Navigate to my HF dataset, download the following files to the data folder and rename them as follows:

waiIllustriousSDXL_v160

waiIllustriousSDXL_v160_characters.csv

waiIllustriousSDXL_v160_tag_assist.json

waiIllustriousSDXL_v120

waiNSFWIllustrious_v120_characters.csv

waiNSFWIllustrious_v120_tag_assist.json

waiANIMA_v10Base10

waiANIMA_v10Base10_characters.csv

waiANIMA_v10Base10_tag_assist.json

Template chooses how the reply is written. Source chooses which text is sent. Picking a template also sets that template's source and character slot. Change either one afterwards and it is stored on the template.

Three templates are included:

• Tag scene reads Brief and answers with one line of English comma-separated tags, about 8 to 16 tags. Gender words are left out.
• Natural scene reads Brief and answers with one English paragraph for diffusion models. Width and height stay with the app, so the paragraph does not contain an aspect ratio.
• Character natural reads character slot 4. It polishes the text it actually received. The usual reply is one English paragraph. If that text explicitly asks for tags, the reply is one line of tags. A place, time, light, weather, background, prop, or extra person is added only when the source text already has it.

Generate runs the request. After a short pause the button becomes Cancel and can stop the request. The reply stays in Result until you choose where it goes. Overwrite replaces Common, Positive, Right positive, or Negative. Append adds the reply to the end of that box. Character files are left unchanged.

Each template keeps its own system prompt, model, timeout, temperature, max tokens, source, and character slot. Model settings and the system prompt start folded. The save icon asks for a name. The same name updates the current template. A new name saves a copy and selects it. Delete asks first. Deleting the last template restores the three original templates. Restore defaults, under Backend, adds a missing original template and replaces an edited original with the stored one. Templates you added yourself stay.

Backend is shared by every template.

• Remote expects an OpenAI-compatible base URL and an API key. A base such as https://nano-gpt.com/api/v1 is called at /chat/completions. A URL that already ends with /chat/completions is used as it is.
• Local expects a llama.cpp server address, for example http://127.0.0.1:8080/chat/completions. Build llama.cpp or take a release, then start it with your GGUF model: llama-server.exe -c 16384 --port 8080 -m "".

Drag on empty space to create a box. When the canvas is crowded, Add inserts a default centered rectangle (with a small nudge if that spot is already taken). Drag inside a box to move it, and drag an edge to resize it. A selected box is solid; the others stay semi-transparent. Boxes are numbered from 01, up to 128. Deleting one renumbers the rest from 01 in list order, and a new box is appended at the end.

Click selects the smallest box under the pointer. The area outside that box still selects the larger one. After the canvas has focus, Tab walks outward through the nested boxes under the pointer, and Shift+Tab walks back. If the pointer is not over two or more boxes, Tab uses the boxes around the selected box, then the full list in number order. Enter or a double-click edits the description. Inside the editor, Enter saves, Shift+Enter adds a line, and Esc cancels. Delete removes the selection.

Each box is object or text. object is the default. For Moss, object generates the description as an item inside the box, and text generates the description as annotation text inside the box.

• An empty object sends its number as caption and has no text field.
• An object with a description sends that item text as caption and has no text field. In the example below, burger and price list are generated as items.
• A text box keeps the number as caption and sends the description in text, so Moss generates that description as annotation text.

A long description wraps to the box width. Lines that do not fit stay hidden. The editor opened by Enter or a double-click grows with the text and stays inside the canvas.

Plate loads the image currently shown in the gallery. Image Info can also send its loaded picture with To Bounding, under the tagger row. In the split view that is the large image. In grid view it is the opened image, or the newest image when none is opened. Click Plate again to clear it. A matching aspect ratio scales to the canvas. A different ratio scales by the short side and stays centered, and the canvas clips the overflow. Opacity runs from 0 to 100 and stays disabled until a plate is loaded. The opacity is saved with settings. The image itself lasts for the current session. While a plate is loaded, box fills follow that slider and thin out as the image gets stronger, so the picture, the region edge, and the description stay visible together.

Lock keeps the selected box from being moved or resized. A locked box can still be selected, recolored, and edited, and Delete still removes it. Add sits next to Clear. Clear asks first. It removes unlocked boxes and leaves locked ones. When every remaining box is locked, Clear removes those too.

