<h1 align="center"> ComfyUI-Flow-Assistor <br> _{^{<i>Essential utility nodes for ComfyUI</i>}} </h1> <p align="center"> <img width="100%" alt="ComfyUI-Flow-Assistor" src="https://github.com/user-attachments/assets/c9306ce8-96b0-4535-91f7-144f2ac27840" /> </p> <br>

Get Started

Install

1. Install ComfyUI.

2. Clone this repository into ComfyUI/custom_nodes:

   cd ComfyUI/custom_nodes
   git clone https://github.com/Merserk/ComfyUI-Flow-Assistor.git
   

3. Restart ComfyUI.

Requirements: Python 3.10+ and a current ComfyUI build with the V3 node API (comfy_api.latest). Caption Creator additionally requires native CLIPType.KREA2 and Qwen3-VL ConvRot support. Version 2.4.1 is V3-only.

All 26 nodes are organized under:

flow-assistor/
├── diagnostics/
├── flow/
├── image/
│   └── caption/
├── loaders/
├── sampling/
├── text/
└── utils/

<br>

🛠️ The Nodes

1. 📝 Prompt Queue

Output prompt lines in sequence.

Paste a multi-line prompt list and receive the next usable line on every queue. Supports empty, repeat_last, and loop, plus reset, whitespace stripping, and empty-line skipping.

---

2. 📂 Prompt Queue (From Folder)

Read prompt files one by one.

Loads supported files from a selected folder in deterministic filename order and outputs both the file content and filename. Supports extension filters, reset, loop, empty, and hold-last behavior.

---

3. 🎥 Camera Angle Control

Build consistent camera-direction prompts.

Create natural, technical, or keyword-style descriptions from rotation, vertical position, subject distance, and focal length controls.

---

4. 🎨 CLIP Text Encode (Prompt Enrichment)

CLIP encoding with 20 built-in style presets.

Adds a selected preset—such as Cinematic, Anime, Cyberpunk, Line Art, Dark Fantasy, or White Background—to your prompt before creating conditioning.

---

5. 📺 Show Text

Display incoming text directly on the graph.

Useful for checking prompt queues, generated camera descriptions, filenames, and other string outputs without opening the console.

---

6. ✍️ Caption Creator

Generate detailed captions from images with a local Qwen3-VL model.

Accepts one image or an image batch and returns one precise caption per image, separated by newlines. Choose the int8 or int4 ConvRot model and set an approximate caption length from 1–200 words; words = 0 requests an unrestricted detailed caption. The selected word count is a target, not a cutoff: generation ends naturally after a complete sentence.

Missing models can be downloaded automatically and are stored in:

ComfyUI/models/text_encoders/flow-assistor/

Disable auto_download to manage the model files manually. Caption Creator checks for ComfyUI's native CLIPType.KREA2 support and reports a clear upgrade error when the installed text-encoder loader is too old for the selected format. It performs one generation pass per image, keeps Qwen3-VL thinking disabled, and returns the model decoder's caption directly without custom cleanup, repetition trimming, punctuation repair, or word-count truncation. A fixed 512-token emergency ceiling remains independent of the requested word count. Oversized caption inputs are reduced to a maximum 784-pixel edge for faster vision processing. On an accelerator-enabled ComfyUI installation, the node asks ComfyUI's model manager to keep the text encoder fully resident on the configured GPU when VRAM permits, then falls back to managed GPU offloading if a full load is not possible. Device residency, token count, timing, and emergency-ceiling status are written to the console. The exact output also appears in the node's scrollable preview. Version 2.4.1 uses Qwen3-VL Instruct-style sampling and an explicit caption-list UI payload so a single caption is never split into numbered characters.

---

7. 📐 Resolution Selector (Groups)

Choose standard dimensions by megapixel group.

Includes common aspect ratios from 0.25MP through 4MP, returns width, height, and an empty latent, and supports batch sizes up to 64.

---

8. 🖼️ Image Resolution Tools

Inspect or fit image and latent dimensions.

• Image Resolution Fit — Resizes an image toward a selected megapixel target while preserving aspect ratio.
• Image Resolution Extractor — Outputs width, height, matching latent, and the unchanged image.
• Image Latent Resolution Extractor — Reads the effective pixel resolution from latent samples.

---

9. 🖱️ Visual Marquee (Interactive)

Pause, select, and process one image region.

Opens an interactive crop selector in the browser and returns the cropped image, mask, and shared TILE_DATA metadata. Crops can remain at original size or be resized to a chosen maximum resolution.

---

10. 🧩 Tile Tools

Support crop-and-merge workflows.

• Tile Manager (Crop) — Preserves the current image, mask, and TILE_DATA workflow contract.
• Tile Compositor (Merge) — Places a processed tile back into the original image, automatically resizing it when needed and optionally feathering the edges.

---

11. ☁️ LoRA Online

Load a LoRA directly from a URL.

Accepts direct file links or Civitai model URLs, downloads asynchronously, applies the LoRA to the model, and can either keep the file or delete it after loading.

---

12. 💎 Detail Enhancer

Adjust sampling to emphasize fine detail.

• Detail Enhancer (Ultimate) — Wraps a sampler.
• Detail Enhancer (Sigmas) — Modifies a sigma schedule directly.

Both provide enhancement strength, texture boost, and start/end percentage controls.

---

13. 🔀 Any Passthrough

Universal rerouting helpers.

• Any Passthrough (6 → 1) — Returns the first connected non-null input.
• Any Passthrough (1 → 6) — Duplicates one input to six matching outputs.

---

14. 🎮 Flow Control (Sidecar Bypass)

Control connected nodes without placing the controller in the data path.

Use the frontend switches to change connected nodes between normal execution and bypass mode.

---

15. ⏱️ Add Delay

Wait before passing data onward.

Adds an asynchronous delay while preserving the connected data type. Useful for timing, interactive workflows, and debugging.

---

16. 🧹 VRAM/RAM Cleaner

Run memory cleanup during a workflow.

Passes the connected object through unchanged and provides three modes: clean the current object, unload other models, or unload everything.

---

17. ✖️ Multiplication (Dual & Latent)

Scale dimensions and optional latent data.

Multiplies two integer inputs by a shared factor and resizes an optional latent with the same multiplier.

---

18. 🐞 Debug Data (Any Input)

Inspect almost any connected value.

Outputs readable information such as Python type, tensor shape, dtype, device, image resolution, batch size, and latent pixel resolution.

---

19. 🔬 Precision Detectors

Inspect active model precision at runtime.

• Detect Model Precision
• Detect CLIP Precision
• Detect VAE Precision

These nodes report compute dtype, stored weight dtype, mixed precision, and supported quantized layouts when detectable.

<br>

🚀 Workflow Examples

The Storyteller

Connect Prompt Queue to your text-conditioning workflow, set ComfyUI's batch count to the number of prompt lines, and generate each scene in order.

Interactive Detail Refinement

Send an image through Visual Marquee, process the selected crop, then reconnect it to Tile Compositor with the original image and tile_data.

Dynamic Upscaling

Extract width and height, pass them through Multiplication, and use the results for a proportional latent or image upscale.

<br>

🤝 Contributing

Ideas, bug reports, and pull requests are welcome. Please keep node IDs and workflow-facing input/output contracts stable whenever possible.

📄 License

MIT License. Precision Detector attribution is included in NOTICE.