable image from the latent output
- Explanation: The node returns
MOSS_IMAGE_LATENT, not a finished image. Sampling may have completed even though there is nothing to preview directly from this socket. - Solution: Connect
latentto a compatible downstream node that turns the latent into a viewable result.
