Comfyui-Anima-Wardrobe-V4 Introduction
Comfyui-Anima-Wardrobe-V4 is a prompt-generation extension for ComfyUI, designed around ANIMA3 prompting and the author’s Anima_工作流V3 workflow. It combines character, appearance, clothing, pose, expression, camera, setting, and lighting tags into a complete English prompt.
Instead of repeatedly writing or copying prompts, you can choose a character, set a content level, adjust your preferences, and let the extension fill in the remaining details. You can also apply a prepared scene or close-up composition while keeping other parts of the prompt random.
Current documented release: v4.0.0.
Adult-content notice
The included vocabulary contains adult material, organized into five content levels. Use is restricted to adults who meet the legal age requirement in their region. Do not use the extension to create sexualized depictions of minors, including fictional characters who are minors. The author’s terms also prohibit use involving real people, non-consensual content, or violent harm.
Content-level filtering classifies prompt text, not the finished image. Always review your prompts and generated images before saving, sharing, or publishing them.
What you can use it for
- Explore new ideas: generate combinations of outfits, expressions, environments, and lighting.
- Keep a character while changing the scene: explicitly select your character and randomize the surrounding details.
- Create a series: reuse a saved configuration, a scene recipe, or a restricted character pool.
- Reduce contradictory instructions: the extension checks for incompatible clothing, poses, expressions, camera directions, and character counts.
- Reproduce a prompt: use a fixed seed rather than starting from an entirely new random draw.
- Review existing prompts: dedicated nodes can grade, check, and repair text from other sources.
- Keep better image records: supported PNG metadata includes the generated prompt rather than only generic quality tags. The extension uses Python and native frontend JavaScript without third-party runtime dependencies. It can coexist with Comfyui-Anima-Wardrobe v2.x, because its node classes, routes, and frontend components use separate names.
How Comfyui-Anima-Wardrobe-V4 Works
Think of the extension as a wardrobe stylist, scene planner, and prompt proofreader working together.
1. It starts with organized ingredients
The vocabulary is divided into sections and fields. A section is a broad category, such as clothing; a field is a specific part of that category, such as shoes.
A generated prompt can draw from:
- Character count and identity
- Hair and facial appearance
- Body characteristics and makeup
- Clothing, footwear, and accessories
- Pose, hand position, and action
- Expression, gaze, and mouth state
- Camera viewpoint and framing
- Location, props, and weather
- Lighting and visual details
- Scene recipes and additional text These are assembled as tags rather than written as a long narrative.
2. Your content level narrows the available choices
The extension filters the candidate vocabulary before drawing random options. An option above your selected level should not enter the normal selection pool, even if you give it a high weight.
It then checks the assembled prompt again. This second pass also covers text that did not come through ordinary random selection, such as your handwritten additions.
| Level | Key | General meaning |
|---|---|---|
| 0 | sfw | Non-explicit content using the extension’s safe vocabulary |
| 1 | suggestive | Suggestive clothing, posing, or atmosphere without explicit exposure |
| 2 | explicit | Adult exposure and other vocabulary classified at this level |
| 3 | hardcore | More explicit sexual activity and adult themes |
| 4 | extreme | The highest-intensity adult categories in the library |
| Unrestricted | any | No upper content-level limit |
Important: these are the extension’s classifications, not universal publishing standards. Some tags receive ratings that may seem unexpectedly strict or permissive. Consult the report when a result surprises you.
Later releases also try to make unrestricted random generation at levels 1–4 reach the selected level, rather than merely remain below it. If locks or disabled fields prevent that, the extension reports the limitation. This is a prompt-level check, not a guarantee about the image.
3. It combines randomness with your preferences
Your settings determine which ingredients are available and how likely they are to appear.
For example, you could:
- Select one character explicitly.
- Prefer workplace clothing.
- Make a particular hair color mandatory.
- Disable photographic-device effects.
- Let the setting and lighting change with each draw. The extension also applies compatibility rules. A poolside setting can favor swimwear, while an office can favor workplace clothing. These preferences leave room for variety unless a combination is explicitly blocked.
4. A scene recipe can provide the foundation
A scene recipe fixes a coordinated group of details, such as:
A seated pose, a café location, a particular camera angle, and a matching expression.
The remaining fields can continue to vary. Some recipes also contain their own candidate lists, so the composition stays recognizable while selected details change.
Ordinary scene recipes can appear during random generation. Close-up recipes are manual-only, so an unexpected body-part crop should not replace a normal composition during an ordinary draw.
5. It checks and organizes the result
Before output, the extension:
- Resolves incompatible field combinations.
- Adjusts prompt density toward an appropriate tag count.
- Protects important recipe and character details from ordinary trimming.
- Removes recognized over-level or contradictory tags.
- Normalizes formatting and removes duplicates.
- Orders tags according to the ANIMA3 prompt structure.