Coordinates use the full generation pixels. The origin is the canvas center, x points right, y points down, and each box is [y_min, x_min, y_max, x_max]. On a 1024×1360 canvas the packed prompt looks like this:

{
  "summary": {
    "caption": "japanese anime style, solo anthro, aged down, fluffy,\nrestaurant, huge burger,two-tone fur, (grey fur:1.1), white cheek fur, white paws, grey wolf ears, grey wolf tail, blonde hair, grey eyes, wolf girl, hair between eyes, long sidelocks, short hair, (fluffy, chest fluff:1.4), cheek fluff, furry, petite, fangs, ear fluff, claws, sitting,"
  },
  "items": [
    { "label": "object", "caption": "furry", "bbox": [-680, -96, 680, 512] },
    { "label": "object", "caption": "burger", "bbox": [-238, -173, 101, 182] },
    { "label": "object", "caption": "price list", "bbox": [-619, -480, -234, -224] },
    { "label": "object", "caption": "table", "bbox": [124, -512, 363, 224] },
    { "label": "object", "caption": "soda drink", "bbox": [-69, -313, 168, -221] },
    { "label": "object", "caption": "Seasoning", "bbox": [98, -484, 166, -320] },
    { "label": "object", "caption": "Dinner Plate", "bbox": [204, -277, 273, 179] }
  ],
  "bbox_coord_space": "abs_centralized",
  "bbox_format": "yxyx"
}

summary.caption is the merged positive prompt after replacement and AI. The JSON is written onto the positive prompt that is actually submitted.

Drag a panel by its collapse dot (the small circle on the header) to reorder it up/down, or drop it on the other column to swap left/right. Click the collapse dot to fold/unfold as before. To move a panel, hover the drag area for about 1 second until the border turns green, then drag. A quick pass or click will not start a layout move.

The generate bars (width/CFG and seed/run buttons) and the prompt box have no collapse dot. Hover empty space (especially the padding at the top) for about 1 second until the border turns green, then drag. Don't drag sliders, dropdowns, buttons, or prompt text — those keep their normal behavior.

The layout icon next to Save Settings cycles how the layout is stored:

• Default color - Shared UI layout for all settings files
• Green - Independent UI layout for the current settings file only
• Orange - Restore the original default layout. Drag any panel afterwards to save it as an Independent layout (Green)

For ComfyUI, check Comfy-Org@HF GGUF model requires custom node gguf the small-case one.

|---models
|   |---checkpoints
|   |---diffusion_models
|   |---text_encoders
|   |---vae

For Forge Neo, check Download Models@Haoming02 Forge also supports the GGUF model, but the Diffusion models use the same Checkpoint folder. Therefore, manage those models with correct folder yourself.

|---models
|   |---Stable-diffusion
|   |---text_encoder
|   |---VAE

Supports magnification scales from 1.5x to 8x. By adjusting different tile sizes, it can optimize VRAM usage to the maximum, enabling 8GB graphics cards to achieve image upscaling at 1.5x to 3x (or higher) scales, while also unleashing the full potential of high-end graphics cards with 24GB or more VRAM to accomplish 4x to 8x image upscaling.

Requires ComfyUI_Mira 0.5.6.0 or above Requires MiraSubPack latest

And Image Tagger for Mira

| Settings | Drag and Drop | Flux.2 | | --- | --- | --- | | | | |

The generated results will be affected by different models. The SDXL model requires a tagger to provide more precise content descriptions, and using an upscale model is recommended. Models such as Flux2, which allow a reference latent to connected, can skip the upscale model and stretch the original image directly. Then, simply use the Positive Prompt to configure all tiles consistently.

| Upscale Ratio 1024x1360(1536) | Tile Size | VAE 8G | VAE 24G~ | | --- | --- | --- | --- | | x1.5 | 768 | Full | Full | | x2 | 1280 | Full/Tiled | Full | | x3 | 1536 | Tiled | Full | | x4 | 2048 | Tiled | Full | | x6 | 2048/2560 | Tiled | Full | | x8 | 2560/3072 | Tiled | Full/Tiled |

Note: Due to limitations of the SDXL model, exceeding 8x magnification will result in overly fragmented visual slices. For higher magnification requirements, it is recommended to use a dedicated ComfyUI workflow instead. Comfyui Workflow MiraITU

Download models with tags from HF, manually rename them according to the following rules, then copy them into models/tagger folder: - cl_tagger_v2.onnx + cl_tagger_v2_tag_mapping.json - wd-eva02-large-tagger-v3.onnx + wd-eva02-large-tagger-v3_selected_tags.csv - wd-v1-4-convnext-tagger.onnx + wd-v1-4-convnext-tagger_selected_tags.csv

SAA
|---models
|   |---tagger
|       |---cl_tagger_1_02.onnx
|       |---cl_tagger_1_02_tag_mapping.json
|       |---wd-eva02-large-tagger-v3.onnx
|       |---wd-eva02-large-tagger-v3_selected_tags.csv
|       |---wd-vit-large-tagger-v3.onnx
|       |---wd-vit-large-tagger-v3_selected_tags.csv
|       |---camie-tagger-v2.onnx
|       |---camie-tagger-v2-metadata.json
Options:
Model Name  >>>  General Threshold(CL/WD/Camie)  >>>  Character Threshold(CL/WD)  >>> Categories(CL/Camie) or mCut(WD)

