ComfyUI image utilities

Image manipulation that doesn't go through a diffusion model. Plus image-to-text inference (Florence-2, WD14, BLIP, DeepDanbooru), which is included here because the input is an image and the node-level setup (model downloads, ONNX dependencies, HF cache paths) is the bulk of the work.

The split:

When to Use This Skill

Sources of truth

• custom_nodes/comfyui-kjnodes/nodes/image_nodes.py — batch / resize / channel / size+count
• custom_nodes/comfyui_essentials/image.py and mask.py — Image*/Mask* family
• custom_nodes/comfyui-tooling-nodes/ — base64, cache, websocket, tiling
• custom_nodes/comfyui-florence2/ — Florence-2 loaders + caption nodes
• custom_nodes/comfyui-wd14-tagger/ — WD14 ONNX wrapper
• custom_nodes/comfyui-art-venture/modules/interrogate/ — BLIP, DeepDanbooru

Resize decision

Latent-friendly sizing

Many models require dimensions divisible by 8 (or 16 for some WanVideo configs). ImageResizeKJv2 has a divisible_by widget; essentials ImageResize does not. For latent-aligned resize, prefer kjnodes' v2; otherwise pre-compute the target size via SimpleMath (see comfy-math-strings).

Batch operations

The batch ↔ list distinction matters: LIST is a Python list of single-image tensors processed by INPUT_IS_LIST=True nodes; BATCH is a single 4D tensor (B, H, W, C). Many downstream nodes (samplers, VAE) expect BATCH. Convert with ImageListToBatch before the sampler.

Mask utilities

For mask shape detection (face mask covers a region; is it empty?), the easy isMaskEmpty probe lives in comfy-conditionals.

Image I/O

For batch / cross-output-directory metadata analysis (figure out which model produced a directory of outputs), use the comfy-metadata skill — it covers PNG tEXt / iTXt, WebP EXIF, MP4 container metadata (kijai WanVideoWrapper's comment blob), and .latent safetensors.

Tiling

For tiled diffusion (split a large image into tiles, run each through a sampler, stitch), that's a model-level pattern — see comfyui-tiled-diffusion (separate pack) or the comfyui-inpaint-cropandstitch skill referenced in portrait-outpaint.

Channel ops

The kjnodes split/merge pair is the workhorse for any per-channel image processing.

Image → prompt (image-to-text inference)

These nodes are inference nodes in the sense that they run a neural network, but they're not diffusion — they produce text from an image. They live in this skill because the input is an image and because users coming from comfy-prompting need the node-level setup details.

Decision: which captioner?

Florence-2 setup

Nodes:

• DownloadAndLoadFlorence2Model — auto-fetch from HF on first run
• Florence2ModelLoader — load from a local path (after the first run, point at the cached path)
• DownloadAndLoadFlorence2Lora — apply a PEFT LoRA adapter (PromptGen variants are LoRAs on top of base Florence-2)

The first run downloads to the HF cache. On a small-root-disk install, set HF_HOME to a larger data disk to keep models off the small root — see ~/.claude/rules/huggingface-downloads.md. Florence-2 weights for the base model are ~1.5 GB; PromptGen LoRA adapters are ~50 MB.

Florence-2 variants and their use:

WD14 setup

Single node: WD14Tagger. ONNX-based, requires onnxruntime (CPU) or onnxruntime-gpu for CUDA acceleration. The ONNX model auto-downloads to the WD14 pack's model directory on first run.

Configurables:

• model: WD14 model variant (Convnext, ViT, SwinV2, MoAT) — newer variants are slightly more accurate
• threshold (general tags): 0.35 default; raise to filter weak tags
• threshold_character: tags for specific known characters
• exclude_tags: comma-separated tags to skip ("1girl, solo")
• replace_underscore: convert booru underscores to spaces
• trailing_comma: append , after the output

BLIP and DeepDanbooru (art-venture)

BlipLoader (or DownloadAndLoadBlip for one-step) + BlipCaption. First-run download ~500 MB. Output is a short caption.

DeepDanbooruCaption for anime — single node, auto-downloads model weights.

Caption chaining

For best-quality prompts, chain:

LoadImage ──► Florence-2 PromptGen ──► STRING (caption)
                                          │
                                          ▼ (optional)
                                  Searge_LLM_Node
                                  (add cinematic detail, camera direction)
                                          │
                                          ▼
                                  CLIPTextEncode (or model-specific encoder)

The Florence-2 → LLM chain produces more nuanced prompts than either alone. See comfy-prompting for the LLM-side details.

Recipes

Sortable per-day output filename with source basename

User wants outputs at output/2026-05-13/143055_michael.png where michael is stripped from the source filename michael.jpg. Project CLAUDE.md covers the native %date:%/%NodeName.widget% syntax; when the native approach fails (e.g. extension stripping), use:

LoadAndResizeImage (kjnodes) ─► image_path (STRING, full path)
                                       │
                                       ▼
                          StringFunction (pysssss, regex)
                          find:    "^.*/|\.(png|jpg|jpeg|webp)$"
                          replace: ""
                                       │
                                       ▼ (bare basename, no path, no ext)
                          JoinStringMulti (kjnodes)
                          in_1: "<bucket>/%date:yyyy-MM-dd%/%date:hhmmss%_%ksampler.sampler_name%_%ksampler.scheduler%_s%ksampler.seed%_"
                          in_2: <bare basename>   # the <descriptor> segment
                                       │
                                       ▼
                          easy imageSave (filename_prefix STRING input)

%date:% tokens pass through verbatim into the SaveImage prefix substitution (resolved at save time). String manipulation chain lives in comfy-math-strings; the save node lives in easy-use. The prefix convention itself is per-install — see "Naming conventions are per-install" in comfy-workflow-json.

