A text-to-image workflow for Illustrious XL in ComfyUI. One base pass, a hires fix, a face fix, and two PNGs per run: the raw base picture and the finished one. The prompt is built from ten small green blocks (Quality, Subject, Character, Style, Appearance, Outfit / Wear, Pose & Action, Camera & Framing, Scene / Background, Lighting & Effects), or from one full prompt if you prefer that.

It is my old ILL workflow with the face fix block, which I deleted once, rebuilt with the green prompt blocks. Every group is explained, one block at a time, in the guide article.

TAG DICTIONARY

The Illustrious Tag Dictionary has about 1,200 Danbooru tags sorted into the same ten blocks, with the same names as the green nodes, plus 61 negative tags, 73 sub-lists in all. It also has six starter recipes, a Tips tab, and a Research tab with my notes and nine sources. Click tags, copy each block, paste it into the node with the same name. Or use Copy full prompt and paste it into the Full Prompt node. Add group takes a whole sub-list and skips tags you already have. Brackets are escaped for you, because ComfyUI reads ( ) as weight. Want expressions in their own block? Tips shows the three steps.

QUICK START

1. Install the six node packs below, then update ComfyUI. The file was saved on ComfyUI 0.44.0.

2. Put the model files in their folders. My files sit in subfolders, so open Checkpoint, Upscale Model, Face Detector, SAM Segmenter and LoRAs: Detail once and pick your own file. Set any LoRA you do not have to None.

3. Embeddings: download the three, or delete their lines in Positive + Embeddings and Negative + Embeddings.

4. Queue once. The Knight in the ruins recipe is already in the green blocks. Two PNGs land in your output folder, the base one and the _final one.

5. Open the tag dictionary, pick tags or a recipe, and paste each block into its green node with Use Template Blocks on true. For one full prompt, paste it into Full Prompt and set the switch to false.

REQUIRED CUSTOM NODES

• Comfyroll Studio: text replace for the template, both LoRA stacks, aspect ratio.

• ComfyUI Image Saver: checkpoint loader, seed, steps and CFG, sampler, scheduler, both Save nodes. Version 1.23.0 or newer.

• Efficiency Nodes: one node, LoRA names for the metadata.

• SDXL Prompt Styler by twri: the Prompt Styler.

• Impact Pack: Face Detailer and SAM loader.

• Impact Subpack: the face detector.

MODEL FILES

• Required: an Illustrious checkpoint. The file is set to WAI-Illustrious-SDXL v16, wai-illustrious-v16.safetensors, in models/checkpoints.

• For the hires fix: 4x-AnimeSharp.pth from OpenModelDB, in models/upscale_models.

• For the face fix: face_yolov8m.pt from Bingsu/adetailer, in models/ultralytics/bbox, and sam_vit_b_01ec64.pth (Segment Anything ViT-B), in models/sams.

• Embeddings: lazypos, lazyneg and lazyhand, in models/embeddings/Illustr/. Without them, delete their lines in the two Embeddings nodes.

• Optional: three detail LoRAs at 0.4, 8.BSS_Detail_Enhancer_v3Best, Consistency_Corrector_iLL_v1_Stable and add-detail-xl. Without them, set each name to None. A missing file stops the queue even with its switch Off.

Only want the base picture? Mute Save: Final (Ctrl+M). Then the upscale and face fix groups do not run, and you do not need the upscaler, the face detector or SAM.

DEFAULTS

• 896x1152 portrait (CR Aspect Ratio preset, Swap Dimensions On for 1152x896). 28 steps, CFG 6, euler_ancestral, simple.

• Hires fix: 4x-AnimeSharp then 0.38, so 1.52x, final 1360x1744. 20 steps, denoise 0.35.

• Face fix: 20 steps, CFG 6, denoise 0.4, its own sampler settings. Only the seed is shared with the base pass.

• The Seed node adds its increment (9982 in the file) to the seed. The seed in the file name is the sum. To repeat an old image, paste its seed and set increment to 0.

• Clip skip: nothing to set. ComfyUI's SDXL encoders already use the second-to-last layer, which is A1111's clip skip 2. The savers record Clip skip: 2 as metadata, for people who repeat a picture in A1111 or Forge.

KNOWN LIMITS

• The dictionary tags are only tested in the six starter recipes. The rest are standard Danbooru tags, sorted, not rendered one by one.

• The _final PNG's metadata describes the base pass (896x1152, 28 steps, denoise 1). The workflow embedded in the PNG has the hires and face values.

• Rendered on WAI-Illustrious v16 only. Other checkpoints load, I did not test them with this file.

Rendered on an AMD card under ROCm. Nothing in the graph is AMD-specific.

If something breaks, tell me in the comments with the error text. If it works, a thumbs-up here helps the next person find it.

Free to use. Credit is appreciated.