# aestx workflows — drop-in versions

Reworked versions of the original `aestx8-*` / `aestx-final-*` workflows,
updated Jul 2026 to run on current ComfyUI (0.27) + current custom-node packs,
and converted from raw API format to proper UI workflows you can **download and
drag-drop into ComfyUI**. Image inputs are now regular **LoadImage** nodes
(pick/upload a file in the node) instead of `ETN_LoadImageBase64` base64
injection.

## The workflows

| file | what it does |
|---|---|
| `aestx-single-ipadapter.json` | SDXL (Juggernaut v8) txt2img, IPAdapter *style transfer* from a style image + depth ControlNet composition lock from a composition image. Batch of 4. |
| `aestx-prev-single-ipadapter.json` | Same as above plus a PreviewImage node showing the generated depth map. |
| `aestx-style-composition.json` | SDXL Lightning (Juggernaut v9 RDPhoto2 Lightning, 5 steps) with `IPAdapterStyleComposition` — separate style and composition reference images — plus depth ControlNet. Batch of 4. |

`api/` holds the equivalent API-format JSONs (for `/prompt` automation — these
are the exact payloads that were test-executed).

## Changes vs the originals (needed to run on current packs)

- `ETN_LoadImageBase64` → core `LoadImage` (defaults: `aestx-style.png`,
  `aestx-composition.png` — just pick your own images in the two nodes).
- IPAdapter_plus renamed weight type `style transfer (SDXL)` → `style transfer`.
- ComfyUI_essentials `ImageResize+`: `keep_proportion: true` → `method: "keep proportion"`.
- Legacy `Get image size` node → essentials `GetImageSize+` (width output wired
  to the DepthAnything resolution input, as before).

## Dependencies

Custom node packs (auto-installable via `download_missing_models.sh`, which
reads the cnr_id/aux_id metadata embedded in the workflow nodes):

- `cubiq/ComfyUI_IPAdapter_plus` (tested @ `a0f451a`)
- `Fannovel16/comfyui_controlnet_aux` (tested @ `e8b689a`)
- `cubiq/ComfyUI_essentials` (tested @ `9d9f4be`)

Python deps installed for the packs: `opencv-python-headless scikit-image
python-dateutil fvcore yapf omegaconf ftfy addict yacs albumentations
scikit-learn matplotlib numba colour-science` (mediapipe/onnxruntime skipped —
not needed by these workflows).

Models (all mirrored in this repo under `models/`, sources noted):

| path in `ComfyUI/models/` | size | source |
|---|---|---|
| `checkpoints/juggernautXL_v8Rundiffusion.safetensors` | 7.1 GB | `lllyasviel/fav_models` |
| `checkpoints/juggernautXL_v9Rdphoto2Lightning.safetensors` | 7.1 GB | `RunDiffusion/Juggernaut-XL-Lightning` (`Juggernaut_RunDiffusionPhoto2_Lightning_4Steps.safetensors`, renamed to the ckpt name the workflows expect) |
| `controlnet/diffusers_xl_depth_full.safetensors` | 2.5 GB | `lllyasviel/sd_control_collection` |
| `ipadapter/ip-adapter-plus_sdxl_vit-h.safetensors` | 0.85 GB | `h94/IP-Adapter` |
| `clip_vision/CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors` | 2.5 GB | `h94/IP-Adapter` image encoder (renamed to the exact filename `IPAdapterUnifiedLoader` searches for) |

`depth_anything_vitl14.pth` (1.3 GB, mirrored at `models/depthanything/`) is
auto-downloaded by controlnet_aux on first run to
`custom_nodes/comfyui_controlnet_aux/ckpts/LiheYoung/Depth-Anything/checkpoints/`
— copy it there manually for offline pods.

All three workflows were test-executed end-to-end on the H100 pod
(ComfyUI 0.27.0, torch 2.8.0+cu128) and outputs visually verified:
IPAdapter style transfer, depth-map conditioning, and batch generation all
confirmed working.