- Produces a report describing checks, warnings, and repairs. The final order follows the general progression:
Count → character → appearance → clothing → pose → expression → camera → setting → lighting → details → natural-language additions.
6. A seed lets you repeat the draw
seed = -1: redraw on each execution.seed >= 0: reproduce the prompt with the same settings and data. A seed is like a bookmark for the random selection process. Reproducibility depends on retaining the relevant configuration and vocabulary: changing character records, weights, recipes, or the extension version can change the result.
A repeatable prompt does not by itself guarantee an identical image. Your image-generation workflow has its own model, sampling, and seed settings.
Comfyui-Anima-Wardrobe-V4 Workflow and Nodes
Preview prompts without generating images
You can explore the prompt generator without running an image-generation workflow:
- Open Anima 提示词随机生成(V4).
- Click ⚙ 打开随机面板 — “Open random panel.”
- Choose a content level and character settings.
- Adjust weights or apply a scene recipe.
- Run a preview.
- Read the generated English text, summary, and check report. The panel and workflow node use the same generation engine.
Use the included workflows
The workflows directory contains four examples:
- 29B base model, 8 steps
- 29B base model, 28 steps
- 12B base model, 8 steps
- 12B base model, 28 steps The generator node is already embedded in these examples. Generating images requires the corresponding checkpoint and the other resources expected by that workflow.
Connect it to the author’s ANIMA workflow
For the documented Anima_工作流V3 layout:
- Locate the CR Text node numbered 1004, labeled 全局正文 — “Global body text.”
- Connect the generator’s
textoutput to that node’s unusedtextinput. - Preview your prompt, then run the image workflow. In this documented workflow, that connection supplies the shared body prompt to all six downstream base-model branches. Disconnecting it returns the CR Text node to its manually entered text.
The node number and six-branch behavior describe the author’s workflow, not every ComfyUI graph.
Available nodes
The four nodes appear under the category AnimaWardrobeV3. The category retains its older name even in V4.
| Node | Purpose | Main outputs |
|---|---|---|
| Anima 提示词随机生成(V4) — AnimaV4RandomPrompt | Generates a complete prompt from your settings | text, report, level, cn_summary |
| Anima 提示词分级审查(V4) — AnimaV4PromptGrader | Classifies an incoming prompt and passes its text onward | text, numeric rating, level key, report |
| Anima 提示词拼接台(V4) — AnimaV4PromptAssemble | Combines a main prompt with supplementary text; includes mode and maximum-level controls | text |
| Anima 提示词检查与修复(V4) — AnimaV4PromptCheck | Checks arbitrary prompt text against a target level, with optional automatic repair | Repaired text, report, target level, actual numeric rating |
The generator’s cn_summary output is a Chinese summary. Its text output is the English tag string used by the workflow.
Panel controls and interface language
The node includes action buttons for:
- Opening the random panel
- Changing the seed
- Enabling redraw on every execution
- Clearing weights Most detailed settings live inside the panel rather than appearing twice on the node.
The supplied interface is Chinese, including the strings in locales/en/. An English-language browser therefore does not automatically produce an English interface. The English explanations in this README are translations for reference.
Comfyui-Anima-Wardrobe-V4 Features
Three-level weighting: broad preferences or precise control
Weights operate at three levels:
| Weight layer | What it controls | 0 | 1.0 | 3.0 |
|---|---|---|---|---|
| Section | A broad category | Disables the section | Normal behavior | Multiplies its probability by three |
| Field | One detail within a category | Disables the field | Normal behavior | Makes the field mandatory when otherwise eligible |
| Option | A specific choice within a field | Excludes the option | Normal preference | Gives the option three times the relative preference |
An option weight of 3.0 does not, by itself, guarantee that option.
To require one eligible choice:
- Set its field weight to
3.0. - Leave the desired option enabled.
- Set the other options in that field to
0. For example, to require white hair, make the hair-color field mandatory and exclude the other hair-color options.
Weights cannot bypass content-level filtering. An unavailable higher-level option remains unavailable.
The baseline selection probability is adjustable and defaults to 0.65. Automatic density adjustment can scale ordinary selection probabilities, but a mandatory field remains mandatory when its other requirements are satisfied.
Explicit selections have exceptions to ordinary weight behavior. Selected characters are not canceled by role weights, and a specified body-size override takes priority over automatic appearance choices.
Character browser
The documented built-in library contains 2,383 female character records across 288 works. The repository also includes a user-character overlay with corrections and supplementary records, so the effective browser total can differ.
You can browse using:
- Thumbnail cards
- Chinese and English names
- Popularity ordering
- Work/IP filters
- Keyword searches across names, tags, and works
- Favorites
- Direct selection or filtered random selection Some names have no Chinese translation and appear in English. Missing thumbnails fall back to a colored initial tile.
The library is designed around female characters. Male participants, where a permitted composition uses them, are represented by generic count tags rather than named male characters.
