Krea2 Multi-LoRA Composer:
Krea2CharacterRouter helps you place multiple Krea 2 character LoRAs in one scene while reducing the chance that their identities blend together. It composes a scene-wide prompt with character descriptions, associates each active LoRA with relevant concept tokens, and uses box-guided attention bias to favor the intended character in its designated area. You can also include a Canvas LoRA alongside the character LoRAs. The node prepares a mask preview and a generation plan so you can inspect the layout and carry the configured scene into a downstream workflow. Its goal is more controlled multi-character composition, not a guarantee that every generation will keep identities perfectly separate.
Krea2 Multi-LoRA Composer Input Parameters:
model
The Krea 2 model you want to use for generation. The node loads the configured LoRA adapters against this model, then clones it and installs routing on the clone when adapters are loaded. Connect a compatible model; this input has no default, minimum, or maximum in the node interface.
clip
The CLIP component used to tokenize the composed prompt and produce conditioning for generation. The node also uses it to locate the concept tokens associated with active LoRAs, so compatibility matters for character routing. Connect the CLIP component intended for your Krea 2 workflow; this input has no default, minimum, or maximum in the node interface.
scene_prompt
Your description of the overall scene, such as the setting, interaction, and lighting. The node combines it with enabled character descriptions and, when active, the Canvas LoRA description to form the final positive prompt. This is a multiline string with dynamic prompts enabled. Its default is “Two people in the same scene, natural interaction, coherent lighting”; the interface specifies no length limit.
width
The intended canvas width in pixels. It determines the horizontal size used for region placement and the mask preview, and contributes to the supersampling plan. The default is 1024, the minimum is 256, the maximum is 4096, and the widget changes in steps of 8. This is the target canvas dimension; the working dimensions used to create the latent may differ when supersampling is enabled.
height
The intended canvas height in pixels. It determines the vertical size used for region placement and the mask preview, and contributes to the supersampling plan. The default is 1024, the minimum is 256, the maximum is 4096, and the widget changes in steps of 8. This is the target canvas dimension; supersampling may produce a different working height.
regions_json
A multiline JSON string defining the character regions. The node parses it to determine which characters are enabled, their LoRA selections and strengths, descriptions, placement boxes, and schedules; enabled regions also contribute to the prompt and mask preview. Its default is the built-in DEFAULT_REGIONS_JSON, whose contents are not included in the supplied source, so no specific default characters or JSON layout can be stated here. The interface specifies no length limit; keep the JSON valid and use the node’s supplied default as a starting template.
feather
Controls how softly region masks transition at their edges. A larger value can make boundaries less abrupt, while a smaller value keeps them more defined; the setting is passed to the mask builder and routing configuration. The default is 0.08, the minimum is 0.0, the maximum is 0.5, and the step is 0.01.
overlap_policy
Selects how overlapping character regions are handled when region masks are built and routing is configured. The available options are nearest, normalize, and allow; the default is nearest. Choose a policy deliberately when character boxes overlap, and check the mask preview to see how the configured regions are represented. The supplied source does not define the precise mask calculation for each option.
schedule_softness
Sets the softness of the routing schedule’s transitions. It is passed into the routing configuration, allowing you to adjust how abruptly or gradually scheduled routing changes occur. The default is 0.04, the minimum is 0.0, the maximum is 0.25, and the step is 0.01. The supplied source does not specify a precise time-to-value curve.
strict
Determines whether certain routing problems stop execution or are reported as warnings. When enabled, the node raises an error if there is no active character or Canvas LoRA, if concept-token alignment fails, or if required concept tokens cannot be found. When disabled, the shown token-alignment failures become warnings; adapter loading also receives this setting, but its internal error-handling details are not supplied. The default is True; the options are True and False.
debug
Controls whether the node writes its diagnostics to the application log. Enable it when you need additional information while troubleshooting; the diagnostics output is returned regardless of this setting. The default is False; the options are True and False.
