Character appearance (image / text):
MossCharacterAppearance prepares a reusable description of a character’s appearance for MOSS manga workflows, helping you establish a consistent visual reference across different panels. It can preserve a description you write, extract visible appearance details from a reference image using MOSS’s Qwen model, or combine an extracted description with your written additions. Image analysis focuses on relatively stable visual traits such as face shape, build, eyes, hair, accessories, and clothing, while deliberately excluding pose, expression, gaze, composition, background, lighting, and art style. If several characters appear, the model is instructed to describe only the largest foreground character. The main appearance description is requested as one concise English paragraph, and a separate analysis may add an apparent age-group prefix when a recognized age group is returned. This node prepares descriptive text only; it does not generate, edit, or repair images, and it does not itself guarantee that later panels will reproduce the character consistently.
Character appearance (image / text) Input Parameters:
指定方法
Required selection that determines how the appearance description is prepared. The default is 文章のみ, meaning Text only; this mode returns your written appearance description unchanged and does not use the connected model or reference image. The image-extraction mode, 画像から抽出, analyzes the reference image and ignores the written appearance prompt. The combined mode, 画像から抽出+文章で補足, performs the same image analysis and then appends any nonempty written description on a new line after “Additional appearance details: ”. Both image-based modes require a reference image and a compatible MOSS Qwen connection, even though those sockets are declared optional in the interface. Choose text-only mode when you already have an established description, image extraction when you need a description from artwork, or combined mode when you want to add or clarify details yourself; written additions are appended rather than automatically reconciled with the extracted text.
外見プロンプト
Required multiline STRING field with an empty-string default and no explicit length limit defined in the provided source. Enter an English description of the character’s appearance, such as “Short black bob, round glasses, and a dark coat with a wide collar.” In 文章のみ mode, this field must contain something other than whitespace, and its original spacing and line breaks are preserved in the result. In 画像から抽出 mode, this field is ignored. In 画像から抽出+文章で補足 mode, it is optional in practice: nonempty text is added after the extracted description, with leading and trailing whitespace removed, while an empty field simply leaves the extracted description unchanged. Your text is not translated or rewritten by this node, so use clear appearance-focused wording and avoid contradictory additions.
最大トークン数
Required INT setting that controls the maximum generation length for the main image-based appearance description. The default is 256, the minimum is 64, the maximum is 1024, and the interface increment is 32. Tokens are pieces of text rather than a fixed number of words, so this value is not an exact word count. A higher setting gives the model more room to finish detailed descriptions, but can require more generation time; a lower setting restricts the available response length and may leave the description incomplete. This setting has no effect in text-only mode. The separate apparent-age analysis uses its own fixed maximum length of 24 and is not controlled by this parameter.
Qwen
Optional CLIP connection in the interface, with no explicit default value declared; the execution method treats an unconnected socket as None. For either image-based mode, connect the MOSS CLIPLoader configured with its type set to moss_image. Although the socket type is CLIP, the connected object must support the model operations needed to accept an image, generate text, and decode the answer, so an arbitrary CLIP connection is not necessarily suitable. The node uses this model for two separate requests: one for visible appearance and another for apparent age group. It disables sampling and requests no thinking for these analyses. In text-only mode, this connection is not requested or used, allowing you to prepare manually written descriptions without running the model.
参考画像
Optional IMAGE connection in the interface, with no explicit default value declared; the execution method treats an unconnected socket as None. This image is required for either image-based mode and can be supplied by a Load Image node. Supply a single image rather than an image batch: the image-preparation helper rejects batches containing multiple images. Large images are resized while preserving their proportions; the supplied test demonstrates a 1200-by-600 image becoming 768-by-384 before analysis. The image supplies the evidence for both the appearance description and the separate apparent-age assessment. If multiple characters are visible, the model is instructed to select the largest foreground character, so a clear single-character reference is preferable when you need to control which character is described. This connection is ignored in text-only mode.