The browser is not an age-verification system. Independently verify that every character used in adult material is an adult.
Fixed characters versus an automatic character pool
These two controls serve different purposes:
| Mode | What happens | Best for |
|---|---|---|
| Select cards / draw once | Characters are added to selected_roles and remain selected | A consistent character or cast |
| Add works to the random pool | role_works restricts automatic selection to specified works | A series that changes characters between runs |
Character-source priority is:
-
Manually entered character:
force_role -
Selected character cards:
selected_roles -
Restricted work pool:
role_works -
Unrestricted automatic selection To change characters automatically:
-
Clear manually entered characters.
-
Clear selected cards.
-
Add one or more works to the random pool.
-
Use
seed = -1. The work pool accepts recognized slugs, English names, and Chinese names. Multiple works can be separated by|or commas.
Unknown work names produce report warnings. If the resulting pool is empty, the extension can fall back to the full library and explain why.
Selection is genuinely random: consecutive repeats are possible.
Character count and multi-character behavior
When you select characters explicitly, the extension adjusts count tags to match:
| Selected characters | General behavior | |---|---| | One | Single-character composition | | Two | Two-character count tags, with level-dependent wording | | Three or more | Group composition |
When characters are chosen automatically, the count combination determines how many female character identities are drawn.
For multiple named characters:
- Single-character appearance alignment is skipped.
- The extension avoids applying one character’s hair or eye color to everyone.
- Incompatible individual or one-to-one action fields are filtered according to the current rules.
Scene recipes can also declare a required count through
count_lock. Some later-release limitations remain; see troubleshooting.
Appearance alignment and fixed character features
Appearance alignment uses available character data to guide supported appearance fields.
Later releases also include a dedicated character-feature table for selected characters’ recognizable details, such as:
- Pointed ears
- Distinctive hats
- Horns or tails
- Special pupils
- Character-specific accessories
This fixed-feature behavior follows
appearance_followand is skipped for multi-character alignment.
Not every character has complete data. Missing information may remain random rather than being guessed.
Editing a character’s ordinary feature-tag list does not mean every tag will appear in the prompt. That list primarily supplies supported body-size information and fallback hair/eye colors. Recognizable accessories and similar special features use the separate fixed-feature system.
Current hairstyle selectors
The current interface uses four hairstyle dropdowns introduced in v3.47.0.
Each offers:
- Character-tag hairstyle: use available official hairstyle data.
- Random: let that aspect follow normal random selection.
- A specific vocabulary option: explicitly choose that aspect. Options display Chinese text alongside the English tag. Older workflows containing plain English values remain supported.
This lets you keep a character’s appearance while intentionally changing part of the hairstyle. Explicit selections are protected against several older behaviors that previously removed or contradicted them.
The older hair_follow checkbox remains stored for compatibility but is hidden in the current interface. Hairstyle alignment is separate from general appearance alignment.
Body-size override
The former large-breast checkbox has been replaced by a size selector with:
- Unspecified
- Flat
- Small
- Medium
- Large
- Huge
- Gigantic Unspecified leaves the field to normal random and character-alignment behavior. A specific choice overrides the corresponding automatic body-size selection, including official character data and ordinary weights.
The legacy big_breasts value remains supported for older workflows but is hidden in the current interface.
Scene recipes
The 🎬 场景搭配 tab provides coordinated scene recipes rather than unrelated individual tags.
You can:
- Search by name, theme, description, field value, or scene ID
- Filter by theme or pose-focused recipes
- Collapse content-level groups
- Expand cards to inspect their ingredients
- Preview a recipe without applying it
- Apply it to the current configuration
- Apply and immediately preview it
- Cancel the applied recipe
- Favorite frequently used recipes
- Use a compact browsing layout Preview is non-destructive: it lets you inspect a temporary result without replacing the current weight configuration.
Apply fixes the recipe’s declared fields. Undeclared fields remain available for random generation unless the recipe explicitly disables them.
The documentation records 200 ordinary scene recipes in its later feature sections, alongside 62 close-up recipes. Some older sections contain obsolete totals; use the current panel’s counts as the practical reference.
Scene variation
Recipes can include:
- Fixed fields: a value stays constant.
random_fields: a field draws from a restricted list of vocabulary options.random_variants: a recipe draws additional phrases from one or more variation pools.drop_fields: named fields are disabled to prevent interference.count_lock: a declared participant count. For example, a recipe can keep a seated composition while varying expression and footwear.
Automatic scene selection
Ordinary recipes, including pose-focused recipes, can appear during random generation.
For levels 1–4, later releases target roughly 60% scene-recipe presence under the documented automatic-selection conditions and favor recipes matching the selected level. This is a statistical target, not an exact quota for a small batch.
To stop ordinary recipe selection, set the n_canon field weight to 0.
