feat: Update qwen to support in loras version 1.0.0 (standalone) #665

GavChap · 2025-10-17T20:47:35Z

Updates only the ComfyUI package to support Qwen Loras without updating nunchaku version from 1.0.0.

Gives the following new nodes:

Lora loader
Lora Stacker (up to 6 loras)

These may be daisy-chained in order to load more.

Built upon the work of:
nunchaku-tech/nunchaku#680
and
nunchaku-tech/nunchaku#754

In my tests this loads almost all the loras I've tried with the correct weighting so images come out very well.

Issues known:

Unloads loras between runs to allow for chaining KSamplers.
Doesn't work with compile
~~Doesn't work with cpu offloading~~

…y nunchaku update)

zhangyi90825-tech · 2025-10-18T11:24:43Z

@GavChap How can I install it? Can you provide a tutorial

kajfblsdkgnlsndg · 2025-10-18T12:23:56Z

Looking forward to the official release of the merger

GavChap · 2025-10-18T17:11:24Z

@GavChap How can I install it? Can you provide a tutorial

It's way easier on Linux than windows, but the general gist is:

In the current custom_nodes/ComfyUI-nunchaku directory run these commands

git fetch origin pull/665/head:pr-665
git checkout pr-665

If you already have the nunchaku comfyui nodes installed, otherwise the easiest way is download the source .zip

https://github.com/GavChap/ComfyUI-nunchaku/archive/refs/heads/qwen-lora-suport-standalone.zip

and unzip that into a directory into ComfyUI/custom_nodes/ComfyUI-nunchaku (though it doesn't really matter what directory name, you just have to make sure it's a sub directory of custom_nodes)

Then run pip install -r requirements.txt

You also need to have the latest release (1.0.0) of Nunchaku installed via a wheel as you would for getting Qwen running anyway.

zhangyi90825-tech · 2025-10-19T17:32:07Z

@GavChap我该如何安装？你能提供一下教程吗？

在 Linux 上比在 Windows 上更容易，但总体要点是：

在当前custom_nodes/ComfyUI-nunchaku目录中运行以下命令

git fetch origin pull/665/head:pr-665 git checkout pr-665

如果您已经安装了 nunchaku comfyui 节点，否则最简单的方法是下载源 .zip

https://github.com/GavChap/ComfyUI-nunchaku/archive/refs/heads/qwen-lora-suport-standalone.zip

并将其解压到 ComfyUI/custom_nodes/ComfyUI-nunchaku 目录中（虽然目录名称并不重要，但你只需要确保它是 custom_nodes 的子目录）

然后运行pip install -r requirements.txt

无论如何，您还需要通过轮子安装最新版本 (1.0.0) 的 Nunchaku，就像运行 Qwen 一样。

I am using version 1.0.1, but calling Lora does not work

zhangyi90825-tech · 2025-10-19T17:33:17Z

@GavChap我安装了吗？你能提供一些教程吗？

在 Linux 上比在 Windows 上更容易，但总体要点是：
在当前custom_nodes/ComfyUI-nunchaku目录中运行以下命令
git fetch origin pull/665/head:pr-665 git checkout pr-665
如果您已经安装了 nunchaku comfyui 节点，否则最简单的方法是下载源 .zip
https://github.com/GavChap/ComfyUI-nunchaku/archive/refs/heads/qwen-lora-suport-standalone.zip
把它的解压到 ComfyUI/custom_nodes/ComfyUI-nunchaku 目录中（虽然目录名称并不重要，但你只需要确保它是 custom_nodes 的子目录）
然后运行pip install -r requirements.txt
无论如何，您还需要通过轮子安装最新版本 (1.0.0) 的双节棍，就像运行 Qwen 一样。

我使用的是 1.0.1 版本，但是无法调用 Lora

I tried multiple Loras, but the generated images still have the same effect as Qwen Image

GavChap · 2025-10-19T19:24:46Z

@GavChap我安装了吗？你能提供一些教程吗？

在 Linux 上比在 Windows 上更容易，但总体要点是：
在当前custom_nodes/ComfyUI-nunchaku目录中运行以下命令
git fetch origin pull/665/head:pr-665 git checkout pr-665
如果您已经安装了 nunchaku comfyui 节点，否则最简单的方法是下载源 .zip
https://github.com/GavChap/ComfyUI-nunchaku/archive/refs/heads/qwen-lora-suport-standalone.zip
把它的解压到 ComfyUI/custom_nodes/ComfyUI-nunchaku 目录中（虽然目录名称并不重要，但你只需要确保它是 custom_nodes 的子目录）
然后运行pip install -r requirements.txt
无论如何，您还需要通过轮子安装最新版本 (1.0.0) 的双节棍，就像运行 Qwen 一样。

