olivv-cs-ComfyUI-FunPack/docs/how-to-start.md

# How to start generating with FunUI

This is the short path from "installed" to "first clip". It does not explain every node or loader: the app does that where you use them (hover a control, read its hint).

## 1. Install

With git, inside `ComfyUI/custom_nodes`:

`git clone https://github.com/digital-garbage/FunUI`

For the classic 4.x add `-b v4-classic`. `comfy node install FunUI` and ComfyUI-Manager install what is published on Comfy Registry, which is the current version (5.2) only once it is published there; check Settings ▸ About FunUI after installing.

Then `pip install -r requirements.txt` in the same Python environment ComfyUI uses, and restart ComfyUI. FunUI uses your existing ComfyUI and PyTorch. If pip wants to upgrade a package ComfyUI compiles against (torch, numpy, transformers, comfy-kitchen), read what it will change first: FunUI's own updater only installs what is missing and reports the rest.

## 2. Open FunUI

Start ComfyUI as usual and open `/fun/` on the same address, for example `http://127.0.0.1:8188/fun/` (the number after the colon is whatever port your ComfyUI uses). The chip at the top right says whether ComfyUI is live.

## 3. Make a project

**File ▸ Project Setup Wizard** has three steps (Back and Continue sit in its footer): choose your look, name the project and pick **Start from** (the pipeline of the model you will use), then say yes to the tour if you want a walk through the screen (Help ▸ Start tour runs it again).

| Start from | State |
|---|---|
| MiniMax H3 · Reference to Video, MiniMax H3 · Fast & Simple R2V | The model FunUI is built and tested around. |
| Stable Diffusion 1.5 · Text to Image (held as a clip) | Experimental. |
| MiniMax H3-Image · Text to Image and Edit | Experimental. Add pictures as references: the first is the picture to edit (`<Picture 1>` in the prompt), the others are references (`<Picture 2>`, ...), up to 16. The result is a picture file (PNG), and Length and FPS are not offered (both are set to 1). Uses 6 steps at CFG 1; Width and Height set the size (about 4 megapixels is a good start). |
| LTX 2.3 · Image to Video, from References, LTX 2.3 · Video to Video, direct | Tested on a real model, with LoRAs, Step merging and Attention explorer. Other enhancements are not tested on LTX yet; please report problems in [Issues](https://github.com/digital-garbage/FunUI/issues). |

## 4. Pick your files

The model's pipeline starts empty: it needs your files.

1. Open **Settings** (in the menu bar, next to Help) and then **Ready to generate?**. It checks the machine: ffmpeg, GPU, model folders, node packs.
2. Open **Settings ▸ Models & Pipeline** and choose a file in each card marked with an orange dot (diffusion model, text encoder, VAE, and so on). The page says how many cards still need something and says "Every slot is filled" when you are ready.

Put model files where ComfyUI expects them (`models/diffusion_models`, `models/text_encoders`, `models/vae`, `models/loras`). 

`.gguf` text encoders need the ComfyUI-GGUF node pack (FunUI does not install it) and work only for architectures it understands; otherwise they are refused with a message (Gemma 4 `.gguf` is one such case). Use the `.safetensors` file of the same model. A diffusion-model `.gguf` also loads without the pack, but is expanded to full size in memory.

## 5. Write the story

Click **◆ Composer** (right end of the timeline header) and write all scenes in one box, separated by the cut word (see the Cuts tab). Each piece becomes one clip on the timeline. Composer also keeps `$variables`, templates and shortcuts, and has an optional **Enhance** tab: a language model rewrites each finished prompt. On H3, pick a chat encoder (for example Gemma 4) under Enhance ▸ Encoder, because H3's own encoder cannot write prose. Leave Encoder family at its default unless your encoder belongs to another family (Gemma 4 worked with the default).

## 6. Generate

- **▶ Generate** (timeline header) makes every scene on the timeline (not left-out scenes or video clips). An earlier render of a scene stays as a take.
- **Selected** makes only the picked scene. **Takes ×N…** makes the same scene several times, differing only by seed.
- **⏱ Sampler** (right end of the same header) holds steps, sampler, scheduler, seed and the common switches. While a run goes, the step count shows to its left.
- No picture needed: on H3, a scene with no source image is generated from text alone.

Watch the Log window if nothing happens: any module that fails says why there, and the rest of the run goes on without it.

## 7. Cut, render, export

Trim and split clips on the timeline. The **＋ Add** menu in the timeline toolbar adds a scene, a video clip or audio track from the bin, an effect or transition, or a text or picture overlay. **⧉ Render** (timeline header) joins the timeline into one file; **⤓ Export** (toolbar) downloads the picked clip (picture only: a clip whose sound was separated is silent there, Render mixes it).

## 8. Teach it your taste (optional)

Rate a result with **Rate scene...** in the timeline header, next to **Takes ×N...** (liked, disliked, bad image, bad composition). **Upscale** and **Interpolate** appear beside it once you set them to "When I press" in their settings. Set a **Taste key** (the ⚠ chip on the timeline points to it: **Settings ▸ Engine ▸ System**), a name for what the ratings teach, such as a style or a character. Without a key nothing is remembered. **Settings ▸ Refinement & Taste** lists, exports and deletes keys.

**What learns from ratings.** Only features that say so in their hint: for example Taste steering (REINS), Attention explorer, Decisiveness, Taste prompt nudge, Velocity bias, DynaShift and First-step seed search. Each run keeps a compact numeric summary of how the model behaved (no pictures or text). When you rate that run, the summary is filed under your key as liked or disliked. A learning feature then pushes new runs toward the liked summaries and away from the disliked ones. Newer ratings count more than old ones, so your taste can move. Everything stays on your computer.

**How much you need.** It differs per feature and the hint says so. Most need at least 2 liked and 2 disliked runs under the same key. Taste prompt nudge and similar prompt-based ones need 3 liked clips. First-step seed search needs 10 rated clips with both kinds. Until there is enough, the feature does nothing and the Log tells you it needs more ratings. "Learn what I like" in LoRA training works differently: it trains a LoRA on your liked clips.

## 9. Features marked experimental

**Settings ▸ Modules** switches each feature on or off. Several features are experimental: built and run, but not yet proven over many runs. If a module fails while generating, the run carries on without it, the Log says why, and the module stays off until its code changes and ComfyUI restarts, or you turn it back on in Settings ▸ Modules.

## If something goes wrong

- Nothing happens: check the ComfyUI chip and the Log window.
- Wrong look: Settings ▸ Appearance. Lost a layout: View ▸ Reset Layout. Lost a project: File ▸ Open recent.
- A bug or a missing feature: please open an [issue](https://github.com/digital-garbage/FunUI/issues).

## Known limits in 5.2

- LTX is tested with the two presets above (LTX 2.3, and LTX 2.5 by hand); most other enhancements have not been tried on LTX yet and are hidden there until they have.
- In the H3 preset the scene picture sets the video's shape only; it is not a start frame.
- A separate "second pass" H3 preset is not offered, because it was never verified on a GPU. For the same idea (first steps on a smaller picture, the rest at full size) use **Two stages** in Settings > Engine > Sampling. It works on H3 and is marked experimental.