Close-up recipes
The 🔍 特写搭配 tab organizes compositions by visible area, such as face, eyes and lips, upper body, hands, legs, feet, and other close-up groups.
These recipes are manual-only.
Their central rule is:
Describe what belongs inside the crop, not everything about the whole character.
For a face close-up, distant footwear and full-body actions are suppressed. For a leg close-up, face and hair details may be disabled and the character identity temporarily cleared to avoid pulling the framing back toward a portrait.
Cards indicate whether they:
- Tighten framing
- Clear character identity
- Retain character identity Changing or canceling the recipe restores the saved configuration baseline. That baseline is stored with the workflow, rather than only in temporary panel memory.
Pose-focused recipes
Pose recipes provide two approaches:
- Pose-first recipes: fix the pose and supporting camera or expression details while leaving the environment open.
- Complete-look recipes: combine the pose with a coordinated location, lighting, and clothing direction. The documented v3.22.0 expansion brought this group to 36 recipes. Later releases added or reorganized further entries.
Clothing coordination and color
Clothing coordination helps avoid unrelated layers and incompatible combinations.
The outfit_style control offers:
- Random: coordination enabled, with the style inferred automatically
- Unrestricted: coordination disabled for pure random behavior
- Specific styles: school, workplace, traditional, sports, daily, home, nightlife, fantasy, and swimwear When inferring a style, the extension generally considers:
Setting → complete outfit → top → bottom.
A complete outfit also reduces the probability of unnecessary additional tops and bottoms, helping avoid combinations such as a dress, shirt, and trousers all appearing together.
Independent clothing and legwear colors
| Field | Controls |
|---|---|
| cloth_color | Bottoms, tops, or complete outfits |
| sock_color | Socks and hosiery |
The general color system uses twelve basic colors:
black, white, grey, red, orange, yellow, green, aqua, blue, purple, pink, and brown.
Colors are attached to compatible garment tags, not emitted as loose color words. If the garment already contains a color, or a supported combination is unavailable, coloring can be skipped with a reason.
Separate curated vocabulary entries include additional named hosiery colors and finishes. Those are not the same as extending the general twelve-color selector.
Compatibility checks and automatic repair
The extension checks combinations including:
- Incompatible character counts
- Conflicting hand instructions
- Contradictory expressions
- Clothing that conflicts with the selected state
- Footwear or legwear contradictions
- Opposing camera directions
- Face-focused instructions in inappropriate crops
- Device effects that conflict with photographic depth-of-field instructions
- Activities that do not fit the declared participant count Automatic repair also normalizes underscores, removes duplicates, and filters recognized over-level tags.
Repairs are recorded in the report. Unknown handwritten phrases are not necessarily removed merely because they are long or absent from the vocabulary.
These checks improve text consistency; they cannot guarantee correct anatomy or composition.
Prompt length and density
Automatic tag limits are based on composition complexity:
- Single character: approximately 29 tags
- Two characters: approximately 37 tags
- Complex compositions: approximately 45 tags Rather than only deleting a large number of details afterward, the engine retries with adjusted selection density. Important recipe fields and supported aligned hairstyle details receive trimming protection.
Favorites, saved configurations, and panel memory
These features store different things:
| Feature | Purpose | Storage |
|---|---|---|
| Character favorites | Characters you frequently browse or use | presets/user_favorites.json |
| Scene favorites | Favorite scenes and close-ups | presets/user_scene_favorites.json |
| Named presets | Reusable prompt-generation configurations | presets/user_presets.json |
| Selected characters | The cast for this workflow | Workflow JSON |
| Panel memory | Last tab, filters, scrolling, and expanded groups | Workflow settings |
Favoriting a character does not select it for generation.
Named presets save configuration rather than one particular random draw. They intentionally do not save the seed. Save your workflow and fixed seed when you need to preserve an exact prompt-generation setup.
Character editing and custom characters
Use ✎ on a card to edit a character, or + 添加角色 to add one.
Supported information includes:
- Prompt trigger
- Chinese display name
- Work/IP
- Preview image
- Feature tags Later releases also support hair and eye-color overrides in the user-character data layer.
Key limits:
- Trigger: lowercase English, no more than 120 characters
- Preview: PNG, JPG, or WEBP, up to 2 MB
- Custom characters: up to 500
- Feature tags: up to 40, each no more than 40 characters The internal character key cannot be edited like a display name. It identifies the record and is used by saved selections.
Ordinary edits are stored separately from the generated built-in library in presets/user_characters.json. Uploaded previews use presets/user_thumbs/.
Built-in and custom characters can be deleted, including through batch selection. Built-in deletions are recorded in an archive, but rebuilding the built-in character library can restore those records.
PNG prompt metadata
The extension writes prompt information into newly generated PNG metadata through the supported workflow path.
V4 provides:
anima_v4_prompt- A legacy
anima_v3_prompttext block for compatibility The dedicated metadata includes the generated body prompt, a best-effort full positive prompt, and available paired negative-prompt information.