我使用的是 1.0.1 版本，但是无法调用 Lora

I tried multiple Loras, but the generated images still have the same effect as Qwen Image

It should log to the ComfyUI log any errors it has loading loras. Can you link to one that doesn't work? I've tested it with Lightning loras and several realism loras and they all seem to work fine.

…_unet_

GavChap · 2025-10-19T20:24:22Z

**zhangyi90825-tech ** commented

@zhangyi90825-tech I believe I've now fixed it, please do another pull to update it.

zhangyi90825-tech · 2025-10-20T06:40:13Z

** zhangyi90825-tech ** 评论道

@zhangyi90825-tech我相信我现在已经修复了它，请再次拉取以更新它。

thank! I am already using it and the effect is very good But there are several bugs, 1. After switching Lora multiple times, the graphics memory will be overloaded, causing the graphics card driver to restart. 2. When using Nunkaku Qwen Image and switching back to Nunkaku Flux to generate images, it will crash

MarkShark2 · 2025-10-20T09:05:45Z

Works pretty good for me and loads most keys, but it did not work when I had CPU offloading enabled. I got this error:

RuntimeError: The size of tensor a (128) must match the size of tensor b (160) at non-singleton dimension 1

Gemini managed to fix this issue by disabling and re-enabling the offloader in the forward function of the ComfyQwenImageWrapper class of wrappers/qwenimage.py like so:

    # Check if the LoRA stack has been changed by a loader node
    if self._applied_loras != self.loras:
        # The compose function handles resetting before applying the new stack
        reset_lora_v2(self.model)
        self._applied_loras = self.loras.copy()
        compose_loras_v2(self.model, self.loras)

        # BUG FIX v2: Force a full teardown and rebuild of the OffloadManager
        # This ensures all old buffers are deallocated before creating new ones.
        if hasattr(self.model, "offload_manager") and self.model.offload_manager is not None:
            # Store the settings from the old manager before destroying it
            manager = self.model.offload_manager
            offload_settings = {
                "num_blocks_on_gpu": manager.num_blocks_on_gpu,
                "use_pin_memory": manager.use_pin_memory,
            }

            # Step 1: Completely disable and clear the old offloader
            self.model.set_offload(False)

            # Step 2: Re-enable offloading, forcing it to rebuild
            # with the new tensor shapes and the original settings.
            self.model.set_offload(True, **offload_settings)

And now it works great with CPU offloading. I don't know if this is the best solution, but it works for me.

GavChap · 2025-10-20T09:28:13Z

@zhangyi90825-tech When it crashes, please paste the log from comfyui in here

zhangyi90825-tech · 2025-10-20T09:30:14Z

@zhangyi90825-tech当它崩溃时，请将 comfyui 的日志粘贴到这里

Okay! I will send it here next time

GavChap · 2025-10-20T09:33:22Z

Works pretty good for me and loads most keys, but it did not work when I had CPU offloading enabled. I got this error:

RuntimeError: The size of tensor a (128) must match the size of tensor b (160) at non-singleton dimension 1

Gemini managed to fix this issue by disabling and re-enabling the offloader in the forward function of the ComfyQwenImageWrapper class of wrappers/qwenimage.py like so:

@MarkShark2 Thanks for that, I've included that fix.

cazs521 · 2025-10-20T10:53:31Z

'🔥 - 18 Nodes not included in prompt but is activated'
model weight dtype torch.bfloat16, manual cast: None
model_type FLUX
Disabling CPU offload
Total LoRAs in stack: 2
Requested to load NunchakuQwenImage
Composing 2 LoRAs...
Applied LoRA compositions to 480 modules.
Could not find/apply LoRA to 120 modules, showing first 5: ['transformer_blocks.0.img_mod.1', 'transformer_blocks.0.txt_mod.1', 'transformer_blocks.1.img_mod.1', 'transformer_blocks.1.txt_mod.1', 'transformer_blocks.10.img_mod.1']
loaded completely 4473.0 242.02829551696777 True
Prompt executed in 54.52 seconds