Character appearance (image / text) Output Parameters:
Appearance for shared characters
STRING containing the completed appearance description, intended for an appearance field in Shared characters, represented by MossMangaCharacters in the supplied code. In text-only mode, this is exactly the text you entered, including its original whitespace and line breaks. In either image-based mode, it contains the model’s extracted appearance paragraph and may begin with “Apparent age group: middle-aged. ” or another recognized age-group label. Recognized groups are child, young teenager, older teenager, young adult, adult in their thirties, middle-aged, and elderly; if the age answer is not recognized, no age prefix is added. In combined mode, any nonempty written supplement follows on a new line beginning with “Additional appearance details: ”. The node also displays this same text in its interface so you can inspect it. Treat an image-derived result as a model interpretation of visible features and apparent age, not as verified facts about the character.
Character appearance (image / text) Usage Tips:
- Start with 文章のみ when you already have a reliable character description. This avoids image analysis entirely, even if Qwen and an image are connected, and preserves your wording exactly.
- For image-based extraction, connect a Load Image node to 参考画像 and a MOSS CLIPLoader with type moss_image to Qwen. A generic CLIP model should not be assumed to support the required image-to-text operations.
- Use a single-character reference with clearly visible hair, face, accessories, and clothing. When several characters are present, the analysis targets the largest foreground character rather than offering a character-selection control.
- Start with 最大トークン数 at its default of 256. Increase it in increments of 32 if extraction stops while reasoning or the appearance description is incomplete; the allowed maximum is 1024.
- Use 画像から抽出+文章で補足 for concise additions such as “Red rectangular glasses and a silver star-shaped brooch.” Review the extracted description first, because additions are appended and do not replace conflicting details.
- Keep reusable descriptions focused on stable appearance rather than a particular panel’s pose, expression, or setting. The image-analysis instruction already excludes those scene-specific details, and following the same approach in manual text makes the description easier to reuse.
- Inspect the displayed result before connecting it to Shared characters. Check that the intended character was described, that important clothing details are present, and that any apparent age-group prefix is appropriate; you can move corrected wording into text-only mode for precise control.
- Provide one image at a time. If your upstream workflow produces a batch, select the intended reference image before connecting it to this node.
Character appearance (image / text) Common Errors and Solutions:
"Method" is "Text only" but "Appearance prompt" is empty. Describe the appearance in English, or choose a method that extracts it from an image.
- Explanation: 文章のみ is selected, but 外見プロンプト is empty or contains only spaces or line breaks. A connected image does not substitute for text in this mode.
- Solution: Enter a nonempty English appearance description, or switch to an image-based method and connect the required image and Qwen model.
"Method" extracts from an image, but "Reference image" is not connected. Connect a "Load Image" node.
- Explanation: An image-based method is selected, but 参考画像 has no image. The node cannot extract appearance from written text alone in these modes.
- Solution: Connect the IMAGE output of a Load Image node to 参考画像. If you only want to use your written description, select 文章のみ instead.
To extract from an image, connect the MOSS CLIPLoader to "Qwen". Set the CLIPLoader type to moss_image.
- Explanation: Qwen is missing or the connected object does not provide the required tokenize, generate, and decode operations. The CLIP socket type alone does not establish compatibility with this workflow.
- Solution: Connect the MOSS CLIPLoader to Qwen and set the loader type to moss_image. If a model is already connected, verify that it is the appropriate MOSS loader configuration rather than another kind of CLIP model.
Appearance extraction stopped while still reasoning. Increase "Max tokens".
- Explanation: The extracted model response begins with an opening thinking marker after leading whitespace is removed, which the node treats as reasoning that did not reach a usable appearance answer.
- Solution: Increase 最大トークン数 above its current setting and retry, up to the supported maximum of 1024. If the error persists, check that your Qwen connection uses the required MOSS configuration.
The extracted appearance is empty. Check the image and Qwen connections.
- Explanation: The model’s extracted answer is empty after whitespace is removed, so there is no appearance description to return.
- Solution: Verify that 参考画像 contains a valid image and Qwen is connected to the correctly configured MOSS CLIPLoader. Retry with a clear single-character reference and inspect the model-loading workflow if the result remains empty.
Multiple-image batch rejected during image preparation
- Explanation: The image-preparation helper rejects a batch containing more than one image. This behavior is confirmed by the supplied tests, but the exact exception message is not included in the provided source excerpts.
- Solution: Select one image from the batch or load a single reference image, then connect that image to 参考画像 and run the node again.