Limitations:
- It affects new images, not previously saved PNGs.
- Runtime LoRA trigger text supplied by another extension may be missing from the reconstructed full prompt.
- Different image viewers inspect different metadata locations.
- Metadata handling does not change the actual execution connections used to generate the image.
Comfyui-Anima-Wardrobe-V4 Models and Example Variants
The extension is not an image model and does not provide its own image-generation checkpoint. It generates text for a downstream image workflow.
The supplied examples cover these configurations:
| Example | Base-model size | Sampling steps | |---|---|---| | 29B / 8-step workflow | 29B | 8 | | 29B / 28-step workflow | 29B | 28 | | 12B / 8-step workflow | 12B | 8 | | 12B / 28-step workflow | 12B | 28 |
Use the checkpoint required by the selected workflow. The provided context does not establish a complete checkpoint download list or a universal quality and speed comparison between the 12B and 29B options.
The author’s testing notes repeatedly identify 8-step turbo generation as a limitation for difficult poses, small facial expressions, mirrors, and extreme perspective. For those compositions, the author recommends checking results at approximately 20–30 steps in an appropriate workflow.
Prompt previewing does not require image generation or a loaded image checkpoint.
What’s New with Comfyui-Anima-Wardrobe-V4
The following history summarizes the documented releases. Later fixes supersede older behavior, terminology, counts, and implementation notes.
Current release and recent controls
| Version | Changes and practical benefit |
|---|---|
| v4.0.0 | Renamed the extension and its modules, data, routes, and metadata to V4. V3 node-class aliases, routes, and PNG metadata remain supported. Prompt-generation behavior is unchanged from v3.47.1. |
| v3.47.1 | Added Chinese/English hairstyle option labels. Hid retired big_breasts and hair_follow widgets while retaining saved values and widget positions for old workflows. |
| v3.47.0 | Added four hairstyle dropdowns. Fixed cases where explicit hairstyle choices were removed by character-feature suppression or hair incompatibility rules. |
| v3.46.2 | Added weights_json.forced support at the node layer, allowing direct scene-ID selection. Documented a remaining count-lock issue in some highest-level scenes. |
| v3.46.1 | Fixed shared-trigger lookup selecting a variant instead of the base character. Added missing appearance data and corrected several character details and names. |
| v3.46.0 | Added the dedicated fixed-character-feature table for recognizable accessories and traits, following appearance alignment. |
| v3.45.0 | Introduced official hairstyle alignment, independent from general appearance alignment. Later replaced in the visible interface by the hairstyle dropdowns. |
| v3.44.4 | Enabled hair and eye-color overrides in user-character data, addressing cases where editing feature tags alone could not correct appearance. |
Browsing, legwear, and composition refinements
| Version | Changes and practical benefit | |---|---| | v3.44.3 | Reworked a mirror-based scene into a hallway composition and replaced several reflection viewpoints to reduce duplicate bodies and anatomy problems. | | v3.44.2 | Expanded dedicated hosiery-scene candidate lists and increased variety in glittery finishes. | | v3.44.1 | Added ten hosiery-focused scenes with coordinated shoes and optional thigh accessories. | | v3.44.0 | Made complete theme and body-part filter bars collapsible while keeping the active filter visible. | | v3.43.0 | Added scene search, separate scene favorites, theme filters, collapsible level groups, incremental loading, and compact mode. | | v3.42.26 | Reclassified two pose scenes from L2 to L1 after revising their vocabulary and suppressing conflicting hand instructions. | | v3.42.25 | Added ten glittery thigh-high and garter variations to the global legwear vocabulary. | | v3.42.24 | Refined pale hosiery colors, added two colors, and synchronized scene candidates with the global vocabulary. | | v3.42.23 | Added three glittery pantyhose options to the global vocabulary. | | v3.42.22 | Added a foot-forward, low-angle scene and refined perspective, footwear suppression, and legwear variation. | | v3.42.21 | Added a low-angle pose scene with stronger expression control and suppression of interfering actions, props, and accessories. |
Scene consistency and level matching
| Version | Changes and practical benefit | |---|---| | v3.42.11 | Revised four adult scenes for clearer viewpoints, supported poses, and independently randomized details. | | v3.42.9 | Added one adult scene and corrected clothing-state and expression conflicts in existing entries. | | v3.42.8 | Strengthened expression consistency in ten highest-level scenes by suppressing competing gaze and emotion fields. | | v3.42.7 | Clarified held props and hand appearance in adult scenes and replaced three recipes. | | v3.42.6 | Simplified several crowded adult compositions, clarified prop ownership, and reduced unnecessary visual instructions. | | v3.42.2 | Added ten highest-level adult scene recipes. Several were subsequently revised for stability. | | v3.42.1 | Improved the scene-selection retry process to approach the documented 60% presence target more consistently. | | v3.42.0 | Restricted automatic L1–L4 scene choices to the selected level and introduced the 60% scene-presence target. | | v3.41.2 | Prevented partner-dependent actions in established solo compositions and protected recipe extras from trimming. | | v3.41.1 | Prevented group-only tags from appearing in non-group compositions. | | v3.41.0 | Added selected-level attainment checks and refined multi-character filtering to avoid unnecessarily removing valid content. | | v3.40.0 | Introduced level-layered scene selection so higher-level scenes were not overwhelmed by larger lower-level pools. Later releases tightened matching further. | | v3.39.0 | Added a shared hand-action compatibility system and an additional check for conflicting upper-body instructions. |
