[Bug] Inconsistencies in Kontext Dev Support and Performance Between INT4 and FP8 Versions

[philipy1219]

Checklist

• 1. I have searched for related issues and FAQs (https://github.com/mit-han-lab/nunchaku/blob/main/docs/faq.md) but was unable to find a solution.
• 2. The issue persists in the latest version.
• 3. Please note that without environment information and a minimal reproducible example, it will be difficult for us to reproduce and address the issue, which may delay our response.
• 4. If your report is a question rather than a bug, please submit it as a discussion at https://github.com/mit-han-lab/ComfyUI-nunchaku/discussions/new/choose. Otherwise, this issue will be closed.
• 5. I will do my best to describe the issue in English.

Describe the Bug

We have conducted tests on the latest release with respect to Kontext Dev support and identified several issues in both single-reference and multi-reference configurations. Compared to the FP8 version, the output results differ noticeably.

Specifically, we observed that the INT4 version maintains a constant iteration speed of approximately 6 it/s, which matches the baseline text-to-image performance. However, in the case of Kontext, each additional reference image typically introduces a performance drop by a factor of around 1.7×, a behavior that is clearly present in the FP8 version but absent in the current INT4 implementation.

Furthermore, we noted that the main Nunchaku package remains at version 0.3.1, whereas comfyui-nunchaku has already been updated to version 0.3.3. Given that Kontext models support referenceLatent injection—a feature introduced in recent diffusers updates—we suspect that the root cause of the observed discrepancies may lie in the fact that the mainline Nunchaku package has not yet been updated accordingly.

We would appreciate any clarification on this, and whether an update to the main Nunchaku version is planned to align with the newer features.

Environment

Torch 2.6, nunchaku 0.3.1, comfyui-nunchaku 0.3.3

Reproduction Steps

Single reference

Workflow:

kontext-dev ComfyUI single reference.json

Input Image:

Output Image:

Multiple reference

Workflow:

kontext-dev ComfyUI multi reference.json

Input Image:

Output Image:

FP8 Output:

[0filler]

main Nunchaku package remains at version 0.3.1

Jun 15, 2025 release main Nunchaku package v0.3.1
Jun 16, 2025 release ComfyUI-nunchaku v0.3.2
The newest Pre-release is [Nunchaku Nightly v0.3.2dev20250630](later than ComfyUI-nunchaku v0.3.3), then [Nunchaku Nightly v0.3.2dev20250620]
So it's almost sure these two version numbers has no strict matching relationship.

[philipy1219]

main Nunchaku package remains at version 0.3.1

Jun 15, 2025 release main Nunchaku package v0.3.1 Jun 16, 2025 release ComfyUI-nunchaku v0.3.2 The newest Pre-release is [Nunchaku Nightly v0.3.2dev20250630](later than ComfyUI-nunchaku v0.3.3), then [Nunchaku Nightly v0.3.2dev20250620] So it's almost sure these two version numbers has no strict matching relationship.

ComfyUI has updated forward function for latent reference in kontext dev, I am not sure whether nunchaku support
it.

[lmxyy]

Thanks for letting us know. We haven't tested the multiple image cases. Will try this soon.

[philipy1219]

Thanks for letting us know. We haven't tested the multiple image cases. Will try this soon.

The single image example just print an image with single color, which is weird.

[philipy1219]

@lmxyy Hello, any updates for this issue?

[lmxyy]

Not yet for now. I think we can try requantizing the model with your provided examples in the future.

[github-actions]

This issue has been automatically closed due to 30-day inactivity. Please feel free to reopen it with /reopen if needed.