iGEN One — RB9.1

The key design principle is modularity : every feature lives in its own group that can be independently enabled or disabled by simply muting or bypassing it. You never need to reconnect anything — the pipeline automatically adapts to whatever groups are active.

The RB8 version is a large-scale, full-featured release containing the complete suite of image loaders, ControlNet processors, style transfer networks, detailers, upscalers, and advanced filter adjustments. It also flattens 14 component subgraphs directly onto the main canvas, exposing their internal controls and intermediate preview/stop nodes so you can inspect and verify each stage without digging into subgraph interiors.

How It Works — The Basics

4-Quadrant Layout

To make this massive workflow easy to navigate, the workspace is arranged as a 4-quadrant grid comprising three horizontal rows split into Left and Right halves:

Top-Left Quadrant (Row 1 & 2, Left): Configuration & Model Loading (Groups 0–9) and Image Inputs / Pre-processors (Groups 10–17).

Bottom-Left Quadrant (Row 3, Left): Prompting & Core Rendering (Groups 18–26). Includes prompt raffle, wildcard processors, initial rendering, and 2nd pass latent upscale.

Bottom-Right Quadrant (Row 3, Right): Post-Processing Stage 1 (Groups 27–32). Contains Flux2 Refine (3rd pass), BFS Face Swap, Flux2 Edit, Tile Upscale (4th pass), Upscale / Sharpen, and Detailer 1 (Face).

Top-Right Quadrant (Row 1 & 2, Right): Detailers & Finishing (Groups 33–44). Contains Detailers 2–6, SeedVR2 Upscale, Final Crop, Rescale Image, Image Adjustments, Watermarks, and Save Image.

S-Curve (Snake) Post-Processing Flow

The post-processing pipeline on the right half flows in an S-curve (snake) pattern moving upwards — so when you read across the bottom-right then back across the middle-right then back across the top-right, you are following the order images travel through the pipeline:

Row 3 (Bottom, Left→Right) : Flux2 Refine → Flux2: Face Swap → Flux2 Edit → Tile Upscale (4th Pass) → Upscale / Sharpen → Detailer: 1 (Face)

Row 2 (Middle, Right→Left) : Detailer: 2 (Eye) → Detailer: 3 (Mouth) → Detailer: 4 (Hand) → Detailer: 5 (Placeholder) → Detailer: 6 (Placeholder)

Row 1 (Top, Left→Right) : SeedVR2 (Upscale) → Image Final Crop → Rescale Image → Image Adjustments → Create Watermark (Text) → Create Watermark (Logo) → Save Image

Toggling Features & Flattened Subgraphs

Each group has a Fast Mode Toggle panel — a small control panel that lets you mute or bypass individual sub-features within the group. To disable an entire group, you mute/bypass the group itself in the ComfyUI canvas.

To mute/activate an entire group: right-click the group header → "Set Group Nodes to Never" (mute all) or "Set Group Nodes to Always" (activate all).

In RB8, 14 subgraphs from previous versions have been flattened — their internal nodes are placed directly on the main canvas instead of being hidden inside a component box. This means you can see intermediate preview images, interact with Stop nodes, and adjust settings without ever opening a subgraph. Despite this expansion, the layout stays compact: all groups are standardized to a vertical height of 808px , keeping everything neatly aligned.

When a group is muted or bypassed, downstream groups automatically skip it and pick up from the last active group. This works because of a priority-based fallback system: each group tries a list of possible input sources in order and uses the first one that's actually active.

You can enable any combination of groups and the pipeline will always find the right data path. There is no need to manually reconnect anything.

Data Routing

Instead of visible noodle connections between groups, iGEN ONE uses Set/Get nodes (Eclipse Version) — named value channels that work like wireless connections. A SetNode in one group publishes a value (like ref_image or model_init ), and a GetNode in another group retrieves it by name. This keeps the visual layout clean and makes it easy to rearrange groups. Two additional variants handle the priority fallback logic: GetAllActive collects all currently active values published under a name (used to chain model sources), and GetFirst picks the first non-empty value from a list — this is what powers the automatic "last active group wins" behaviour throughout the pipeline.

Compatibility with KJNodes Set/Get nodes is one-way : Eclipse Get nodes can read values published by KJNodes Set nodes, but KJNodes Get nodes cannot read values published by Eclipse Set nodes. In practice this means you can use KJNodes Set nodes as a source and pick them up with Eclipse Get nodes, but not the other way around.

Left Half — Generation & Configuration

Model Loading & Configuration (Groups 0–9)

This section at the top-left houses all the settings and model loading pathways. It runs once at the start and feeds everything downstream.

0. Folder / Resolution

The configuration hub for the entire workflow. This is where you set:

Output folder — uses a date-based template ( %Y-%m-%d ) so your outputs automatically organize into day-stamped subfolders.

Image dimensions — default is 832×1248 (2:3 aspect ratio, optimal for Flux2 and Z-Image-Turbo).

Batch size — how many images to generate in a single run.

VRAM purge — whether to aggressively free GPU memory between operations (useful on tighter setups).