cazs521 · 2025-10-20T10:54:55Z

@GavChap This appeared for me

GavChap · 2025-10-20T11:08:48Z

'🔥 - 18 Nodes not included in prompt but is activated' model weight dtype torch.bfloat16, manual cast: None model_type FLUX Disabling CPU offload Total LoRAs in stack: 2 Requested to load NunchakuQwenImage Composing 2 LoRAs... Applied LoRA compositions to 480 modules. Could not find/apply LoRA to 120 modules, showing first 5: ['transformer_blocks.0.img_mod.1', 'transformer_blocks.0.txt_mod.1', 'transformer_blocks.1.img_mod.1', 'transformer_blocks.1.txt_mod.1', 'transformer_blocks.10.img_mod.1'] loaded completely 4473.0 242.02829551696777 True Prompt executed in 54.52 seconds

Where it states "Could not find/apply LoRA to 120 modules" it doesn't seem to mean much as the lora is still applied to the right number of modules and the right weights. However I have noticed that it detects as model Flux, which is weird. I'll look in to that

cazs521 · 2025-10-20T11:22:25Z

Yes, the model is correctly loaded. I tested it, and it's roughly the same as the original version with LoRA.

GavChap · 2025-10-20T11:29:52Z

Yes, the model is correctly loaded. I tested it, and it's roughly the same as the original version with LoRA.

It seems to be certain lora trainers that create layers that shouldn't be there, because some of mine trained with AI toolkit don't have any errors about layers and work perfectly.

cazs521 · 2025-10-20T11:50:50Z

You're right. I loaded two LoRAs: one is Lightning's acceleration LoRA, which runs perfectly, and the other one I downloaded randomly.

GavChap · 2025-10-20T14:49:33Z

I have noticed that it detects as model Flux, which is weird. I'll look in to that

As an aside, apparently that's correct for ComfyUI, Qwen is a "Flux" model type.

Firetheft · 2025-10-20T15:42:43Z

Great to finally see a workable solution—thanks for your effort, and we hope Dr. Li can take some time to review it!
and a few fix:
Unloads loras between runs to allow for chaining KSamplers.

change

        # Check if the LoRA stack has been changed by a loader node
        if self._applied_loras != self.loras:
            # The compose function handles resetting before applying the new stack
            reset_lora_v2(self.model)
            self._applied_loras = self.loras.copy()
            compose_loras_v2(self.model, self.loras)

            # BUG FIX v2: Force a full teardown and rebuild of the OffloadManager
            # This ensures all old buffers are deallocated before creating new ones.
            if hasattr(self.model, "offload_manager") and self.model.offload_manager is not None:

to

        # Check if the LoRA stack has been changed by a loader node
        loras_changed = self._applied_loras != self.loras
        
        # Additional checks: check if the underlying model is "dirty" (lora applied),
        # But this wrapper instance expects it to be clean (the loras list is empty).
        # This handles state out-of-sync issues caused by model deepcopy.
        model_is_dirty = (
            not self.loras and # We expect no LoRA
            hasattr(self.model, "_lora_slots") and self.model._lora_slots # But the model actually has LoRA
        )

        if loras_changed or model_is_dirty:
            # The compose function handles resetting before applying the new stack
            reset_lora_v2(self.model)
            self._applied_loras = self.loras.copy()
            compose_loras_v2(self.model, self.loras)

            # BUG FIX v2: Force a full teardown and rebuild of the OffloadManager
            if hasattr(self.model, "offload_manager") and self.model.offload_manager is not None:

…ra are going to cause a VRAM allocation issue and add safety margin parameter.

cazs521 · 2025-10-21T08:28:55Z

@GavChap 【Applied LoRA compositions to 480 modules.
Could not find/apply LoRA to 120 modules, showing first 5: ['transformer_blocks.0.img_mod.1', 'transformer_blocks.0.txt_mod.1', 'transformer_blocks.1.img_mod.1', 'transformer_blocks.1.txt_mod.1', 'transformer_blocks.10.img_mod.1']】 Actually, there is still a problem. The layers that weren't loaded in this part cause the output images to be insufficiently stylized. By increasing the weight of the LoRA, the original stylization effect of LoRA can appear. For example, when I increased the LoRA weight from 1 to 1.5, it worked completely.

