NegPiP Prompt (Multi-Model)

ComfyUI NegPiP Prompt

<p align="center"> <img src="assets/icon.png" alt="NegPiP Prompt icon" width="200"> </p>

A standalone, single-node NegPiP implementation for Z-Image, Z-Image Turbo, Krea 2, MOSS-Image / MOSS Turbo, SD1, SDXL, and Anima. The node patches the connected model, merges separate positive and negative text strings, and returns the conditioning required by a normal ComfyUI sampling workflow.

The package keeps its original Registry/repository name, comfyui-negpip-zimage, but version 2.x is no longer limited to Z-Image.

Installation

Open a terminal in ComfyUI/custom_nodes and run:

git clone https://github.com/cyberdeliaAI/comfyui-negpip-zimage

Restart ComfyUI afterwards. No additional Python packages are required beyond the dependencies included with an up-to-date ComfyUI installation.

Krea 2 support requires a ComfyUI version that includes the native Krea 2 model and krea2 CLIP loader type.

Usage

Place NegPiP Prompt (Multi-Model) directly after the matching model and CLIP loaders. It has four inputs:

• model: the loaded diffusion model;
• clip: the matching text encoder;
• positive: a connected positive STRING;
• negative: a connected negative STRING, optionally containing weights.

Both prompt inputs are sockets. They can be connected to a multiline text node, wildcard node, prompt generator, or any other STRING output.

The node converts the negative string to negative NegPiP weights. For example:

positive: a sharp portrait, detailed eyes
negative: blurry background, (text:1.3)

is compiled to:

a sharp portrait, detailed eyes, (blurry background:-1), (text:-1.3)

The outputs are:

• patched_model: the patched MODEL that must be connected to the sampler or guider;
• positive CONDITIONING containing the compiled prompt;
• empty negative CONDITIONING for the existing guider/sampler workflow;
• compiled_prompt as a STRING, showing exactly what was sent to the patched text encoder.

Connect compiled_prompt to Preview Any to inspect the conversion. The negative conditioning is intentionally empty: NegPiP processes negative concepts inside the compiled positive conditioning.

The node consumes and encodes the connected CLIP internally, so a CLIP output is not required. This also applies to Krea 2: its patched CLIP clone is used immediately for the node's conditioning output.

Prompt strength and length

You can enter either plain negative text or positive magnitude weights:

blurry, background blur, bokeh

(blurry, background blur, bokeh:0.4)

The node changes these to weights of -1 and -0.4, respectively. Do not enter a negative weight in the negative input unless you deliberately want the same absolute strength; the node always makes negative-input weights negative.

Explicit weights use parentheses, such as (detail:1.3). Bare text such as :3, 16:9, 12:30, or detail:1.3 stays literal in both prompt inputs.

Large negative lists at strength 1.0 can dominate the prompt and produce unexpected results. Group related concepts and start with lower strengths:

(3D, CGI, render, blender, video game screenshot, illustration:0.25),
(text, writing, subtitle, watermark, logo:0.7),
(blurry, low quality, jpeg artifacts, grainy:0.4)

There is no node-level character limit. The effective token/context limit depends on the connected model and text encoder. The Qwen-based Z-Image and Krea 2 encoders support a much longer context than CLIP-based SD models, but long prompts can still use substantially more memory.

Z-Image Turbo normally remains at CFG 1.0. Use the normal CFG settings for SD1/SDXL workflows.

For Z-Image, compare prompt changes with the same seed. Broad semantic categories can react non-linearly, so start with a strength around 0.25 to 0.5 before trying 1.0.

Z-Image troubleshooting

• Connect patched_model from this node to the sampler or guider. Do not keep that input connected directly to the original model loader.
• Connect the node's positive and negative CONDITIONING outputs directly to the corresponding sampler/guider inputs.
• Use compiled_prompt only for inspection. Do not send it through another CLIP Text Encode node; Z-Image's normal tokenizer does not interpret NegPiP weights.
• With Asian in the negative input, compiled_prompt must show (Asian:-1). For a weaker test, enter (Asian:0.4) in the negative input, which compiles to (Asian:-0.4).

Krea 2 notes

• Load the text encoder with ComfyUI's CLIPLoader type krea2; a regular Qwen, Z-Image, or Flux CLIP is not interchangeable.
• Connect patched_model to the sampler and use the node's conditioning outputs directly, just as for Z-Image.
• The integrated Krea 2 path applies NegPiP to all 28 main transformer blocks and both text-fusion refiner blocks. Enabling the refiners is intentionally stronger than the standalone upstream node's default and is intended to reduce positive leakage into neighboring text tokens before the main model blocks. Start with lower prompt magnitudes if the effect is too strong.
• Prompt magnitudes remain available in the negative input. For example, (blurry:0.4) compiles to (blurry:-0.4).
• Complex conditioning transforms that normalize or clamp the Krea 2 conditioning tensor may destroy its embedded NegPiP sidecar. A metadata fallback is included, but direct connections from this node remain the safest path.