supersample_scale
Requests a supersampling plan based on the target canvas dimensions. At the default of 1.0, no scale increase is requested; higher values request a larger working canvas, which also increases the latent dimensions created by this node. The default is 1.0, the minimum is 1.0, the maximum is 2.0, and the step is 0.05. Consult the returned plan for the actual working dimensions rather than assuming they are an exact multiplication.
canvas_lora_json
A multiline JSON string configuring an optional Canvas LoRA. When its configuration is enabled, names a selected LoRA, and has nonzero strength, the node includes it in routing and prompt composition. The returned share configuration records its trigger, description, strength, coverage, and schedule. Its default is the built-in DEFAULT_CANVAS_LORA_JSON, whose contents are not included in the supplied source; the interface specifies no length limit. Use the supplied default as a template rather than assuming an undocumented JSON schema.
Krea2 Multi-LoRA Composer Output Parameters:
model
A clone of the input MODEL. When adapters have been loaded, the node installs a routing wrapper on this clone so downstream generation can use the configured LoRAs and spatial routing. Connect this output, rather than the original model connection, to the downstream sampling workflow when you want the router’s configuration applied.
conditioning
The CONDITIONING produced when CLIP encodes the composed prompt using scheduled token encoding. It carries the prompt conditioning intended for downstream generation; it is not an image or a diagnostic report.
prompt
The final composed positive prompt as a STRING. It combines the scene description with applicable configured character content and, when active, Canvas LoRA content. Inspect it to confirm that the scene and intended character descriptions are present before sampling.
mask_preview
An IMAGE preview created from masks for the enabled character regions. It helps you inspect the configured spatial layout and its feather and overlap settings before generation. It is a region-mask preview, not a rendered prediction of the final artwork; the source does not establish that an active Canvas LoRA receives its own preview mask here.
diagnostics
A formatted JSON STRING describing the router’s configuration and what happened during setup. It includes the composed prompt, canvas and supersampling details, routed-region information such as token positions and matched adapter counts, attention-bias status, and warnings. Use it to investigate missing token matches, skipped adapter targets, or other configuration concerns.
latent
A LATENT containing a zero-filled starting tensor for downstream sampling. Its spatial dimensions are derived from the supersampling plan’s working width and height, divided by eight, rather than directly from the target canvas dimensions. It is a starting latent, not a generated image.
share_prompt_json
A formatted JSON STRING intended to capture the scene’s shareable configuration. It records the scene and composed prompts, target canvas and supersampling information, router settings, Canvas LoRA configuration, character configurations and placements, schedules, and warnings. Character placement is represented in both normalized coordinates and pixel coordinates based on the target canvas.
target_width
The INT target width returned by the supersampling plan. Use it as the planned final canvas width when coordinating later workflow steps; distinguish it from the potentially larger working width used for supersampling.
target_height
The INT target height returned by the supersampling plan. Use it as the planned final canvas height when coordinating later workflow steps; distinguish it from the potentially larger working height used for supersampling.
supersample_plan
The KREA2_SUPERSAMPLE_PLAN object created from the requested canvas dimensions and supersample scale. It carries the plan used to size the starting latent and reports target and working dimensions through the node’s diagnostics. Pass it to downstream components that explicitly accept this plan type rather than treating it as an image or prompt.
Krea2 Multi-LoRA Composer Usage Tips:
- Start with a compatible Krea 2 model and CLIP pairing, then inspect the returned prompt and diagnostics before sampling. Concept-token matching depends on how the connected CLIP tokenizes the composed prompt.
- Begin with character boxes that have limited overlap and the default
nearestoverlap policy. Checkmask_previewafter changing placements, feather, or overlap policy so you can spot an unintended spatial layout early. - Keep
strictenabled while setting up a workflow. It makes a missing active LoRA or required concept-token match immediately visible; turn it off only when you have a reason to continue and review the resulting warnings. - Leave
supersample_scaleat 1.0 for an initial test. If you increase it, check the plan’s working dimensions and ensure downstream steps use the returned target dimensions where the final canvas size matters. - Use
canvas_lora_jsonfor a Canvas LoRA only when you intend to add one. Its active state requires an enabled configuration, a selected LoRA, and nonzero strength; otherwise it does not join the routing list. - Enable
debugwhen investigating setup behavior, and readdiagnosticsfor token positions, matched adapter counts, skipped targets, attention-bias status, and warnings. Debug logging supplements that output rather than replacing it.
