NIJIKAIDO W4A4 - ComfyUI Compatibility Fix

So I recently tried running NIJIKAIDO W4A4 in ComfyUI and kept getting completely black images. The model would load, go through all the sampling steps, and finish like everything was fine, but the actual output was just black.

After messing around with it for a while, I finally figured out a way to get it working.

Turns out the issue was related to how the W4A4 matrix multiplications were being handled. I managed to work around it by changing a setting in the model's quantization metadata.

No model weights were changed, and nothing was retrained or merged. It's still the same NIJIKAIDO model, just with a different computation path that seems to work better on certain GPUs.

Figured I'd upload it here in case anyone else is having the same issue.

Original Model & Credits

First off, all credit for the original NIJIKAIDO model goes to Vince_AI.

He's the one who created and released the original model. I only worked on making this particular W4A4 version more compatible with ComfyUI.

He also gave me permission to share this modified version, so huge thanks to him for being cool about it.

Original model:
https://civarchive.com/models/2984651/krea2turbonijikaidow4a4?modelVersionId=3383767

Original creator:
https://civarchive.com/user/Vince_AI

Definitely check out his original release.

About the Model

NIJIKAIDO is a stylized Krea 2 Turbo model focused on anime-inspired fantasy, expressive characters, dramatic lighting, and cinematic compositions.

It's particularly good for samurai, warriors, fantasy characters, creatures, dark fantasy, Japanese-inspired scenes, and colorful illustrations.

The actual model weights are exactly the same as the original release, so this version isn't supposed to change its visual style.

Tested GPUs

So far, the modified version has been successfully tested on:

• AMD Radeon RX 6900 XT (ROCm)

• NVIDIA GeForce GTX 1080

Both were able to generate images with the modified model.

I can't guarantee it'll work on every GPU, but if you're getting black images with the original W4A4 version, this might be worth trying.

If you test it on another GPU, feel free to leave a comment with your specs and whether it worked. I'm genuinely curious how many different cards can run it.

Recommended ComfyUI Settings

These are the settings I used as a starting point:

• Steps: 8

• CFG: 1.0

• Sampler: Euler

• Scheduler: Simple

• Text Encoder: Qwen3-VL 4B (compatible Krea 2 encoder)

• VAE: Qwen Image VAE

• CLIP Type: krea2

You'll need a ComfyUI version that supports this type of ConvRot W4A4 quantization.

Technical Details - What Actually Changed?

For anyone interested in how the fix works, here's a more detailed explanation.

The original model contains 224 ConvRot W4A4 quantized layers.

Each of these layers had the following setting in its quantization metadata:

full_precision_matrix_mult = false

I changed that setting to:

full_precision_matrix_mult = true

This tells ComfyUI to dequantize the affected weights for regular matrix multiplication instead of using the original quantized matrix multiplication path.

The weights are still stored in their W4A4-compressed form. They aren't permanently converted to FP16 or FP32. The difference is how those weights are handled during computation.

This can help on GPUs that don't properly support the original quantized computation path.

What changed in the file:

• 224 ConvRot W4A4 metadata flags changed from false to true

• Original Safetensors header: 95,768 bytes

• Modified header: 103,832 bytes

• Total tensors: 656, unchanged

• All original tensor weights preserved

• Tensor data verified using SHA-256

The original and modified files have identical tensor payloads.

Tensor payload SHA-256:

055f7c4255dde18aef160a925191d9677cf6e5424ccd64530b10af8f02a2f3ea

The full file hash is different because the header was modified, which is expected.

A quick note about performance:

Since this version uses dequantized weights for the affected matrix multiplications, it might be slower or require more temporary memory than the original W4A4 version running with a properly supported optimized kernel.

Also, full_precision_matrix_mult doesn't necessarily mean FP32. The actual computation dtype depends on your ComfyUI and PyTorch configuration.

This is mainly a compatibility workaround, not a performance optimization.

For more information, check out the official ComfyUI quantization documentation:

https://github.com/Comfy-Org/ComfyUI/blob/master/QUANTIZATION.md

That's pretty much it. Nothing fancy, just a small metadata change that ended up fixing a really annoying issue.

Hopefully it helps someone else who was stuck with the same black-image problem.

And again, thanks to Vince_AI for making the original model and letting me share this version.