Recipe and viewpoint development
| Version | Changes and practical benefit | |---|---| | v3.37.0 | Refined an adult recipe’s footwear, clothing, pose, and action consistency. | | v3.36.0 | Clarified required props and action wording in an adult recipe. | | v3.35.1 | Revised blanket placement and framing in an adult recipe. | | v3.35.0 | Changed that recipe to a first-person composition and suppressed conflicting fields. | | v3.34.1 | Expanded a solo recipe’s framing and poses; fixed trimming that removed recipe-selected random fields. | | v3.34.0 | Added variation to one adult recipe and changed another to a clearly solo composition. | | v3.33.0 | Expanded self-view pose and hand variation and corrected a separate recipe’s solo count. | | v3.32.0 | Strengthened true self-view framing, added scene-level character suppression, and reduced unwanted camera-interface imagery. | | v3.31.0 | Removed mirror instructions from an inspection recipe and changed it to a side view. | | v3.30.0 | Clarified that recipe’s hand placement and restricted poses to compatible choices. | | v3.29.0 | Separated third-person overhead framing from first-person self-view framing. | | v3.28.0 | Removed camera-object and recording-interface cues from a candid-view recipe; added viewpoint variation. | | v3.27.1 | Added an empty-handed option to a nurse-themed recipe and improved hand-support consistency. | | v3.27.0 | Added independent variation pools, allowing pose and prop choices to vary separately; corrected unrelated props in the nurse recipe. | | v3.26.0 | Removed duplicate-umbrella cues from a rain recipe and stabilized its pose and hand direction. | | v3.25.0 | Added per-node memory for the panel’s tab, scrolling, filters, and expanded groups. | | v3.24.0 | Introduced phrase-based recipe variants and prevented density retries from unfairly favoring one variation. | | v3.23.0 | Added optional upward-only content-level synchronization when manually applying recipes; revised numerous counts, viewpoints, and compositions. | | v3.22.3 | Required explicit counts for ordinary scene recipes, generally using one female character where additional participants were unnecessary. | | v3.22.1 | Connected recipe count locks and field suppression to the engine, fixing many missing scene-defining tags. | | v3.22.0 | Added restricted random fields and field suppression, improved a difficult close-up composition, and expanded pose recipes to 36. |
Close-ups, character control, and vocabulary
| Version | Changes and practical benefit | |---|---| | v3.18.6 | Removed work names from most emitted character prompts while retaining them for ambiguous names. Added built-in character deletion and batch deletion. | | v3.18.5 | Persisted the close-up restoration baseline to fix disabled clothing fields remaining disabled after refresh. Expanded close-up recipes to 62. | | v3.18.4 | Added visible role-lock diagnosis and automatic unlocking for character-preserving close-ups; expanded two-character close-ups. | | v3.18.3 | Distinguished character-preserving close-ups from identity-clearing crops and added recipe count handling. | | v3.18.2 | Added work-restricted character sourcing to close-ups and expanded leg and adult close-up groups. | | v3.18.1 | Introduced visible-field allowlists and tighter body-part framing, plus more complete restoration when canceling. | | v3.18.0 | Introduced the separate manual-only close-up tab with 32 initial recipes. | | v3.17.2 | Added ten camera-and-pose experiment recipes. | | v3.17.1 | Expanded suggestive recipes across everyday, summer, professional, and fitness themes. | | v3.17.0 | Replaced the legacy body-size checkbox with a multi-choice selector and preserved backward compatibility. | | v3.16.0 | Unified calibrated rating behavior between reports and prompt assembly while retaining raw ratings for diagnosis. | | v3.15.1 | Added the standing-split pose. | | v3.15.0 | Added action, expression, and setting compatibility; expanded scenes and expressions; corrected invalid tags and level handling. | | v3.14.0 | Expanded poses and added suggestive poses. Fixed mixed-level sections being incorrectly disabled at lower levels. | | v3.13.0 | Improved playable-character filtering and coverage for targeted games, reducing unwanted NPC and non-character entries. | | v3.12.0 | Expanded clothing, footwear, accessories, tops, and bottoms; added curated character forms and simplified-Chinese name normalization. |
Foundation releases
