Advanced real-time control for Z-Image models and Qwen3-4B Text Encoders. These nodes perform "live surgery" on your model's weights to boost or lower specific generative phases without any training.
UPDATE 2.2.0
Now compatible with Z-image Base
Hi everyone!
I developed these nodes to slice the monolithic S3-DiT of Z-Image to give you more control over different slices of both the Model and Text Encoder. This transform your Z-Image model into a set of "Sliders," allowing you to tweak composition, lighting, or micro-details in real-time.
You can finally play with the models!
Work in Progress: Please keep in mind that the boundaries for these functional blocks are experimental. Since Z-Image is a single-stream architecture... it's quite chaotic. If you want more freedom and help us all understand what's really inside each slice of Z-image and Qwen, feel free to play with the "LAB" version, where you can control every layer indipendently.
Prompt for the examples:
American comic book art, Western illustration style, highly detailed, dynamic line work. A powerful female elf standing in a dynamic pose, raising one arm high towards the sky. Her hand is encased in a bulky, complex mechanical gauntlet that crackles with arcs of blue electricity. She wears a sharp, fitted black uniform adorned with a thick, luxurious fur collar. Her vibrant green hair is styled with a heavy side-sweep that completely covers one eye. Small, delicate antlers frame the sides of her head. The perspective is a dramatic low angle looking up, emphasizing her dominance and the height of her reach. The background places her in the center of a cobblestone medieval square, but the low angle highlights a brilliant, clear summer sky dotted with fluffy white clouds behind her. The lighting is bright and natural sunlight, accented by the intense glow of the electrical sparks. The image conveys a strong sense of movement, with wind blowing through her hair and the surging energy of the gauntlet.
The 30 layers of the model are divided into 6 Positional Blocks (5 layers per block), plus controls for the auxiliary components that wrap around the main transformer.
A. The Main Stack (Positional Blocks)
You can push or pull the weights of these blocks relative to the base_strength.
- Block 1 (Start 00-04): Initial composition.
- Block 2 (Early 05-09): Early feature definition.
- Block 3 (Mid 10-14): Mid-level structure.
- Block 4 (Core 15-19): Core processing.
- Block 5 (Late 20-24): High-level details.
- Block 6 (End 25-29): Final output resolution.
B. Global Components Instead of affecting layers, you can scale specific internal mechanisms across the entire model:
- Global Attention: Scales the Attention mechanisms. This impacts how the model relates different parts of the image to each other.
- Global MLP: Scales the Feed-Forward networks. This impacts the processing power of the individual nodes.
C. Auxiliary Models These controls target weights that exist outside the main 30-layer stack:
- Embedders Strength: Scales the internal embedding layers. This controls how strong the initial signal injection is before it enters the main layers.
- Refiners Strength: Scales the dedicated refinement weights, often responsible for noise handling or context refining.
- Unsafe Tune Normalization: By default, this is OFF (Locked). Normalization layers stabilize the neural network; scaling them (changing their math) often leads to artifacts or "fried" images. Enable this only if you want to deliberately break the model's stability.
The Qwen3-4B text encoder has 36 layers. The v2.0 Tuner organizes these into 6 Semantic Zones (6 layers per zone), moving from raw embedding to abstract reasoning.
- Zone 1 (Layers 00-05): Embedding & Tokenization
- Zone 2 (Layers 06-11): Low-Level Syntax
- Zone 3 (Layers 12-17): High-Level Syntax
- Zone 4 (Layers 18-23): Semantics
- Zone 5 (Layers 24-29): Contextual Logic
- Zone 6 (Layers 30-35): Abstract Reasoning
Adjusting these allows you to influence how strictly the model adheres to grammar versus how freely it interprets the "vibe" of your prompt.
But wait, there's more:
For those curious FREAKS that want to look under the hood and customize their model with complete control:
The LAB versions of these nodes discard the groupings entirely. They expose:
- Layers 00-29 individually for the Z-Image Model.
- Layers 00-35 individually for the Qwen Text Encoder.
This is intended for research and fine-tuning. If you want to discover exactly which layer controls "lighting" or "texture" for a specific seed, this is where you experiment.
Once you have tuned your model or text encoder state in memory, you can use the Z-Image Saver or Qwen Saver nodes to serialize your changes into a standalone .safetensors file, ready to be loaded in other workflows.
Required for Z-Image Tuning
This custom loader is critical for the Z-Image Tuner. Standard ComfyUI loaders cache models in your RAM to speed up generation. However, because the Z-Image Tuner modifies weights "in-place" (live editing the model in memory), using the standard loader causes your adjustments to accumulate. If you run the tuner twice, you might accidentally apply your changes on top of the previous changes, "frying" the model.
The Arthemy Tuner Loader forces a "clean refresh" from the disk every time you run the workflow. It ensures that every tuning session starts from the pristine, original model file, preventing "dirty cache" corruption.
For the Qwen Text Encoder, you do not need the custom Loader.
- Navigate to your
ComfyUI/custom_nodes/folder. - Run:
git clone https://github.com/aledelpho/ComfyUI-Arthemy-ZImage-Tuner.git - Restart ComfyUI.