Caption a folder of images for batch dataset prep

You have 100 photos in input/dataset/ and want a .txt next to each with a Florence-2 caption:

easy imagesCountInDirectory ──► count
                                  │
                                  ▼ (drives forLoopStart iteration count)
                          easy forLoopStart (total = count)
                                  │
                                  ▼ (per-iteration)
                  LoadImage (path = `input/dataset/{index}.jpg`)
                                  │
                                  ▼
                          Florence-2 PromptGen ──► caption STRING
                                                       │
                                                       ▼
                                                  SaveText (bjornulf)
                                                  path: `input/dataset/{index}.txt`
                                  │
                                  ▼
                          easy forLoopEnd

Pattern hits two skills: this one for Florence-2; comfy-flow-control for the for-loop primitives.

Mask-driven crop + paste workflow

You have a portrait, a face mask, want to upscale only the face:

LoadImage ──► IMAGE
   │
   ▼
GenerateFaceMask (whatever)
   │
   ▼ MASK
   │
   ▼
ImageCropByMask (kjnodes) ──► cropped IMAGE (face region only)
   │                       └─► crop coords (for paste-back)
   ▼
(upscale chain: sampler, etc.)
   │
   ▼ upscaled face
   │
   ▼
ImageComposite (essentials, alpha-paste using mask) ──► final image

ImageCropByMask returns both the cropped image and the bbox; the bbox is needed to paste the result back to the right location.

Tile-based large-image processing

LoadImage (4096×4096) ──► IMAGE
                            │
                            ▼
                  ExtractImageTile (tooling-nodes, 2×2 grid)
                            │
                            ▼ 4 IMAGE tiles
                            │
                            ▼ (process each through a sampler)
                            │
                            ▼ 4 processed tiles
                            │
                            ▼
                  MergeImageTile (tooling-nodes) ──► final 4096×4096

For sampler-aware tiled diffusion (overlap, blending across tile seams), reach for comfyui-tiled-diffusion instead — these tooling-node tile ops are pure image-level.

Gotchas

• LoadAndResizeImage.image_path is a full path, not just a basename. Strip the dir component via StringFunction regex (see Recipe 1) or use the native %LoadImage.image% substitution if not using LoadAndResizeImage.
• ImageResize defaults are not latent-aligned. The result may be 519×731, which crashes the VAE on some models. Use ImageResizeKJv2 with divisible_by=8 or pre-compute via SimpleMath.
• ImageBatchMulti slots are dynamic. Adding/removing inputs rewrites the slot count. After heavy editing, re-add the node to compact unused slots.
• MaskFromColor is RGB-only. RGBA images need RemoveTransparency first (or split channels and feed RGB).
• WD14Tagger needs onnxruntime, not torch. The pack ships its own ONNX session; CUDA acceleration requires onnxruntime-gpu (which on this install would need: .venv/bin/python -m pip install onnxruntime-gpu).
• Florence-2 first run downloads to ~/.cache/huggingface by default. On a small-root-disk install, set HF_HOME to a larger data disk (a systemd unit's Environment= block, or your shell profile). Otherwise the ~1.5 GB weights fill the small root partition. See ~/.claude/rules/huggingface-downloads.md.
• BLIP captions are short (~10-15 words). For longer prompts, prefer Florence-2 PromptGen or chain BLIP output through a local LLM.
• DeepDanbooru and WD14Tagger produce overlapping but non-identical tag vocabularies. Stick with one for a given workflow; mixing produces redundant 1girl, 1_girl, solo, person pile-ups.
• Tooling-nodes' Load Image Cache is in-memory only. The cache doesn't survive a service restart. For persistent caching, save to disk via SaveImage with a known filename and reload.
• ImageComposite (essentials) requires both images to be the same size. Resize one or use ImagePadKJ to match dimensions first.
• SplitImageChannels always emits 4 MASKs (R, G, B, A). On an RGB input, the alpha channel comes back as a fully-white mask (not None) — wire only the channels you need.
• RepeatImageToCount doesn't broadcast smaller dimensions. If the input batch has 1 image and target count is 5, you get 5 copies of that image — useful when matching a batch dimension to drive a per-frame conditioning input.

Cross-refs

• comfy-prompting — when to reach for image-to-prompt nodes; wildcard / LLM combination with caption output.
• comfy-conditionals — easy isMaskEmpty to gate mask-driven branches; MaskBoundingBox to compute "is the region large enough?" predicates.
• comfy-flow-control — index switches over batches; for-loop iteration over image directories.
• comfy-math-strings — string manipulation for filename templating (the recipe above); resolution math.
• comfy-debug-preview — Get*SizeAndCount, CImageGetResolution, ImageDetails for inspecting batches and source images.
• comfy-metadata — offline / cross-file metadata extraction across output directories; the Crystools nodes here are for inline reads, comfy-metadata is for retrospective analysis.
• photo-restore, portrait-outpaint, video-extend — task-level skills that combine these image utilities with model inference.

Things this skill does NOT cover

• Model inference on images — KSampler, VAEEncode/Decode, ControlNet apply, IPAdapter. Those are model-family work; see wan / z-image / hidream-o1 / task skills.
• Tiled diffusion at the sampler level — see comfyui-tiled-diffusion and the inpaint-cropandstitch workflow referenced in portrait-outpaint.
• Image-format conversion at the file level (JPEG quality, WebP encoding parameters) — that's tools-plugin:imagemagick-conversion territory for external CLI work.
• 3D / depth / mesh manipulation — depthanythingv2, depthflow-nodes (broken on this install per project CLAUDE.md).