GavChap · 2025-10-21T09:12:33Z

@GavChap 【Applied LoRA compositions to 480 modules. Could not find/apply LoRA to 120 modules, showing first 5: ['transformer_blocks.0.img_mod.1', 'transformer_blocks.0.txt_mod.1', 'transformer_blocks.1.img_mod.1', 'transformer_blocks.1.txt_mod.1', 'transformer_blocks.10.img_mod.1']】 Actually, there is still a problem. The layers that weren't loaded in this part cause the output images to be insufficiently stylized. By increasing the weight of the LoRA, the original stylization effect of LoRA can appear. For example, when I increased the LoRA weight from 1 to 1.5, it worked completely.

@cazs521 what lora is it? Can you link me to it?

peter4431 · 2025-10-21T12:04:52Z

@GavChap

An error will be reported when batch size > 1, and wrappers/qwenimage.py needs to be modified

-        timestep_float = timestep.item() if isinstance(timestep, torch.Tensor) else float(timestep)
+        if isinstance(timestep, torch.Tensor):
+            if timestep.numel() == 1:
+                timestep_float = timestep.item()
+            else:
+                timestep_float = timestep.flatten()[0].item()
+        else:
+            timestep_float = float(timestep)

cazs521 · 2025-10-21T13:31:48Z

@GavChap 【Applied LoRA compositions to 480 modules. Could not find/apply LoRA to 120 modules, showing first 5: ['transformer_blocks.0.img_mod.1', 'transformer_blocks.0.txt_mod.1', 'transformer_blocks.1.img_mod.1', 'transformer_blocks.1.txt_mod.1', 'transformer_blocks.10.img_mod.1']】 Actually, there is still a problem. The layers that weren't loaded in this part cause the output images to be insufficiently stylized. By increasing the weight of the LoRA, the original stylization effect of LoRA can appear. For example, when I increased the LoRA weight from 1 to 1.5, it worked completely.

@cazs521 what lora is it? Can you link me to it?

The insufficient image style is due to the lack of steps, CFG, and accelerated models.

…ones with all underscores, and stop yelling about unused blocks as it's fine.

defensezone · 2025-10-24T09:13:42Z

@GavChap How can I install it? Can you provide a tutorial

It's way easier on Linux than windows, but the general gist is:

In the current custom_nodes/ComfyUI-nunchaku directory run these commands

git fetch origin pull/665/head:pr-665 git checkout pr-665

If you already have the nunchaku comfyui nodes installed, otherwise the easiest way is download the source .zip

https://github.com/GavChap/ComfyUI-nunchaku/archive/refs/heads/qwen-lora-suport-standalone.zip

and unzip that into a directory into ComfyUI/custom_nodes/ComfyUI-nunchaku (though it doesn't really matter what directory name, you just have to make sure it's a sub directory of custom_nodes)

Then run pip install -r requirements.txt

You also need to have the latest release (1.0.0) of Nunchaku installed via a wheel as you would for getting Qwen running anyway.

I'm getting this error when trying to load a LoRA for Qwen using the Nunchaku FLUX LoRA Loader node.

`ot prompt
!!! Exception during processing !!!
Traceback (most recent call last):
File "W:\GIT\ComfyUI\execution.py", line 496, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "W:\GIT\ComfyUI\execution.py", line 315, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "W:\GIT\ComfyUI\execution.py", line 289, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "W:\GIT\ComfyUI\execution.py", line 277, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "W:\GIT\ComfyUI\custom_nodes\ComfyUI-nunchaku\nodes\lora\flux.py", line 114, in load_lora
assert isinstance(model_wrapper, ComfyFluxWrapper)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError

Prompt executed in 8.72 seconds`

GavChap · 2025-10-24T10:20:05Z

@GavChap How can I install it? Can you provide a tutorial

It's way easier on Linux than windows, but the general gist is:
In the current custom_nodes/ComfyUI-nunchaku directory run these commands
git fetch origin pull/665/head:pr-665 git checkout pr-665
If you already have the nunchaku comfyui nodes installed, otherwise the easiest way is download the source .zip
https://github.com/GavChap/ComfyUI-nunchaku/archive/refs/heads/qwen-lora-suport-standalone.zip
and unzip that into a directory into ComfyUI/custom_nodes/ComfyUI-nunchaku (though it doesn't really matter what directory name, you just have to make sure it's a sub directory of custom_nodes)
Then run pip install -r requirements.txt
You also need to have the latest release (1.0.0) of Nunchaku installed via a wheel as you would for getting Qwen running anyway.