MOSS-Image / MOSS Turbo notes

MOSS support is new in 2.2.0 and experimental pending generation tests with a real checkpoint. Use ComfyUI's native MOSS-Image support and CLIPLoader type moss_image, with its Qwen3.5-9B text encoder. The older OpenMOSS custom HF loader is not supported by this NegPiP path.

For an SFT checkpoint with an 8-step SFT DMD LoRA:

1. Load the model, apply your model LoRAs including the DMD adapter, then connect the resulting MODEL to NegPiP Prompt (Multi-Model).
2. Connect the MOSS CLIP loader and separate positive/negative STRING inputs.
3. Connect patched_model and the node's conditioning outputs directly to the sampler. For a second latent-refinement pass, use the same patched model and conditioning outputs there as well.
4. Keep the Turbo sampling baseline: CFG 1, 8 steps, Euler / simple. With CyberMOSS Sampler 0.1.1+, choose turbo_reference and verify that schedule_mode is simple. NegPiP does not change the sigma schedule.

Start with a short negative prompt at a low magnitude, for example:

(watermark, text, jpeg artifacts:0.25)

The node encodes the positive prompt normally and negative spans separately, using MOSS's thinking-enabled chat template. It injects the negative embeddings before the template suffix and applies their signed magnitude to attention V projections in all main blocks and both text-refiner blocks. Both split Q/K/V and fused QKV projections are handled. The native attention backend, normalization, rotary positions, model LoRAs and other model patches remain in the computation. No extra negative diffusion pass is needed at CFG 1; the separate negative-span text encodes do add prompt-encoding work and memory use.

Negative magnitude 0 removes that span. Empty negatives keep the native positive-only computation. Positive prompt weighting otherwise follows the native MOSS tokenizer; this path does not add positive-weight syntax to it. Custom chat templates, embedding transforms that discard NegPiP metadata, reference-image editing, compiled-model execution and third-party quantization backends have not been validated. Reference-image editing is rejected explicitly.

Checks cover signed V projections, token/template alignment, batches, fused and split QKV, model/CLIP cloning, repeated prompts and CFG-1 tensor predictions. Passing these checks does not establish image quality or negative-prompt effectiveness on a particular SFT/DMD combination.

Compatibility

| Architecture | Status | Patch path | | --- | --- | --- | | Z-Image / Z-Image Turbo | Supported | Lumina2 / NextDiT / Qwen3-4B | | Krea 2 | Supported | SingleStreamDiT / Qwen3-VL-4B / 12-layer text fusion | | MOSS-Image / MOSS Turbo | Experimental | Native MOSSImage / Qwen3.5-9B / main + text-refiner V projections | | SD1 | Supported | paired CLIP embeddings + cross-attention patch | | SDXL / SDXL Refiner | Supported | paired CLIP embeddings + cross-attention patch | | Anima | Supported | Qwen3-0.6B encoder + internal T5 weight mask + current Cosmos API | | Flux | Not supported | upstream marks its NegPiP path as unmaintained |

The original workflow node ID, ZImageNegPipPrompt, is retained so workflows made with version 1.x continue to load. The visible node name and category are now model-neutral.

This package can be installed next to the original ComfyUI-ppm because it has a unique node ID and internal patch keys. Do not stack two NegPiP nodes on the same model/CLIP branch. If both connected inputs were already patched by ComfyUI-ppm's CLIP NegPip, this node reuses that patch instead of applying a second one.

The Anima implementation uses the current ComfyUI rotary-position API and does not import the removed apply_rotary_pos_emb function from comfy.ldm.cosmos.predict2.

Development checks

With Python 3.11+ and PyTorch installed, run:

python -B -m unittest discover -s tests -v

The MOSS tests use real CPU tensors and small attention models with isolated ComfyUI boundary doubles. They do not load checkpoint weights.

Credits and license

This implementation is derived from:

• ComfyUI-ppm by pamparamm;
• the Z-Image adaptation in BigStationW/ComfyUI-ppm;
• ComfyUI-krea2-negpip by blue-pen5805 (adapted from commit 3740add);
• ComfyUI-NegPipPromptMerge by Deathspike;
• NegPiP by laksjdjf and hako-mikan.

Released under the GNU Affero General Public License v3. See LICENSE.