The Image Tagger runs on Node.JS with onnxruntime-node. It DOES NOT require any backend support. Idle: prefers DirectML / CUDA for the currently selected tagger only (one GPU-resident model). Falls back to CPU if GPU init fails; promoting a model to GPU drops its CPU copy. Generating: prefers CPU so diffusion keeps VRAM. Closing Image Info, or starting Generate / ControlNet, releases any GPU tagger session into the CPU cache when possible. GPU-resident models unused for ~2 minutes are released (into CPU cache when possible) even without Generate. CPU keeps up to the last 3 used ONNX models; unused CPU entries expire after ~5 minutes.

DirectML is still slower than Python onnxruntime-gpu (CUDA), but steady GPU runs are much closer to usable speed than the old per-request CPU path.

In other words with my i9-9960x with Titan RTX The good news is, you can run Image tagger during generate (uses CPU while generating, then can return to GPU when idle)

| Device | Avg Tagging Time | Model | Platform | Resolution | Recommend Value | | --- | --- | --- | --- | --- | --- | | onnxruntime | 1.053s | cl_tagger_1_02 | Python | 448 | 0.55/0.60 | | onnxruntime-gpu | 0.297s | cl_tagger_1_02 | Python | 448 | 0.55/0.60 | | onnxruntime-node | ~0.12s (steady) | cl_tagger_1_02 | Electron(NodeJS@DirectML) | 448 | 0.55/0.60 | | onnxruntime-node | ~1.0s (steady) | cl_tagger_1_02 | Electron(NodeJS@CPU) | 448 | 0.55/0.60 | | onnxruntime-node | (first load slower) | wd / camie | Electron(NodeJS@DirectML) | 448 / 512 | 0.35/0.85 / 0.50 |

For ComfyUI

ControlNet requires comfyui_controlnet_aux 1.1.5 latest Put your ControlNet models in ComfyUI\\models\\controlnet

IP Adapter requires comfyui-art-venture 1.1.7 latest and ComfyUI_IPAdapter_plus 2.0.0 latest Put your Clip Vision models in ComfyUI\\models\\clip_vision Put your IP Adapter models in ComfyUI\\models\\ipadapter I didn't test too much on IPA, but for SDXL/ilXL/NoobXL recommends CLIP-ViT-bigG-14-laion2B-39B-b160k.safetensors with ipa_styleIpadapterFor_NoobAI-XL_v10.safetensors. You may also need CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors Only the first IP Adapter slot will accept by ComfyUI, others set to On will ignore.

For Forge Neo

Usage

1. Drag and drop (or click Add then Paste) your Image(or openPose image) to SAA/SAAC Image Info, select Pre-processor, Resolution, Post-processor and then click Add ControlNet. After it says Added the previre image will swap to your Pre-processed image, close Image info window, check ControlNet tab for more settings. Hover your mouse over a dropdown/text item to view its feature.
2. In ControlNet tab, you can change Pre-processor by select the a one and click Refresh to generate it and update preview. (Or set Method to On but SAA preview will not update).
3. Because images are too big to save in settings file, so ControlNet settings will not save when SAA close, but also select another settings will not override current ControlNet settings.
4. ControlNet works on normal and regional conditon.
5. If you are new to ControlNet, try Canny or OpenPose first.
6. ComfyUI D-O-E-S N-O-T like submitting the same data, you may receive an Empty response error when submit same data.
7. ComfyUI comfyui-art-venture requires a square image, you may get warnings with NOT square image input.
8. In case of your IPA image is too big, the Resolution selection for IP Adapter will resize your input image to target size, 1024 is enough for most case.
9. The Info button is not working when Pre-Process Model select to IP Adapter

Supports English and Chinese(translate file required) tag search. Special thanks to Kiratian(天痕) for helping to translate the tags into Chinese version.

Artist Search for Anima and others Activate it using the @ symbol. The result will be filtered by groups 1 and 8. To apply the Artist tag in Anima Model, you need to add an @ symbol at the start of the tag. E.g. mira -> @mira

| Mark | ID | Category | Group | | --- | --- | --- | --- | | [G] | 0 | General | Danbooru | | [A] | 1 | Artist | Danbooru | | [©] | 3 | Copyright | Danbooru | | [C] | 4 | Character | Danbooru | | [M] | 5 | Meta | Danbooru | | | 7 | General | E621 | | | 8 | Artist | E621 | | <©> | 10 | Copyright | E621 | | | 11 | Character | E621 | | | 12 | Species | E621 | | | 14 | Meta | E621 | | | 15 | Lore | E621 | | Wildcards | 255 | Wildcards | SAA |