1. Model Loader (Main)

Loads your primary generation checkpoint using Eclipse's Smart Model Loader . This is the model that drives the Initial Render and 2nd-pass upscale. The default configuration loads Krea2_MoodyMix_v3 , an external CLIP model (Qwen3vl_4b_fp8_scaled, krea2 type), and an external VAE. All of this is saved in a template so you can swap it in one click.

2. Model Patcher (Main)

A toolkit of optional model-level modifications you can stack on top of your checkpoint without changing the checkpoint itself. Each is independently toggleable, so you can experiment by turning them on one at a time:

Krea2T-Enhancer — A specialized attention patch that improves output quality with Krea2-family models (enabled by default).

ModelSamplingFlux — Adjusts the internal noise schedule for Flux-based models. Enable this if you're using a Flux checkpoint.

DynamicThresholdingFull — Prevents prompt adherence from degrading at high CFG values.

PAG (Perturbed Attention Guidance) — Amplifies fine detail by subtly disrupting self-attention during sampling.

SAG (Self-Attention Guidance) — A complementary technique that enhances feature contrast using attention maps.

DifferentialDiffusion — Allows masked selective denoising, so different areas of the image can be denoised at different strengths.

CFGZeroStar — An alternative CFG method that can improve quality at lower CFG values.

PatchSageAttention — Replaces standard attention with a memory-efficient implementation. Good for reducing VRAM usage.

TorchCompileModel — JIT-compiles the model for faster inference. There's a warmup cost on the first run, but subsequent runs are faster.

TeaCache — Caches token computations across steps to skip redundant work. Default threshold is 0.4.

UNetTemporalAttentionMultiply — Modifies temporal attention layers; primarily for video-aware models.

3. Model Loader (Detailer)

A dedicated checkpoint loader for all the detailer groups (Groups 32–37). By separating detailer models from the main generation model, you can run a lighter, faster model specifically tuned for inpainting and detail enhancement without affecting your primary generation pipeline. Default: darkBeast Blitz6 via the ZIT_DarkBeast_Blitz6 template.

4. Model Patcher (Detailer)

The same set of optional patches as Model Patcher (Main) , but applied to the detailer model instead. All patches are bypassed by default — the detailer model typically doesn't need them, but they're here if you want to experiment.

5. PuLID (Flux)

Identity preservation using PuLID. Load a reference photo of a face and PuLID encodes it using FaceNet/InsightFace, then guides the generation to maintain that person's facial features throughout the diffusion process — without modifying your prompt or checkpoint. This is the standard Flux implementation.

6. PuLID (Flux: Nunchaku)

The same identity preservation as PuLID (Flux) , but optimized for Nunchaku-quantized Flux models. If you're running a quantized model to reduce VRAM usage, use this group instead of PuLID (Flux) .

7. Flux Redux

Style transfer using Flux Redux. Load one or two reference images and the workflow encodes their visual style using CLIP Vision, then blends that style signal into your generation. Unlike img2img, Redux influences the aesthetic without constraining the composition — your prompt still drives the structure.

8. Preprocessor

Prepares a reference image for use with ControlNet or depth-based img2img pathways. It scales the image to a safe resolution (target 1.68M pixels) to avoid out-of-memory errors, then runs the Zoe DepthAnything preprocessor to extract a depth map. You only need this group active when using ControlNet, the i2i (Flux Preproc) toggle in Initial Render, or the i2i (DiffSynth: Qwen Lora) toggle.

9. ControlNet

Structural conditioning — guides where objects and shapes appear in the generated image without needing to describe them in text. Supports:

Standard ControlNet union-promax — General-purpose structural conditioning (strength 0.75).

DiffSynth Qwen/ZIT ControlNet — Alternative ControlNet using the Z-Image-Turbo model (strength 0.65).

Image Sources & Pre-processors (Groups 10–17)

The workflow offers three ways to get a starting image. Only one should be active at a time — the pipeline automatically picks whichever source is enabled. A loaded image can serve two purposes: as a visual reference for img2img generation, or simply as input for the Image to Prompt group to generate a text description.

You can also load an image and skip the Initial Render entirely — disable the Initial Render switch in the Initial Render Group, and the loaded image goes straight to post-processing. This lets you bring in images from anywhere (other workflows, other tools, photographs) and run them through the full detailer/upscale pipeline.

10. Image Load

Loads a single image from disk. This is the simplest option — pick a file and go. It also extracts any embedded generation metadata (model name, prompt, sampler, seed) from the image, and a Show Text [Stop] node labeled Metadata Preview pauses the workflow so you can read it before proceeding. Useful for "remix" workflows where you want to regenerate with the same settings that produced the original.

11. Image Load from Folder

Batch processing mode — loads images one by one from a folder. Like Image Load , it exposes the extracted metadata and a Stop for review.

Set the index to -4 for shuffle mode (random order, no repeats). The optional seed_input slot controls when special modes advance — connect a seed and keep it fixed to freeze the selection while you tweak other settings. Change the seed value to advance to the next image.

12. Image Load from Folder / Video

Native video support and interactive frame selection. It can decode PyAV video frames or load sequential image files from a folder. After loading, the Image Selector node shows a visual grid...

Pack de nœud · Registry