| Version | Changes and practical benefit | |---|---| | v3.11.0 | Added 22 adult-theme recipes, collapsible scene cards, and non-destructive recipe previews. The theme was later renamed in V4. | | v3.10.0 | Added work-restricted automatic character pools and content-level scene groups. | | v3.9.0 | Introduced coordinated scene recipes and protected their fields from trimming; prevented internal recipe IDs from entering prompts. | | v3.8.0 | Added camera options, final-text repair, the fourth check-and-repair node, filtered character draws, and preview/workflow consistency fixes. | | v3.7.0 | Improved image-viewer compatibility by writing prompt text into the upstream concatenation metadata actually inspected by the tested viewer. | | v3.6.0 | Added PNG prompt-metadata injection. | | v3.5.0 | Added full feature-tag editing and on-demand character-detail retrieval. | | v3.4.0 | Added editable characters, custom characters, and a separate user-data layer. | | v3.3.0 | Fixed automatic character counts and introduced the legacy body-size checkbox. | | v3.2.0 | Unified character selection in the browser; added favorites, named presets, and the disclaimer panel. | | v3.1.0 | Introduced the expanded character browser, work filtering, search, thumbnails, multiselection, and Chinese interface text. | | v3.0.0 | Initial release with fifteen vocabulary sections, five-level filtering, three-level weights, three nodes, and the panel. |
Troubleshooting Comfyui-Anima-Wardrobe-V4
My character does not change between runs
- Clear
force_role, if you entered a character manually. - Clear selected characters in the browser.
- Add your desired works to the automatic random pool.
- Set
seedto-1. - Run again and inspect the report. Selected cards are fixed choices, even if you originally selected them through a random-draw button.
An occasional repeated automatic draw is normal; there is no no-repeat tracking.
Random characters or work-restricted selection do not work
- Open the close-up tab.
- Look for a warning that
roleis locked. - If character identity is appropriate for the crop, use 🔓 解除 role 锁 — “Unlock role.”
- Check for a manually entered or already selected character taking priority.
- Check the report for unrecognized work names or an empty-pool fallback. A faceless close-up intentionally locks or clears character identity. This is expected for those recipes.
A detail never appears
- Confirm that the relevant section is enabled.
- Check that its field weight is not
0. - Check that the desired option is not weighted
0. - Confirm that it is allowed at the current content level.
- Cancel any close-up or scene that disables the field.
- Check whether the field is locked.
- Read the report for compatibility removal or trimming. To require one eligible option, make the field mandatory and disable its competitors.
A selected option has weight 3.0, but it is not guaranteed
3.0 has different meanings at different layers:
- Field: mandatory selection, subject to eligibility and other constraints.
- Option: stronger relative preference.
Set the field to
3.0and competing options to0. For scene recipes, use the panel’s Apply action instead of relying on an option weight alone.
A specific scene ID is not selected through custom weights
For manual scene selection, use the scene card’s Apply button.
For workflows already using direct JSON control, v3.46.2 added:
json { "forced": { "n_canon": "sc_your_existing_scene_id" } }
Replace the example with an actual recipe ID from your current library. This still does not bypass level eligibility or other constraints.
Socks, shoes, or clothing disappeared after a close-up
- Cancel the currently applied close-up.
- Apply the intended composition again.
- Inspect the relevant clothing field weights.
- Check for weights left at
0. - If you no longer need your custom weights, use the clear-weights action and reconfigure.
Earlier releases could lose the restoration baseline after refreshing. Current releases store it in
weights_json._nearBase, so updated workflows should restore settings more reliably.
Hair or eye color is incorrect
- Check that appearance alignment is enabled.
- Test with one character, since multi-character alignment is intentionally skipped.
- Confirm that the character record contains usable appearance data.
- Check whether multiple variants share the trigger; v3.46.1 fixed a lookup issue affecting this.
- Remember that changing ordinary tags may not override a dedicated hair/eye field or the higher-priority appearance table. Later releases support hair/eye overrides in user-character data. Some characters still require a dedicated fixed-feature correction where a higher-priority appearance record applies.
The hairstyle does not follow the character
- Use the current hairstyle dropdowns.
- Select Character-tag hairstyle for the aspects you want aligned.
- Confirm that the character has corresponding hairstyle data.
- Check for a field weight of
0. - Test with a single character. Missing hairstyle data remains random. An explicit hairstyle choice can also intentionally override the official style.
My close-up becomes a portrait or full-body image
- Apply a recipe from the close-up tab.
- Confirm that its framing-tightening behavior is active.
- Inspect extra text for full-body poses, footwear, facial descriptions, or other details outside the intended crop.
- Check whether the recipe is supposed to retain character identity.
- Try a more capable sampling configuration for difficult framing. Even a consistent prompt may not be followed perfectly by the image model, especially in an 8-step workflow.