I'm getting this error when trying to load a LoRA for Qwen using the Nunchaku FLUX LoRA Loader node.

`ot prompt !!! Exception during processing !!! Traceback (most recent call last): File "W:\GIT\ComfyUI\execution.py", line 496, in execute output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "W:\GIT\ComfyUI\execution.py", line 315, in get_output_data return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "W:\GIT\ComfyUI\execution.py", line 289, in _async_map_node_over_list await process_inputs(input_dict, i) File "W:\GIT\ComfyUI\execution.py", line 277, in process_inputs result = f(**inputs) ^^^^^^^^^^^ File "W:\GIT\ComfyUI\custom_nodes\ComfyUI-nunchaku\nodes\lora\flux.py", line 114, in load_lora assert isinstance(model_wrapper, ComfyFluxWrapper) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AssertionError

Prompt executed in 8.72 seconds`

Yes, because you need to use the new Nunchaku Qwen Lora Loader / Lora Stack instead.

zhangyi90825-tech · 2025-10-31T07:32:11Z

I have encountered a new problem again. When I use the Nvidia 3060 12G, I can generate images normally without using the Lora node. However, when I enable the Lora node, the error code is displayed The size of tensor a (32) must match the size of tensor b (96) at non-singleton dimension 1, May I ask how to solve this problem?

GavChap · 2025-10-31T12:10:24Z

I have encountered a new problem again. When I use the Nvidia 3060 12G, I can generate images normally without using the Lora node. However, when I enable the Lora node, the error code is displayed The size of tensor a (32) must match the size of tensor b (96) at non-singleton dimension 1, May I ask how to solve this problem?

@zhangyi90825-tech Can you link me to the lora so I can test? Also if you can upload an image with your workflow (with the node disabled) that'd help, also I think you're on an older version, as your dit loader doesn't have the extra field for vram auto-switching to cpu offload

zhangyi90825-tech · 2025-11-01T07:02:23Z

我又遇到了一个新问题。使用 Nvidia 3060 12G 显卡时，不使用 LoRa 节点可以正常生成图像。但是，启用 LoRa 节点后，会显示错误代码“张量 a (32) 的大小必须与张量 b (96) 的大小在非单例维度 1 处匹配”。请问如何解决这个问题？

@zhangyi90825-tech 另外，如果你能上传一张包含你的工作流程图觉得分区）的图片那就更好了。还有，我用你的版本比较旧，因为你的Dit Loader没有用于VRAM自动切换到CPU卸载的额外字段。

thank! After I update the new version, I can use it

C00reNUT · 2025-11-01T08:34:44Z

https://github.com/ussoewwin/ComfyUI-QwenImageLoraLoader is also an alternative

zhangyi90825-tech · 2025-11-03T13:24:07Z

救命，又出问题了，我不知道原因在哪里。

zhangyi90825-tech · 2025-11-03T13:34:34Z

救命，又出问题了，我不知道原因在哪里。

I encountered an error when updating today's version, but if I switch back to the previous version now, there will be no problem

# Conflicts: # __init__.py # nodes/models/flux.py # nodes/models/qwenimage.py

juntaosun · 2025-11-10T12:06:46Z

This image depicts an elegant Asian woman in a red suit, with fair skin. She is on the right side of the frame, leaning slightly forward, smiling as she hands a gold gift box tied with a red ribbon to another person. The box is inscribed with "Nunchaku." She wears a pearl necklace and delicate earrings, her golden curls cascading down her shoulders, exuding a noble air. Beside her lies a black Hermès handbag with a gold clasp, hinting at luxury and taste. The soft lighting and dim background highlight the woman and the gift, creating a warm, romantic, or celebratory atmosphere.