Krea2 Multi-LoRA Composer Common Errors and Solutions:
No enabled character or Canvas LoRA has a selected LoRA
- Explanation: With
strictenabled, the node stops when it finds no active routing entry. An active character must be enabled, have a LoRA other than an empty value orNone, and have nonzero strength; an active Canvas LoRA must meet corresponding conditions. - Solution: In
regions_jsonorcanvas_lora_json, enable an intended entry, select its LoRA, and set a nonzero strength. If you are intentionally testing without active routing, you can disablestrict, but the node will not have an active LoRA to route.
Could not align routed LoRA phrases to Krea2 tokens: …
- Explanation: The token-alignment step raised an
AttributeError,TypeError, orValueError, andstrictconverted it into a stopping error. The text after the colon contains the underlying reason. - Solution: Read the complete error detail, confirm that the connected CLIP component belongs to the intended Krea 2 workflow, and check the active character and Canvas LoRA concept text and triggers. If you disable
strictto inspect the rest of the setup, review the diagnostics warning: token routing is reported as disabled for this failure.
No Krea2 token positions found for: …
- Explanation: The node could not find concept-token positions for the named active character regions, or for a Canvas LoRA whose concept text is required. With
strictenabled, this stops execution; otherwise it becomes a warning. - Solution: Check the listed entries’ descriptions and triggers, inspect the composed
prompt, and confirm the connected CLIP component is appropriate. Keep the identifying phrases clear and present in the composed prompt, then rerun and checktoken_positionsin the diagnostics.
Token routing disabled: …
- Explanation: This warning appears when token alignment fails while
strictis disabled. Execution continues, but the failed alignment means the intended concept-token routing should not be assumed to work for that run. - Solution: Use the detail after the colon to investigate the CLIP connection and configured phrases. Re-enable
strictwhile testing so the underlying failure stops the workflow instead of being easy to overlook.
…: skipped N non-spatial LoRA targets
- Explanation: The diagnostics warning reports that adapter loading skipped the stated number of non-spatial LoRA targets for the named entry. The supplied source does not show the adapter loader’s criteria or establish the visual effect of those skipped targets.
- Solution: Check that you selected the intended LoRA for that entry. Inspect
matched_modulesandskipped_modulesin diagnostics, and use a LoRA suited to the model and routing workflow if the match is not what you expected.
…: skipped N unsupported adapter targets
- Explanation: The diagnostics warning reports unsupported adapter targets skipped during loading for the named entry. Some targets in that LoRA were not used by this routing setup.
- Solution: Confirm that the LoRA is intended for the connected model and this node’s workflow. Review the matched adapter count in diagnostics; if necessary, choose a compatible LoRA rather than assuming every target in the file was applied.
JSON parsing or configuration error in regions_json or canvas_lora_json
- Explanation: Both inputs are parsed before routing begins. The supplied node class does not show the parser’s exact error strings or full schema, so a specific message cannot be guaranteed; malformed JSON or unsupported configuration values may prevent the intended setup.
- Solution: Start from the widget’s supplied default JSON, make one change at a time, and check punctuation such as commas and quotation marks. Preserve the default structure unless you have the configuration schema for the installed version.
Memory or allocation error after increasing canvas size or supersampling
- Explanation: The node creates a starting latent using the supersampling plan’s working dimensions. Larger width, height, or
supersample_scalevalues therefore require a larger allocation; the exact error text depends on the runtime and is not specified in this source. - Solution: Reduce
supersample_scaletoward 1.0 or lower the target width and height, then retry. Check the reported working dimensions in diagnostics to understand the size the workflow is preparing.