The report shows a different rating than another tool
The extension uses calibrated ratings in addition to the underlying raw classifier. For example, wet-hair wording has a documented calibration because a broad “wet” rule can otherwise overrate it.
- Compare the report’s calibrated level with its raw level where available.
- Review the tags identified by each tool.
- Use AnimaV4PromptCheck at your intended target level.
- Review the image independently. A rating difference is not proof that either system has inspected the final image or verified publishing compliance.
The selected content level is not reached
- Read the report for fields that were locked or disabled.
- Check whether the active recipe suppresses the necessary vocabulary.
- Review the participant count and compatibility restrictions.
- Remove handwritten text that conflicts with the intended composition.
- Distinguish a correctly rated prompt from a model that did not render its instructions. Later releases attempt to add a compatible level-defining tag, but cannot do so when every suitable route is blocked.
A scene produces the wrong number of people
- Inspect the generated count tags.
- Check selected characters and manual character overrides.
- Review the recipe’s declared count and the report.
- Test another seed to separate a text-generation issue from an image-model interpretation issue. The latest documented history still notes count-lock failures in some L4 recipes. Count control should not be treated as infallible.
PNG metadata shows only quality tags
- Generate a new PNG using the supported workflow and metadata-preserving save path.
- Inspect the new file, not an older image.
- Look for
anima_v4_promptor the legacyanima_v3_prompttext block. - Try a viewer that can inspect ComfyUI workflow metadata or custom PNG text fields. Some viewers read only particular workflow widgets. Also, the reconstructed “full” positive prompt may omit runtime LoRA trigger text.
Mirror or extreme-perspective images have anatomy problems
The author documented instability with mirror reflections, crowded bodies, and extreme perspectives in 8-step turbo workflows.
Try:
- Using a mirror as a background prop rather than a reflection viewpoint.
- Simplifying the composition.
- Choosing a tighter crop where appropriate.
- Testing at approximately 20–30 steps in a compatible workflow. More steps are a troubleshooting option, not a guarantee.
My custom character or preview image is rejected
Check the documented limits:
- Lowercase English trigger, no Chinese characters, at most 120 characters
- PNG, JPG, or WEBP preview
- Image size at most 2 MB
- At most 500 custom characters
- At most 40 feature tags, each at most 40 characters
Frequently asked questions
Can I use the extension only for non-explicit art?
Yes. Select sfw, review the generated prompt, and inspect the image. The vocabulary package itself still includes adult material, and the author restricts use to adults.
Does generation require an internet connection? Prompt generation runs locally. Separate character-library rebuilding and thumbnail-fetching tools access external sources; those are not the normal prompt-generation process.
Does a custom character entry teach the model a new identity? No. It stores prompt and browser information. The extension does not train an image model, so a custom name alone does not guarantee recognition.
Will a favorite automatically appear in my image? No. A favorite is a browsing bookmark. Select the character card to make it part of the current cast.
Why are work names missing from character prompts? Most are intentionally omitted. Work identifiers are retained when needed to distinguish characters with the same name.
Can old V3 workflows still open? V4 retains V3 class aliases and other compatibility paths. Very old workflows may still require the historical control-layout migration appropriate to their original version.
Why do older module names and category labels remain? Some names are retained for compatibility. Their presence does not necessarily mean the V4 extension failed to load.
Learn More about Comfyui-Anima-Wardrobe-V4
The supplied context does not include a dedicated tutorial site, support forum, or demonstration video. The repository resources below are the documented starting points.
| Resource | What you can learn | |---|---| | Example workflows | How the generator fits into the supplied 12B/29B and 8-step/28-step image workflows | | Preset and character data | The vocabulary, character records, scene recipes, thumbnails, and stored user-data locations | | Documentation directory | Additional design material, including scene-planning references | | Maintainer guide | Data structure, maintenance rules, and regeneration details for advanced customization | | Tests | The behaviors checked by the author, including levels, weights, roles, close-ups, and metadata | | Tools | Advanced data preparation, validation, and workflow-maintenance utilities |
Back up your creative settings
Before replacing plugin files or sharing a customized setup, preserve the relevant user data:
presets/user_characters.jsonpresets/user_thumbs/presets/user_favorites.jsonpresets/user_scene_favorites.jsonpresets/user_presets.json- Your workflow JSON files
The repository’s
user_characters.jsonalso contains supplied corrections and additions. Treat it as an important data file rather than an empty preferences file.
Exclude personal uploaded thumbnails from public sharing packages unless you intentionally want to distribute them.
Attribution and responsible use
The author states that the extension’s underlying framework derives from the authorized open-source script ANIMA提示词生成器.html.
Character and work names remain the intellectual property of their respective rights holders. Character previews are provided for identification; their inclusion does not grant rights to redistribute or commercially reuse them.
Prompt generation is local, but output remains your responsibility. Review each result, verify character adulthood for any adult material, and follow applicable laws and platform policies.