EXPERIMENTAL Expect OOM Credit: nunchaku-tech/ComfyUI-nunchaku#665

Soliver84 · 2025-11-21T21:50:44Z

RAM and dimension FIX:
comfyanonymous/ComfyUI#10802

feat: Update qwen to support in loras version 1.0.0 (standalone of an…

0432a96

…y nunchaku update)

GavChap added 2 commits October 19, 2025 21:15

feat: Update qwen regex conversion to allow for modules starting lora…

1cfe87a

…_unet_

feat: Make module selection more robust

eacbe79

fix: more keys to map

83555cb

fix: cpu offloading

950d66e

revert: flux model changes

dcfa7ff

GavChap added 2 commits October 20, 2025 17:41

feature: Add automatic calculation to use CPU Offloading if base + lo…

5543de2

…ra are going to cause a VRAM allocation issue and add safety margin parameter.

refactor: add dirty model check

8f3038f

fix: batches > 1

50b046e

ivancollider mentioned this pull request Oct 21, 2025

[Feature] Qwen image LORA support #479

Open

2 tasks

GavChap added 2 commits October 21, 2025 17:53

fix: update rules for lora conversion to make it more explicit about …

0f92779

…ones with all underscores, and stop yelling about unused blocks as it's fine.

fix: stop yelling about invalid blocks as it's fine.

e45f9b6

GavChap force-pushed the qwen-lora-suport-standalone branch from 83c670a to e45f9b6 Compare November 3, 2025 15:08

GavChap added 2 commits November 3, 2025 16:25

Merge branch 'main' into qwen-lora-suport-standalone

23636a9

# Conflicts: # __init__.py # nodes/models/flux.py # nodes/models/qwenimage.py

fix: merge updates from upstream

6a4f03d

Haoming02 added a commit to Haoming02/sd-webui-forge-classic that referenced this pull request Nov 12, 2025

support nunchaku qwen-image lora

6e5c405

EXPERIMENTAL Expect OOM Credit: nunchaku-tech/ComfyUI-nunchaku#665

feat: Update qwen to support in loras version 1.0.0 (standalone) #665

Are you sure you want to change the base?

feat: Update qwen to support in loras version 1.0.0 (standalone) #665

Uh oh!

Conversation

GavChap commented Oct 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zhangyi90825-tech commented Oct 18, 2025

Uh oh!

kajfblsdkgnlsndg commented Oct 18, 2025

Uh oh!

GavChap commented Oct 18, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zhangyi90825-tech commented Oct 19, 2025

Uh oh!

zhangyi90825-tech commented Oct 19, 2025

Uh oh!

GavChap commented Oct 19, 2025

Uh oh!

GavChap commented Oct 19, 2025

Uh oh!

zhangyi90825-tech commented Oct 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MarkShark2 commented Oct 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

GavChap commented Oct 20, 2025

Uh oh!

zhangyi90825-tech commented Oct 20, 2025

Uh oh!

GavChap commented Oct 20, 2025

Uh oh!

cazs521 commented Oct 20, 2025

Uh oh!

cazs521 commented Oct 20, 2025

Uh oh!

GavChap commented Oct 20, 2025

Uh oh!

cazs521 commented Oct 20, 2025

Uh oh!

GavChap commented Oct 20, 2025

Uh oh!

cazs521 commented Oct 20, 2025

Uh oh!

GavChap commented Oct 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Firetheft commented Oct 20, 2025

Uh oh!

cazs521 commented Oct 21, 2025

Uh oh!

GavChap commented Oct 21, 2025

Uh oh!

peter4431 commented Oct 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cazs521 commented Oct 21, 2025

Uh oh!

defensezone commented Oct 24, 2025

Uh oh!

GavChap commented Oct 24, 2025

Uh oh!

zhangyi90825-tech commented Oct 31, 2025

Uh oh!

GavChap commented Oct 31, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zhangyi90825-tech commented Nov 1, 2025

Uh oh!

C00reNUT commented Nov 1, 2025

Uh oh!

zhangyi90825-tech commented Nov 3, 2025

Uh oh!

zhangyi90825-tech commented Nov 3, 2025

Uh oh!

juntaosun commented Nov 10, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

GavChap commented Oct 17, 2025 •

edited

Loading

GavChap commented Oct 18, 2025 •

edited

Loading

zhangyi90825-tech commented Oct 20, 2025 •

edited

Loading

MarkShark2 commented Oct 20, 2025 •

edited

Loading

GavChap commented Oct 20, 2025 •

edited

Loading

peter4431 commented Oct 21, 2025 •

edited

Loading

GavChap commented Oct 31, 2025 •

edited

Loading

juntaosun commented Nov 10, 2025 •

edited

Loading