[kijai]

Apparently even if not intended (as far as I know) it used to be possible to use the VL variant as the text encoder, functionality which broke in the Qwen3-VL text generation PR. This should restore that so users can generate text and flux2 prompt embeds with same load.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 3d01420c-5a51-4979-abf8-8f128280ede5

📥 Commits

Reviewing files that changed from the base of the PR and between a590d60 and fc1ca3c.

📒 Files selected for processing (1)

• comfy/sd.py

---

📝 Walkthrough

Walkthrough

In comfy/sd.py, the load_text_encoder_state_dicts function gains a new conditional branch inside the single-TE-model path for Qwen3-VL-detected text encoders. When clip_type is CLIPType.FLUX or CLIPType.FLUX2, the function now loads the Klein text-encoder variant (klein_te), selecting QWEN3VL_8B or QWEN3VL_4B as the model type and pairing it with KleinTokenizer8B or KleinTokenizer based on the detected model size.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and specifically references restoring Qwen3-VL functionality as a flux2 klein text encoder, which directly aligns with the changeset's purpose of re-enabling this capability.
Description check	✅ Passed	The description explains the restoration of previously broken Qwen3-VL text encoder functionality for flux2, directly relating to the code changes that add conditional logic for CLIPType.FLUX and CLIPType.FLUX2.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[LukeG89]

@kijai Something isn't working quite right with qwen3-vl-8b as TE for Klein. I wanted to try both bf16 and fp8_scaled, but the image tends to degrade during generation—sometimes, the more tokens you add to the prompt, the more it degrades. I had mixed results: in some cases, the degradation was barely visible, but other times the image went completely broken, especially in editing mode. With qwen3_8b, there are no issues at all; both generation and editing work perfectly.

Just reporting this to let you know (I will continue using the regular model as usual).